diff --git a/README.md b/README.md index 50bc01a9bddf30d00b2dde2c98daff5e91e71569..fd8a99296da429b619cad709f90dbb580fd90188 100644 --- a/README.md +++ b/README.md @@ -1,2 +1,42 @@ -# vllm-auto-test +# vLLM 0.9.2 Management + +vLLM-0.9.2的软件版本管理及脚本程序管理 + +## 当前版本信息 + +1. 最新镜像 + + docker pull image.sourcefind.cn:5000/dcu/admin/base/vllm:0.9.2-ubuntu22.04-dtk25.04.1-rc5-rocblas104381-0915-das1.6-py3.10-20250916-rc2-ds3.2 + + 该镜像默认使用V1 Engine，并且默认开启了Prefix Caching功能 + +2. 推荐使用的环境变量（可作为测试参考） + + ``` + # 表示在并发数小于阈值时执行cudagraph模式，大于阈值时执行eager模式（产品建议K100AI设置成44，BW1000不设置该参数） + export VLLM_ENFORCE_EAGER_BS_THRESHOLD=44 + + # vllm 0.8.5/0.9.2版本出现精度乱码问题，可以尝试该环境变量解决 + export VLLM_USE_FLASH_ATTN_PA=0 + ``` + +2. 常见模型部署方法 + + 参考智算产品部提供的部署手册（定期更新）：https://r0ddbu55vzx.feishu.cn/docx/LL7KdYsWeoch7PxaS7wcBR5OnLe?from=from_copylink + +3. 常见模型性能摸测结果 + + 【金山文档 | WPS云文档】 大模型推理性能记录表_2025 https://www.kdocs.cn/l/cpcfyAiQx4WW（by 徐晓欧/刘煦/孙中谦） + +4. 通用小参数量的大模型推荐的w8a8精度的量化方法是compressed-tensors、quark、w8a8-dynamic？（by 王凯雄） + + DCU推荐使用: compressed-tensors + + compressed-tensors 量化方法见链接: https://sw4sldkryl8.feishu.cn/docx/RJqldrez2o477Cxyo40cg3Ven7h?from=from_copylink + + w8a8-dynamic 是华为卡上模型量化后的格式 + +## 代码/脚本更新日志 + +2025/10/16：增加最大并发量的测试方法（by 刘煦） diff --git a/depends/.gitkeep b/depends/.gitkeep new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/depends/README.md b/depends/README.md new file mode 100644 index 0000000000000000000000000000000000000000..14bfaf4282dabdd2b9a31f37cf8b716df9958e45 --- /dev/null +++ b/depends/README.md @@ -0,0 +1,16 @@ +# vLLM 0.9.2 Management + +离线环境安装包 + +api server测试方法需要jq和netcat依赖包，在离线环境内需要自行安装 + +jq_depends.zip和netcat_depends.zip是deb包，在ubuntu环境下可以直接安装 + +``` +unzip jq_depends.zip +unzip netcat_depends.zip +dpkg -i *.deb +``` + +jq-1.5.tar.gz和netcat-0.7.1.tar.gz是源码，可以直接源码编译 + diff --git a/depends/jq-1.5.tar.gz b/depends/jq-1.5.tar.gz new file mode 100644 index 0000000000000000000000000000000000000000..d034ef9459e2a8ca010cf080f1a01a28ef3283ab Binary files /dev/null and b/depends/jq-1.5.tar.gz differ diff --git a/depends/jq_depends.zip b/depends/jq_depends.zip new file mode 100644 index 0000000000000000000000000000000000000000..6d7f0364ac427562292b184ae6fb412ba4c0bff3 Binary files /dev/null and b/depends/jq_depends.zip differ diff --git a/depends/netcat-0.7.1.tar.gz b/depends/netcat-0.7.1.tar.gz new file mode 100644 index 0000000000000000000000000000000000000000..d9ddba836d07e8e6905a95e9c164de149e531fdc Binary files /dev/null and b/depends/netcat-0.7.1.tar.gz differ diff --git a/depends/netcat_depends.zip b/depends/netcat_depends.zip new file mode 100644 index 0000000000000000000000000000000000000000..5a19c13f30adc196041cf64b4ff28401c4d83015 Binary files /dev/null and b/depends/netcat_depends.zip differ diff --git a/offline_benchmark_test/README.md b/offline_benchmark_test/README.md new file mode 100644 index 0000000000000000000000000000000000000000..5aeb697667956b2e1440f4674ec8c23e5170cb7f --- /dev/null +++ b/offline_benchmark_test/README.md @@ -0,0 +1,47 @@ +# Offline Benchmark scripts + +vLLM 0.9.2 Benchmark测试的代码/脚本管理 + +## 更新日志 + +2025/10/16：新增vllm0.9.2离线测试的脚本 + +## 使用方法 + +1. 需要安装jq +``` + apt-get install jq +``` + +2. 文件说明： + + 1）auto_quick_check_config.json：配置文件 + + 2）auto_quick_check.sh：自动化启动脚本 + + 3）benchmarks：评测部分的代码，来源为vllm 0.9.2的GitHub；其中benchmark_throughput.py在官方基础上增加ttft、tpot的指标打印 + +``` +# 配置文件说明 auto_quick_check_config.json +{ + "DCU": "BW1000", # DCU型号 + "vllm_version": "0.9.2", # vllm版本号 + "pkg_version": "dtk25.04.1", # DTK版本号 + "dst_path": "./result/", # 结果/日志存放目录 + "items":[ + { + "model_name": "Qwen3-32B", # 模型名称 + "model_path": "/data/models/Qwen3-32B/", # 模型路径 + "dtype": "float16", # 数据类型 + "tensor_parallel": [4], # 使用卡数 + "batch_size": [128, 64, 32, 16], # 并发数 + "seqlen_tuple": ["512 512", "1024 1024"] # 输入输出序列组合 + } + ] +} +``` + +3. 运行指令： +``` +bash auto_quick_check.sh +``` diff --git a/offline_benchmark_test/auto_quick_check.sh b/offline_benchmark_test/auto_quick_check.sh new file mode 100644 index 0000000000000000000000000000000000000000..567d8aaf67840f32af07b973e593436f2ce20d29 --- /dev/null +++ b/offline_benchmark_test/auto_quick_check.sh @@ -0,0 +1,97 @@ +#!/bin/bash + +## 产品部提供的vllm 0.9.2默认环境变量，用于测试非DS 671B模型 +export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +export ALLREDUCE_STREAM_WITH_COMPUTE=1 +export NCCL_MIN_NCHANNELS=16 +export NCCL_MAX_NCHANNELS=16 +# export VLLM_PCIE_USE_CUSTOM_ALLREDUCE=1 #K100-AI需要 + +## 根据TOPO进行NUMA绑核 +export VLLM_NUMA_BIND=1 +export VLLM_RANK0_NUMA=0 +export VLLM_RANK1_NUMA=0 +export VLLM_RANK2_NUMA=0 +export VLLM_RANK3_NUMA=0 +export VLLM_RANK4_NUMA=0 +export VLLM_RANK5_NUMA=0 +export VLLM_RANK6_NUMA=0 +export VLLM_RANK7_NUMA=0 + +## 项目上反馈的可能对性能有提升的环境变量，可酌情使用 +# export HSA_FORCE_FINE_GRAIN_PCIE=1 +# export NCCL_P2P_LEVEL=SYS +# export NCCL_LAUNCH_MODE=GROUP +# export VLLM_RPC_TIMEOUT=1800000 +# export VLLM_SPEC_DECODE_EAGER=1 +# export VLLM_ENFORCE_EAGER_BS_THRESHOLD=44 # K100AI设置为44，BW1000去掉该环境变量 + +# export VLLM_MLA_DISABLE=0 +# export VLLM_USE_FLASH_MLA=1 +# export VLLM_ZERO_OVERHEAD=1 # 该参数对某些模型性能有明显提升；但在某些环境下表现异常，根据实际情况斟酌选用；且使用后对Qwen3模型会关闭其thinking功能 + +# export W8A8_SUPPORT_METHODS=3 # 对W8A8的量化模型有提升 +# export ROCBLAS_INT8_ENABLE=0 # 对W8A8的量化模型有提升 + +# export VLLM_USE_FLASH_ATTN_PA=0 # 解决精度乱码问题 + +# 禁止生成core文件 +ulimit -c 0 + +# 读取json配置文件 +json_data=$(cat auto_quick_check_config.json) + +# 获取基础信息 +DCU=$(echo $json_data | jq -r '.DCU') +vllm_version=$(echo $json_data | jq -r '.vllm_version') +pkg_version=$(echo $json_data | jq -r '.pkg_version') +dst_path=$(echo $json_data | jq -r '.dst_path') + +items=$(echo $json_data | jq -c '.items[]') +while read -r item; do + model_name=$(echo "$item" | jq -r '.model_name') + model_path=$(echo "$item" | jq -r '.model_path') + dtype=$(echo "$item" | jq -r '.dtype') + tensor_parallel=$(echo "$item" | jq -r '.tensor_parallel') + batch_size=$(echo "$item" | jq -r '.batch_size') + seqlen_tuple=$(echo "$item" | jq -r '.seqlen_tuple') + + result_path=${dst_path}/${model_name}/ + if [ ! -f ${result_path} ]; then + mkdir ${result_path} -p + fi + + if [ -e "${result_path}output.csv" ] && [ -s "${result_path}output.csv" ]; then + : + else + echo "model_name,DCU,DCU nums,precision,input_len,output_len,bs,TTFT_mean(s),TPOT_mean(s),GenerateThroughput(tokens/s),total_time(s),version" > ${result_path}output.csv + fi + echo $tensor_parallel | jq -c '.[]' | while read -r tp; do + echo $batch_size | jq -c '.[]' | while read -r bs; do + echo $seqlen_tuple | jq -c '.[]' | while read -r sq; do + IFS=' ' read -r input_len output_len <<< ${sq//\"/} + + case $dtype in + "float16" | "bfloat16") + extra="--dtype ${dtype}" ;; + "gptq") + extra="--quantization gptq" ;; + "awq") + extra="--quantization awq" ;; + esac + + # 当前离线测试脚本仅支持V0 Engine，推荐使用在线测试方法 + VLLM_USE_V1=0 python benchmarks/benchmark_throughput.py --model ${model_path} --tensor-parallel-size ${tp} \ + --num-prompts ${bs} --input-len ${input_len} --output-len ${output_len} ${extra} --trust-remote-code --max-model-len 32768 \ + 2>&1 | tee ${result_path}/vllm_${model_name}_bs_${bs}_inputlen_${input_len}_outputlen_${output_len}_tp_${tp}.log + + throughput=`grep -a "^Generate Throughput:" ${result_path}/vllm_${model_name}_bs_${bs}_inputlen_${input_len}_outputlen_${output_len}_tp_${tp}.log | awk -F ' ' '{print $3}'` + TTFT_mean=`grep -a "^TTFT mean:" ${result_path}/vllm_${model_name}_bs_${bs}_inputlen_${input_len}_outputlen_${output_len}_tp_${tp}.log | awk -F ' ' '{print $3}'` + TPOT_mean=`grep -a "^TPOT mean:" ${result_path}/vllm_${model_name}_bs_${bs}_inputlen_${input_len}_outputlen_${output_len}_tp_${tp}.log | awk -F ' ' '{print $3}'` + total_time=`grep -a "^Elapsed_time:" ${result_path}/vllm_${model_name}_bs_${bs}_inputlen_${input_len}_outputlen_${output_len}_tp_${tp}.log | awk -F ' ' '{print $2}'` + echo "$model_name,$DCU,$tp,$dtype,$input_len,$output_len,$bs,$TTFT_mean,$TPOT_mean,$throughput,$total_time,$pkg_version" >> ${result_path}output.csv + done + done + done + +done <<< "$items" diff --git a/offline_benchmark_test/auto_quick_check_config.json b/offline_benchmark_test/auto_quick_check_config.json new file mode 100644 index 0000000000000000000000000000000000000000..5eba2b094e834c08695d95903ea6dd31b0b7e6b6 --- /dev/null +++ b/offline_benchmark_test/auto_quick_check_config.json @@ -0,0 +1,16 @@ +{ + "DCU": "BW1000", + "vllm_version": "0.9.2", + "pkg_version": "dtk25.04.1", + "dst_path": "./result/", + "items":[ + { + "model_name": "Qwen3-32B", + "model_path": "/data/models/Qwen3-32B/", + "dtype": "float16", + "tensor_parallel": [4], + "batch_size": [128, 64, 32, 16], + "seqlen_tuple": ["512 512", "1024 1024"] + } + ] +} \ No newline at end of file diff --git a/offline_benchmark_test/benchmarks/README.md b/offline_benchmark_test/benchmarks/README.md new file mode 100644 index 0000000000000000000000000000000000000000..fb8690d42db98393c118947d7facce057f3cd48f --- /dev/null +++ b/offline_benchmark_test/benchmarks/README.md @@ -0,0 +1,605 @@ +# Benchmarking vLLM + +This README guides you through running benchmark tests with the extensive +datasets supported on vLLM. It’s a living document, updated as new features and datasets +become available. + +**Dataset Overview** + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
DatasetOnlineOfflineData Path
ShareGPT✅✅wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
BurstGPT✅✅wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv
Sonnet✅✅Local file: benchmarks/sonnet.txt
Random✅✅synthetic
HuggingFace-VisionArena✅✅lmarena-ai/VisionArena-Chat
HuggingFace-InstructCoder✅✅likaixin/InstructCoder
HuggingFace-AIMO✅✅AI-MO/aimo-validation-aime , AI-MO/NuminaMath-1.5, AI-MO/NuminaMath-CoT
HuggingFace-Other✅✅lmms-lab/LLaVA-OneVision-Data, Aeala/ShareGPT_Vicuna_unfiltered
Custom✅✅Local file: data.jsonl
+ +✅: supported + +🟡: Partial support + +🚧: to be supported + +**Note**: HuggingFace dataset's `dataset-name` should be set to `hf` + +--- +
+🚀 Example - Online Benchmark + +
+ +First start serving your model + +```bash +vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests +``` + +Then run the benchmarking script + +```bash +# download dataset +# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json +python3 vllm/benchmarks/benchmark_serving.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --endpoint /v1/completions \ + --dataset-name sharegpt \ + --dataset-path /ShareGPT_V3_unfiltered_cleaned_split.json \ + --num-prompts 10 +``` + +If successful, you will see the following output + +``` +============ Serving Benchmark Result ============ +Successful requests: 10 +Benchmark duration (s): 5.78 +Total input tokens: 1369 +Total generated tokens: 2212 +Request throughput (req/s): 1.73 +Output token throughput (tok/s): 382.89 +Total Token throughput (tok/s): 619.85 +---------------Time to First Token---------------- +Mean TTFT (ms): 71.54 +Median TTFT (ms): 73.88 +P99 TTFT (ms): 79.49 +-----Time per Output Token (excl. 1st token)------ +Mean TPOT (ms): 7.91 +Median TPOT (ms): 7.96 +P99 TPOT (ms): 8.03 +---------------Inter-token Latency---------------- +Mean ITL (ms): 7.74 +Median ITL (ms): 7.70 +P99 ITL (ms): 8.39 +================================================== +``` + +**Custom Dataset** + +If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl + +``` +{"prompt": "What is the capital of India?"} +{"prompt": "What is the capital of Iran?"} +{"prompt": "What is the capital of China?"} +``` + +```bash +# start server +VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct --disable-log-requests +``` + +```bash +# run benchmarking script +python3 benchmarks/benchmark_serving.py --port 9001 --save-result --save-detailed \ + --backend vllm \ + --model meta-llama/Llama-3.1-8B-Instruct \ + --endpoint /v1/completions \ + --dataset-name custom \ + --dataset-path \ + --custom-skip-chat-template \ + --num-prompts 80 \ + --max-concurrency 1 \ + --temperature=0.3 \ + --top-p=0.75 \ + --result-dir "./log/" +``` + +You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`. + +**VisionArena Benchmark for Vision Language Models** + +```bash +# need a model with vision capability here +vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests +``` + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend openai-chat \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --endpoint /v1/chat/completions \ + --dataset-name hf \ + --dataset-path lmarena-ai/VisionArena-Chat \ + --hf-split train \ + --num-prompts 1000 +``` + +**InstructCoder Benchmark with Speculative Decoding** + +``` bash +VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \ + --speculative-config $'{"method": "ngram", + "num_speculative_tokens": 5, "prompt_lookup_max": 5, + "prompt_lookup_min": 2}' +``` + +``` bash +python3 benchmarks/benchmark_serving.py \ + --model meta-llama/Meta-Llama-3-8B-Instruct \ + --dataset-name hf \ + --dataset-path likaixin/InstructCoder \ + --num-prompts 2048 +``` + +**Other HuggingFaceDataset Examples** + +```bash +vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests +``` + +**`lmms-lab/LLaVA-OneVision-Data`** + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend openai-chat \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --endpoint /v1/chat/completions \ + --dataset-name hf \ + --dataset-path lmms-lab/LLaVA-OneVision-Data \ + --hf-split train \ + --hf-subset "chart2text(cauldron)" \ + --num-prompts 10 +``` + +**`Aeala/ShareGPT_Vicuna_unfiltered`** + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend openai-chat \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --endpoint /v1/chat/completions \ + --dataset-name hf \ + --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \ + --hf-split train \ + --num-prompts 10 +``` + +**`AI-MO/aimo-validation-aime`** + +``` bash +python3 vllm/benchmarks/benchmark_serving.py \ + --model Qwen/QwQ-32B \ + --dataset-name hf \ + --dataset-path AI-MO/aimo-validation-aime \ + --num-prompts 10 \ + --seed 42 +``` + +**`philschmid/mt-bench`** + +``` bash +python3 vllm/benchmarks/benchmark_serving.py \ + --model Qwen/QwQ-32B \ + --dataset-name hf \ + --dataset-path philschmid/mt-bench \ + --num-prompts 80 +``` + +**Running With Sampling Parameters** + +When using OpenAI-compatible backends such as `vllm`, optional sampling +parameters can be specified. Example client command: + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --endpoint /v1/completions \ + --dataset-name sharegpt \ + --dataset-path /ShareGPT_V3_unfiltered_cleaned_split.json \ + --top-k 10 \ + --top-p 0.9 \ + --temperature 0.5 \ + --num-prompts 10 +``` + +**Running With Ramp-Up Request Rate** + +The benchmark tool also supports ramping up the request rate over the +duration of the benchmark run. This can be useful for stress testing the +server or finding the maximum throughput that it can handle, given some latency budget. + +Two ramp-up strategies are supported: +- `linear`: Increases the request rate linearly from a start value to an end value. +- `exponential`: Increases the request rate exponentially. + +The following arguments can be used to control the ramp-up: +- `--ramp-up-strategy`: The ramp-up strategy to use (`linear` or `exponential`). +- `--ramp-up-start-rps`: The request rate at the beginning of the benchmark. +- `--ramp-up-end-rps`: The request rate at the end of the benchmark. + +
+ +
+📈 Example - Offline Throughput Benchmark + +
+ +```bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset-name sonnet \ + --dataset-path vllm/benchmarks/sonnet.txt \ + --num-prompts 10 +``` + +If successful, you will see the following output + +``` +Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s +Total num prompt tokens: 5014 +Total num output tokens: 1500 +``` + +**VisionArena Benchmark for Vision Language Models** + +``` bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --backend vllm-chat \ + --dataset-name hf \ + --dataset-path lmarena-ai/VisionArena-Chat \ + --num-prompts 1000 \ + --hf-split train +``` + +The `num prompt tokens` now includes image token counts + +``` +Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s +Total num prompt tokens: 14527 +Total num output tokens: 1280 +``` + +**InstructCoder Benchmark with Speculative Decoding** + +``` bash +VLLM_WORKER_MULTIPROC_METHOD=spawn \ +VLLM_USE_V1=1 \ +python3 vllm/benchmarks/benchmark_throughput.py \ + --dataset-name=hf \ + --dataset-path=likaixin/InstructCoder \ + --model=meta-llama/Meta-Llama-3-8B-Instruct \ + --input-len=1000 \ + --output-len=100 \ + --num-prompts=2048 \ + --async-engine \ + --speculative-config $'{"method": "ngram", + "num_speculative_tokens": 5, "prompt_lookup_max": 5, + "prompt_lookup_min": 2}' +``` + +``` +Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s +Total num prompt tokens: 261136 +Total num output tokens: 204800 +``` + +**Other HuggingFaceDataset Examples** + +**`lmms-lab/LLaVA-OneVision-Data`** + +```bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --backend vllm-chat \ + --dataset-name hf \ + --dataset-path lmms-lab/LLaVA-OneVision-Data \ + --hf-split train \ + --hf-subset "chart2text(cauldron)" \ + --num-prompts 10 +``` + +**`Aeala/ShareGPT_Vicuna_unfiltered`** + +```bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --backend vllm-chat \ + --dataset-name hf \ + --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \ + --hf-split train \ + --num-prompts 10 +``` + +**`AI-MO/aimo-validation-aime`** + +```bash +python3 benchmarks/benchmark_throughput.py \ + --model Qwen/QwQ-32B \ + --backend vllm \ + --dataset-name hf \ + --dataset-path AI-MO/aimo-validation-aime \ + --hf-split train \ + --num-prompts 10 +``` + +**Benchmark with LoRA Adapters** + +``` bash +# download dataset +# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json +python3 vllm/benchmarks/benchmark_throughput.py \ + --model meta-llama/Llama-2-7b-hf \ + --backend vllm \ + --dataset_path /ShareGPT_V3_unfiltered_cleaned_split.json \ + --dataset_name sharegpt \ + --num-prompts 10 \ + --max-loras 2 \ + --max-lora-rank 8 \ + --enable-lora \ + --lora-path yard1/llama-2-7b-sql-lora-test + ``` + +
+ +
+🛠️ Example - Structured Output Benchmark + +
+ +Benchmark the performance of structured output generation (JSON, grammar, regex). + +**Server Setup** + +```bash +vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests +``` + +**JSON Schema Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset json \ + --structured-output-ratio 1.0 \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**Grammar-based Generation Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset grammar \ + --structure-type grammar \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**Regex-based Generation Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset regex \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**Choice-based Generation Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset choice \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**XGrammar Benchmark Dataset** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset xgrammar_bench \ + --request-rate 10 \ + --num-prompts 1000 +``` + +
+ +
+📚 Example - Long Document QA Benchmark + +
+ +Benchmark the performance of long document question-answering with prefix caching. + +**Basic Long Document QA Test** + +```bash +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 16 \ + --document-length 2000 \ + --output-len 50 \ + --repeat-count 5 +``` + +**Different Repeat Modes** + +```bash +# Random mode (default) - shuffle prompts randomly +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --document-length 3000 \ + --repeat-count 3 \ + --repeat-mode random + +# Tile mode - repeat entire prompt list in sequence +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --document-length 3000 \ + --repeat-count 3 \ + --repeat-mode tile + +# Interleave mode - repeat each prompt consecutively +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --document-length 3000 \ + --repeat-count 3 \ + --repeat-mode interleave +``` + +
+ +
+🗂️ Example - Prefix Caching Benchmark + +
+ +Benchmark the efficiency of automatic prefix caching. + +**Fixed Prompt with Prefix Caching** + +```bash +python3 benchmarks/benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-prompts 1 \ + --repeat-count 100 \ + --input-length-range 128:256 +``` + +**ShareGPT Dataset with Prefix Caching** + +```bash +# download dataset +# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json + +python3 benchmarks/benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --dataset-path /path/ShareGPT_V3_unfiltered_cleaned_split.json \ + --enable-prefix-caching \ + --num-prompts 20 \ + --repeat-count 5 \ + --input-length-range 128:256 +``` + +
+ +
+⚡ Example - Request Prioritization Benchmark + +
+ +Benchmark the performance of request prioritization in vLLM. + +**Basic Prioritization Test** + +```bash +python3 benchmarks/benchmark_prioritization.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --input-len 128 \ + --output-len 64 \ + --num-prompts 100 \ + --scheduling-policy priority +``` + +**Multiple Sequences per Prompt** + +```bash +python3 benchmarks/benchmark_prioritization.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --input-len 128 \ + --output-len 64 \ + --num-prompts 100 \ + --scheduling-policy priority \ + --n 2 +``` + +
diff --git a/offline_benchmark_test/benchmarks/auto_tune.sh b/offline_benchmark_test/benchmarks/auto_tune.sh new file mode 100644 index 0000000000000000000000000000000000000000..b257b57ce06f5038338bc8a8f114fcb13c0e203b --- /dev/null +++ b/offline_benchmark_test/benchmarks/auto_tune.sh @@ -0,0 +1,276 @@ +#!/bin/bash + +# This script aims to tune the best server parameter combinations to maximize throughput for given requirement. +# The current server parameter combination is max_num_seqs and max_num_batched_tokens +# It also supports additional requirement: e2e latency and prefix cache. + +# Pre-requisite: +# 1. Checkout to your branch, install/ update the correct running env. For TPU, activate conda env and install the corresponding torch, xla version. +# 2. If the model is customized, replace the MODEL's config with the customized config. +# 3. Set variables (ALL REQUIRED) +# BASE: your directory for vllm repo +# MODEL: the model served by vllm +# SYSTEM: the hardware, choice TPU or GPU, for other systems, "get best profile" might not support. +# TP: ways of tensor parallelism +# DOWNLOAD_DIR: directory to download and load model weights. +# INPUT_LEN: request input len +# OUTPUT_LEN: request output len +# MIN_CACHE_HIT_PCT: prefix cache rate +# MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000 +# NUM_SEQS_LIST: a list of `max-num-seqs` you want to loop with. +# NUM_BATCHED_TOKENS_LIST: a list of `max-num-batched-tokens` you want to loop with. +# Note that the default NUM_SEQS_LIST and NUM_BATCHED_TOKENS_LIST are set for medium size input/output len, for extra short context (such as 20:20), you might need to include larger numbers in NUM_SEQS_LIST. +# 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens. +# 5. The final result will be saved in RESULT file. + + +# Example use cases +# 1. Given input_len=1800, output_len=20, what's the best max_num_seqs and max_num_batched_tokens to get highest throughput? +# Use INPUT_LEN=1800, OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=100000000000 +# 2. If we have latency requirement to be lower than 500ms, what's the best server parameter? +# Use INPUT_LEN=1800, OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=500 +# 3. If we want to reach 60% prefix cache, what's the best server parameter? +# Use INPUT_LEN=1800, OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=60, MAX_LATENCY_ALLOWED_MS=500 + +TAG=$(date +"%Y_%m_%d_%H_%M") +BASE="" +MODEL="meta-llama/Llama-3.1-8B-Instruct" +SYSTEM="TPU" +TP=1 +DOWNLOAD_DIR="" +INPUT_LEN=4000 +OUTPUT_LEN=16 +MIN_CACHE_HIT_PCT=0 +MAX_LATENCY_ALLOWED_MS=100000000000 +NUM_SEQS_LIST="128 256" +NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096" + +LOG_FOLDER="$BASE/auto-benchmark/$TAG" +RESULT="$LOG_FOLDER/result.txt" +PROFILE_PATH="$LOG_FOLDER/profile" + +echo "result file: $RESULT" +echo "model: $MODEL" + +rm -rf $LOG_FOLDER +rm -rf $PROFILE_PATH +mkdir -p $LOG_FOLDER +mkdir -p $PROFILE_PATH + +cd "$BASE/vllm" + +pip install -q datasets + +current_hash=$(git rev-parse HEAD) +echo "hash:$current_hash" >> "$RESULT" +echo "current_hash: $current_hash" + +best_throughput=0 +best_max_num_seqs=0 +best_num_batched_tokens=0 +best_goodput=0 + +start_server() { + local gpu_memory_utilization=$1 + local max_num_seqs=$2 + local max_num_batched_tokens=$3 + local vllm_log=$4 + local profile_dir=$5 + + pkill -f vllm + + VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \ + --disable-log-requests \ + --port 8004 \ + --gpu-memory-utilization $gpu_memory_utilization \ + --max-num-seqs $max_num_seqs \ + --max-num-batched-tokens $max_num_batched_tokens \ + --tensor-parallel-size $TP \ + --enable-prefix-caching \ + --load-format dummy \ + --download-dir "$DOWNLOAD_DIR" \ + --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 & + + # wait for 10 minutes... + server_started=0 + for i in {1..60}; do + RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout) + STATUS_CODE=$(echo "$RESPONSE" | tail -n 1) + if [[ "$STATUS_CODE" -eq 200 ]]; then + server_started=1 + break + else + sleep 10 + fi + done + if (( ! server_started )); then + echo "server did not start within 10 minutes. Please check server log at $vllm_log". + return 1 + else + return 0 + fi +} + +update_best_profile() { + local profile_dir=$1 + local profile_index=$2 + sorted_paths=($(find "$profile_dir" -maxdepth 1 -not -path "$profile_dir" | sort)) + selected_profile_file= + if [[ "$SYSTEM" == "TPU" ]]; then + selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb" + fi + if [[ "$SYSTEM" == "GPU" ]]; then + selected_profile_file="${sorted_paths[$profile_index]}" + fi + rm -f $PROFILE_PATH/* + cp $selected_profile_file $PROFILE_PATH +} + +run_benchmark() { + local max_num_seqs=$1 + local max_num_batched_tokens=$2 + local gpu_memory_utilization=$3 + echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens" + local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt" + local profile_dir="$LOG_FOLDER/profile_${max_num_seqs}_${max_num_batched_tokens}" + echo "vllm_log: $vllm_log" + echo + rm -f $vllm_log + mkdir -p $profile_dir + pkill -f vllm + local profile_index=0 + + echo "starting server..." + start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log $profile_dir + result=$? + if [[ "$result" -eq 1 ]]; then + echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens" + else + echo "server started." + fi + echo + + echo "run benchmark test..." + meet_latency_requirement=0 + # get a basic qps by using request-rate inf + bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt" + prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 )) + python benchmarks/benchmark_serving.py \ + --backend vllm \ + --model $MODEL \ + --dataset-name random \ + --random-input-len $INPUT_LEN \ + --random-output-len $OUTPUT_LEN \ + --ignore-eos \ + --disable-tqdm \ + --request-rate inf \ + --percentile-metrics ttft,tpot,itl,e2el \ + --goodput e2el:$MAX_LATENCY_ALLOWED_MS \ + --num-prompts 1000 \ + --random-prefix-len $prefix_len \ + --port 8004 \ + --profile &> "$bm_log" + throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}') + goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + + if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then + meet_latency_requirement=1 + request_rate=inf + fi + + if (( ! meet_latency_requirement )); then + # start from request-rate as int(throughput) + 1 + request_rate=$((${throughput%.*} + 1)) + while ((request_rate > 0)); do + profile_index=$((profile_index+1)) + # clear prefix cache + curl -X POST http://0.0.0.0:8004/reset_prefix_cache + sleep 5 + bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt" + python benchmarks/benchmark_serving.py \ + --backend vllm \ + --model $MODEL \ + --dataset-name random \ + --random-input-len $INPUT_LEN \ + --random-output-len $OUTPUT_LEN \ + --ignore-eos \ + --disable-tqdm \ + --request-rate $request_rate \ + --percentile-metrics ttft,tpot,itl,e2el \ + --goodput e2el:$MAX_LATENCY_ALLOWED_MS \ + --num-prompts 100 \ + --random-prefix-len $prefix_len \ + --port 8004 &> "$bm_log" + throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}') + goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then + meet_latency_requirement=1 + break + fi + request_rate=$((request_rate-1)) + done + fi + # write the results and update the best result. + if ((meet_latency_requirement)); then + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" >> "$RESULT" + if (( $(echo "$throughput > $best_throughput" | bc -l) )); then + best_throughput=$throughput + best_max_num_seqs=$max_num_seqs + best_num_batched_tokens=$max_num_batched_tokens + best_goodput=$goodput + if [[ "$SYSTEM" == "TPU" ]]; then + update_best_profile "$profile_dir/plugins/profile" $profile_index + fi + if [[ "$SYSTEM" == "GPU" ]]; then + update_best_profile "$profile_dir" $profile_index + fi + fi + else + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" >> "$RESULT" + fi + + echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput" + + pkill vllm + sleep 10 + printf '=%.0s' $(seq 1 20) + return 0 +} + +read -r -a num_seqs_list <<< "$NUM_SEQS_LIST" +read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST" + +# first find out the max gpu-memory-utilization without HBM OOM. +gpu_memory_utilization=0.98 +find_gpu_memory_utilization=0 +while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do + start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log" + result=$? + if [[ "$result" -eq 0 ]]; then + find_gpu_memory_utilization=1 + break + else + gpu_memory_utilization=$(echo "$gpu_memory_utilization - 0.01" | bc) + fi +done + +if [[ "$find_gpu_memory_utilization" -eq 1 ]]; then + echo "Using gpu_memory_utilization=$gpu_memory_utilization to serve model." +else + echo "Cannot find a proper gpu_memory_utilization over 0.9 to serve the model, please check logs in $LOG_FOLDER." + exit 1 +fi + +for num_seqs in "${num_seqs_list[@]}"; do + for num_batched_tokens in "${num_batched_tokens_list[@]}"; do + run_benchmark $num_seqs $num_batched_tokens $gpu_memory_utilization + done +done +echo "finish permutations" +echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" +echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT" + diff --git a/offline_benchmark_test/benchmarks/backend_request_func.py b/offline_benchmark_test/benchmarks/backend_request_func.py new file mode 100644 index 0000000000000000000000000000000000000000..c7229dbb8e90d1167eab8fda1a5fb353e31e81a4 --- /dev/null +++ b/offline_benchmark_test/benchmarks/backend_request_func.py @@ -0,0 +1,628 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import io +import json +import os +import sys +import time +import traceback +from dataclasses import dataclass, field +from typing import Optional, Union + +import aiohttp +import huggingface_hub.constants +from tqdm.asyncio import tqdm +from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast + +# NOTE(simon): do not import vLLM here so the benchmark script +# can run without vLLM installed. + +AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60) + + +@dataclass +class RequestFuncInput: + prompt: str + api_url: str + prompt_len: int + output_len: int + model: str + model_name: Optional[str] = None + logprobs: Optional[int] = None + extra_body: Optional[dict] = None + multi_modal_content: Optional[dict] = None + ignore_eos: bool = False + language: Optional[str] = None + + +@dataclass +class RequestFuncOutput: + generated_text: str = "" + success: bool = False + latency: float = 0.0 + output_tokens: int = 0 + ttft: float = 0.0 # Time to first token + itl: list[float] = field(default_factory=list) # list of inter-token latencies + tpot: float = 0.0 # avg next-token latencies + prompt_len: int = 0 + error: str = "" + + +async def async_request_tgi( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith("generate_stream") + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + params = { + "max_new_tokens": request_func_input.output_len, + "do_sample": True, + "temperature": 0.01, # TGI does not accept 0.0 temperature. + "top_p": 0.99, # TGI does not accept 1.0 top_p. + "truncate": request_func_input.prompt_len, + "ignore_eos_token": request_func_input.ignore_eos, + } + payload = { + "inputs": request_func_input.prompt, + "parameters": params, + } + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + if request_func_input.ignore_eos: + output.output_tokens = request_func_input.output_len + else: + output.output_tokens = None + + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post(url=api_url, json=payload) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + chunk_bytes = chunk_bytes.decode("utf-8") + + # NOTE: Sometimes TGI returns a ping response without + # any data, we should skip it. + if chunk_bytes.startswith(":"): + continue + chunk = chunk_bytes.removeprefix("data:") + + data = json.loads(chunk) + timestamp = time.perf_counter() + # First token + if ttft == 0.0: + ttft = time.perf_counter() - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + most_recent_timestamp = timestamp + + output.latency = most_recent_timestamp - st + output.success = True + output.generated_text = data["generated_text"] + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_trt_llm( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith("generate_stream") + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + payload = { + "accumulate_tokens": True, + "text_input": request_func_input.prompt, + "temperature": 0.0, + "top_p": 1.0, + "max_tokens": request_func_input.output_len, + "stream": True, + } + if request_func_input.ignore_eos: + payload["min_length"] = request_func_input.output_len + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post(url=api_url, json=payload) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + + chunk = chunk_bytes.decode("utf-8").removeprefix("data:") + + data = json.loads(chunk) + output.generated_text += data["text_output"] + timestamp = time.perf_counter() + # First token + if ttft == 0.0: + ttft = timestamp - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + most_recent_timestamp = timestamp + + output.latency = most_recent_timestamp - st + output.success = True + + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_deepspeed_mii( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith(("completions", "profile")), ( + "OpenAI Completions API URL must end with 'completions' or 'profile'." + ) + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + payload = { + "model": request_func_input.model, + "prompt": request_func_input.prompt, + "max_tokens": request_func_input.output_len, + "temperature": 0.01, # deepspeed-mii does not accept 0.0 temp. + "top_p": 1.0, + } + headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"} + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + # NOTE: DeepSpeed-MII doesn't support streaming as of Jan 28 2024, + # will use 0 as placeholder. + # See https://github.com/microsoft/DeepSpeed-MII/pull/311 + output.ttft = 0 + + st = time.perf_counter() + try: + async with session.post( + url=api_url, json=payload, headers=headers + ) as response: + if response.status == 200: + parsed_resp = await response.json() + output.latency = time.perf_counter() - st + if "choices" in parsed_resp: + output.generated_text = parsed_resp["choices"][0]["text"] + elif "text" in parsed_resp: + output.generated_text = parsed_resp["text"][0] + else: + output.error = ( + "Unexpected response format: " + "neither 'choices' nor 'text' found" + ) + output.success = False + output.success = True + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_openai_completions( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith(("completions", "profile")), ( + "OpenAI Completions API URL must end with 'completions' or 'profile'." + ) + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + payload = { + "model": request_func_input.model_name + if request_func_input.model_name + else request_func_input.model, + "prompt": request_func_input.prompt, + "temperature": 0.0, + "repetition_penalty": 1.0, + "max_tokens": request_func_input.output_len, + "logprobs": request_func_input.logprobs, + "stream": True, + "stream_options": { + "include_usage": True, + }, + } + if request_func_input.ignore_eos: + payload["ignore_eos"] = request_func_input.ignore_eos + if request_func_input.extra_body: + payload.update(request_func_input.extra_body) + headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"} + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + generated_text = "" + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post( + url=api_url, json=payload, headers=headers + ) as response: + if response.status == 200: + first_chunk_received = False + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + + chunk = chunk_bytes.decode("utf-8").removeprefix("data: ") + if chunk != "[DONE]": + data = json.loads(chunk) + + # NOTE: Some completion API might have a last + # usage summary response without a token so we + # want to check a token was generated + if choices := data.get("choices"): + # Note that text could be empty here + # e.g. for special tokens + text = choices[0].get("text") + timestamp = time.perf_counter() + # First token + if not first_chunk_received: + first_chunk_received = True + ttft = time.perf_counter() - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + most_recent_timestamp = timestamp + generated_text += text or "" + if usage := data.get("usage"): + output.output_tokens = usage.get("completion_tokens") + if first_chunk_received: + output.success = True + else: + output.success = False + output.error = ( + "Never received a valid chunk to calculate TTFT." + "This response will be marked as failed!" + ) + output.generated_text = generated_text + output.latency = most_recent_timestamp - st + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_openai_chat_completions( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith(("chat/completions", "profile")), ( + "OpenAI Chat Completions API URL must end with 'chat/completions'." + ) + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + content = [{"type": "text", "text": request_func_input.prompt}] + if request_func_input.multi_modal_content: + content.append(request_func_input.multi_modal_content) + payload = { + "model": request_func_input.model_name + if request_func_input.model_name + else request_func_input.model, + "messages": [ + {"role": "user", "content": content}, + ], + "temperature": 0.0, + "max_completion_tokens": request_func_input.output_len, + "stream": True, + "stream_options": { + "include_usage": True, + }, + } + if request_func_input.ignore_eos: + payload["ignore_eos"] = request_func_input.ignore_eos + if request_func_input.extra_body: + payload.update(request_func_input.extra_body) + headers = { + "Content-Type": "application/json", + "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}", + } + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + generated_text = "" + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post( + url=api_url, json=payload, headers=headers + ) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + chunk_bytes = chunk_bytes.decode("utf-8") + # NOTE: SSE comments (often used as pings) start with a colon. + # These are not JSON data payload and should be skipped. + if chunk_bytes.startswith(":"): + continue + + chunk = chunk_bytes.removeprefix("data: ") + + if chunk != "[DONE]": + timestamp = time.perf_counter() + data = json.loads(chunk) + + if choices := data.get("choices"): + content = choices[0]["delta"].get("content") + # First token + if ttft == 0.0: + ttft = timestamp - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + generated_text += content or "" + elif usage := data.get("usage"): + output.output_tokens = usage.get("completion_tokens") + + most_recent_timestamp = timestamp + + output.generated_text = generated_text + output.success = True + output.latency = most_recent_timestamp - st + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_openai_audio( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + # Lazy import without PlaceholderModule to avoid vllm dep. + import soundfile + + api_url = request_func_input.api_url + assert api_url.endswith(("transcriptions", "translations")), ( + "OpenAI Chat Completions API URL must end with 'transcriptions' " + ) + "or `translations`." + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + content = [{"type": "text", "text": request_func_input.prompt}] + payload = { + "model": request_func_input.model_name + if request_func_input.model_name + else request_func_input.model, + "temperature": 0.0, + "max_completion_tokens": request_func_input.output_len, + "stream": True, + "language": "en", + # Flattened due to multipart/form-data + "stream_include_usage": True, + "stream_continuous_usage_stats": True, + } + if request_func_input.extra_body: + payload.update(request_func_input.extra_body) + headers = { + "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}", + } + + # Send audio file + def to_bytes(y, sr): + buffer = io.BytesIO() + soundfile.write(buffer, y, sr, format="WAV") + buffer.seek(0) + return buffer + + with to_bytes(*request_func_input.multi_modal_content["audio"]) as f: + form = aiohttp.FormData() + form.add_field("file", f, content_type="audio/wav") + for key, value in payload.items(): + form.add_field(key, str(value)) + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + generated_text = "" + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post( + url=api_url, data=form, headers=headers + ) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + + chunk = chunk_bytes.decode("utf-8").removeprefix("data: ") + if chunk != "[DONE]": + timestamp = time.perf_counter() + data = json.loads(chunk) + + if choices := data.get("choices"): + content = choices[0]["delta"].get("content") + # First token + if ttft == 0.0: + ttft = timestamp - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append( + timestamp - most_recent_timestamp + ) + + generated_text += content or "" + elif usage := data.get("usage"): + output.output_tokens = usage.get( + "completion_tokens" + ) + + most_recent_timestamp = timestamp + + output.generated_text = generated_text + output.success = True + output.latency = most_recent_timestamp - st + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +def get_model(pretrained_model_name_or_path: str) -> str: + if os.getenv("VLLM_USE_MODELSCOPE", "False").lower() == "true": + from modelscope import snapshot_download + + from vllm.model_executor.model_loader.weight_utils import get_lock + + # Use file lock to prevent multiple processes from + # downloading the same model weights at the same time. + with get_lock(pretrained_model_name_or_path): + model_path = snapshot_download( + model_id=pretrained_model_name_or_path, + local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, + ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"], + ) + + return model_path + return pretrained_model_name_or_path + + +def get_tokenizer( + pretrained_model_name_or_path: str, + tokenizer_mode: str = "auto", + trust_remote_code: bool = False, + **kwargs, +) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]: + if pretrained_model_name_or_path is not None and not os.path.exists( + pretrained_model_name_or_path + ): + pretrained_model_name_or_path = get_model(pretrained_model_name_or_path) + if tokenizer_mode == "slow": + if kwargs.get("use_fast", False): + raise ValueError("Cannot use the fast tokenizer in slow tokenizer mode.") + kwargs["use_fast"] = False + if tokenizer_mode == "mistral": + try: + from vllm.transformers_utils.tokenizer import MistralTokenizer + except ImportError as e: + raise ImportError( + "MistralTokenizer requires vllm package.\n" + "Please install it with `pip install vllm` " + "to use mistral tokenizer mode." + ) from e + return MistralTokenizer.from_pretrained(str(pretrained_model_name_or_path)) + else: + return AutoTokenizer.from_pretrained( + pretrained_model_name_or_path, + trust_remote_code=trust_remote_code, + **kwargs, + ) + + +ASYNC_REQUEST_FUNCS = { + "tgi": async_request_tgi, + "vllm": async_request_openai_completions, + "lmdeploy": async_request_openai_completions, + "deepspeed-mii": async_request_deepspeed_mii, + "openai": async_request_openai_completions, + "openai-chat": async_request_openai_chat_completions, + "openai-audio": async_request_openai_audio, + "tensorrt-llm": async_request_trt_llm, + "scalellm": async_request_openai_completions, + "sglang": async_request_openai_completions, + "llama.cpp": async_request_openai_completions, +} + +OPENAI_COMPATIBLE_BACKENDS = [ + k + for k, v in ASYNC_REQUEST_FUNCS.items() + if v in (async_request_openai_completions, async_request_openai_chat_completions) +] diff --git a/offline_benchmark_test/benchmarks/benchmark_dataset.py b/offline_benchmark_test/benchmarks/benchmark_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..55c0cf851264f37d271f7a8cd1fc5ad17451f386 --- /dev/null +++ b/offline_benchmark_test/benchmarks/benchmark_dataset.py @@ -0,0 +1,1168 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +This module defines a framework for sampling benchmark requests from various +datasets. Each dataset subclass of BenchmarkDataset must implement sample +generation. Supported dataset types include: + - ShareGPT + - Random (synthetic) + - Sonnet + - BurstGPT + - HuggingFace + - VisionArena +""" + +import base64 +import io +import json +import logging +import random +from abc import ABC, abstractmethod +from collections.abc import Mapping +from dataclasses import dataclass +from functools import cache +from io import BytesIO +from typing import Any, Callable, Optional, Union + +import numpy as np +import pandas as pd +from datasets import load_dataset +from PIL import Image +from transformers import PreTrainedTokenizerBase + +from vllm.lora.request import LoRARequest +from vllm.lora.utils import get_adapter_absolute_path +from vllm.multimodal import MultiModalDataDict +from vllm.multimodal.image import convert_image_mode +from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer + +logger = logging.getLogger(__name__) + +# ----------------------------------------------------------------------------- +# Data Classes +# ----------------------------------------------------------------------------- + + +@dataclass +class SampleRequest: + """ + Represents a single inference request for benchmarking. + """ + + prompt: Union[str, Any] + prompt_len: int + expected_output_len: int + multi_modal_data: Optional[Union[MultiModalDataDict, dict]] = None + lora_request: Optional[LoRARequest] = None + + +# ----------------------------------------------------------------------------- +# Benchmark Dataset Base Class +# ----------------------------------------------------------------------------- + + +class BenchmarkDataset(ABC): + DEFAULT_SEED = 0 + IS_MULTIMODAL = False + + def __init__( + self, + dataset_path: Optional[str] = None, + random_seed: int = DEFAULT_SEED, + ) -> None: + """ + Initialize the BenchmarkDataset with an optional dataset path and random + seed. Args: + dataset_path (Optional[str]): Path to the dataset. If None, it + indicates that a default or random dataset might be used. + random_seed (int): Seed value for reproducible shuffling or + sampling. Defaults to DEFAULT_SEED. + """ + self.dataset_path = dataset_path + # Set the random seed, ensuring that a None value is replaced with the + # default seed. + self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED + self.data = None + + def apply_multimodal_chat_transformation( + self, prompt: str, mm_content: Optional[MultiModalDataDict] = None + ) -> list[dict]: + """ + Transform a prompt and optional multimodal content into a chat format. + This method is used for chat models that expect a specific conversation + format. + """ + content = [{"text": prompt, "type": "text"}] + if mm_content is not None: + content.append(mm_content) + return [{"role": "user", "content": content}] + + def load_data(self) -> None: + """ + Load data from the dataset path into self.data. + + This method must be overridden by subclasses since the method to load + data will vary depending on the dataset format and source. + + Raises: + NotImplementedError: If a subclass does not implement this method. + """ + # TODO (jenniferzhao): add support for downloading data + raise NotImplementedError("load_data must be implemented in subclasses.") + + def get_random_lora_request( + self, + tokenizer: PreTrainedTokenizerBase, + max_loras: Optional[int] = None, + lora_path: Optional[str] = None, + ) -> tuple[Optional[LoRARequest], AnyTokenizer]: + """ + Optionally select a random LoRA request and return its associated + tokenizer. + + This method is used when LoRA parameters are provided. It randomly + selects a LoRA based on max_loras and retrieves a cached tokenizer for + that LoRA if available. Otherwise, it returns the base tokenizer. + + Args: + tokenizer (PreTrainedTokenizerBase): The base tokenizer to use if no + LoRA is selected. max_loras (Optional[int]): The maximum number of + LoRAs available. If None, LoRA is not used. lora_path + (Optional[str]): Path to the LoRA parameters on disk. If None, LoRA + is not used. + + Returns: + tuple[Optional[LoRARequest], AnyTokenizer]: A tuple where the first + element is a LoRARequest (or None if not applicable) and the second + element is the tokenizer associated with the LoRA request (or the + base tokenizer). + """ + if max_loras is None or lora_path is None: + return None, tokenizer + + # Generate a random LoRA ID in the range [1, max_loras]. + lora_id = random.randint(1, max_loras) + lora_request = LoRARequest( + lora_name=str(lora_id), + lora_int_id=lora_id, + lora_path=lora_path_on_disk(lora_path), + ) + if lora_id not in lora_tokenizer_cache: + lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request) + # Return lora_request and the cached tokenizer if available; otherwise, + # return the base tokenizer + return lora_request, lora_tokenizer_cache[lora_id] or tokenizer + + @abstractmethod + def sample( + self, tokenizer: PreTrainedTokenizerBase, num_requests: int + ) -> list[SampleRequest]: + """ + Abstract method to generate sample requests from the dataset. + + Subclasses must override this method to implement dataset-specific logic + for generating a list of SampleRequest objects. + + Args: + tokenizer (PreTrainedTokenizerBase): The tokenizer to be used + for processing the dataset's text. + num_requests (int): The number of sample requests to generate. + + Returns: + list[SampleRequest]: A list of sample requests generated from the + dataset. + """ + raise NotImplementedError("sample must be implemented in subclasses.") + + def maybe_oversample_requests( + self, requests: list[SampleRequest], num_requests: int + ) -> None: + """ + Oversamples the list of requests if its size is less than the desired + number. + + Args: + requests (List[SampleRequest]): The current list of sampled + requests. num_requests (int): The target number of requests. + """ + if len(requests) < num_requests: + random.seed(self.random_seed) + additional = random.choices(requests, k=num_requests - len(requests)) + requests.extend(additional) + logger.info("Oversampled requests to reach %d total samples.", num_requests) + + +# ----------------------------------------------------------------------------- +# Utility Functions and Global Caches +# ----------------------------------------------------------------------------- + + +def is_valid_sequence( + prompt_len: int, + output_len: int, + min_len: int = 4, + max_prompt_len: int = 1024, + max_total_len: int = 2048, + skip_min_output_len_check: bool = False, +) -> bool: + """ + Validate a sequence based on prompt and output lengths. + + Default pruning criteria are copied from the original `sample_hf_requests` + and `sample_sharegpt_requests` functions in benchmark_serving.py, as well as + from `sample_requests` in benchmark_throughput.py. + """ + # Check for invalid conditions + prompt_too_short = prompt_len < min_len + output_too_short = (not skip_min_output_len_check) and (output_len < min_len) + prompt_too_long = prompt_len > max_prompt_len + combined_too_long = (prompt_len + output_len) > max_total_len + + # Return True if none of the invalid conditions are met + return not ( + prompt_too_short or output_too_short or prompt_too_long or combined_too_long + ) + + +@cache +def lora_path_on_disk(lora_path: str) -> str: + return get_adapter_absolute_path(lora_path) + + +# Global cache for LoRA tokenizers. +lora_tokenizer_cache: dict[int, AnyTokenizer] = {} + + +def process_image(image: Any) -> Mapping[str, Any]: + """ + Process a single image input and return a multimedia content dictionary. + + Supports three input types: + + 1. Dictionary with raw image bytes: - Expects a dict with a 'bytes' key + containing raw image data. - Loads the bytes as a PIL.Image.Image. + + 2. PIL.Image.Image input: - Converts the image to RGB. - Saves the image as + a JPEG in memory. - Encodes the JPEG data as a base64 string. - Returns + a dictionary with the image as a base64 data URL. + + 3. String input: - Treats the string as a URL or local file path. - + Prepends "file://" if the string doesn't start with "http://" or + "file://". - Returns a dictionary with the image URL. + + Raises: + ValueError: If the input is not a supported type. + """ + if isinstance(image, dict) and "bytes" in image: + image = Image.open(BytesIO(image["bytes"])) + if isinstance(image, Image.Image): + image = convert_image_mode(image, "RGB") + with io.BytesIO() as image_data: + image.save(image_data, format="JPEG") + image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8") + return { + "type": "image_url", + "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"}, + } + + if isinstance(image, str): + image_url = ( + image if image.startswith(("http://", "file://")) else f"file://{image}" + ) + return {"type": "image_url", "image_url": {"url": image_url}} + + raise ValueError( + f"Invalid image input {image}. Must be a PIL.Image.Image" + " or str or dictionary with raw image bytes." + ) + + +# ----------------------------------------------------------------------------- +# Random Dataset Implementation (Synthetic Data) +# ----------------------------------------------------------------------------- + + +class RandomDataset(BenchmarkDataset): + # Default values copied from benchmark_serving.py for the random dataset. + DEFAULT_PREFIX_LEN = 0 + DEFAULT_RANGE_RATIO = 0.0 + DEFAULT_INPUT_LEN = 1024 + DEFAULT_OUTPUT_LEN = 128 + + def __init__( + self, + **kwargs, + ) -> None: + super().__init__(**kwargs) + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + prefix_len: int = DEFAULT_PREFIX_LEN, + range_ratio: float = DEFAULT_RANGE_RATIO, + input_len: int = DEFAULT_INPUT_LEN, + output_len: int = DEFAULT_OUTPUT_LEN, + **kwargs, + ) -> list[SampleRequest]: + # Enforce range_ratio < 1 + assert range_ratio < 1.0, ( + "random_range_ratio must be < 1.0 to ensure a valid sampling range" + ) + + vocab_size = tokenizer.vocab_size + num_special_tokens = tokenizer.num_special_tokens_to_add() + real_input_len = input_len - num_special_tokens + + prefix_token_ids = ( + np.random.randint(0, vocab_size, size=prefix_len).tolist() + if prefix_len > 0 + else [] + ) + + # New sampling logic: [X * (1 - b), X * (1 + b)] + input_low = int(real_input_len * (1 - range_ratio)) + input_high = int(real_input_len * (1 + range_ratio)) + output_low = int(output_len * (1 - range_ratio)) + output_high = int(output_len * (1 + range_ratio)) + + # Add logging for debugging + logger.info("Sampling input_len from [%s, %s]", input_low, input_high) + logger.info("Sampling output_len from [%s, %s]", output_low, output_high) + + input_lens = np.random.randint(input_low, input_high + 1, size=num_requests) + output_lens = np.random.randint(output_low, output_high + 1, size=num_requests) + offsets = np.random.randint(0, vocab_size, size=num_requests) + + requests = [] + for i in range(num_requests): + inner_seq = ( + (offsets[i] + i + np.arange(input_lens[i])) % vocab_size + ).tolist() + token_sequence = prefix_token_ids + inner_seq + prompt = tokenizer.decode(token_sequence) + # After decoding the prompt we have to encode and decode it again. + # This is done because in some cases N consecutive tokens + # give a string tokenized into != N number of tokens. + # For example for GPT2Tokenizer: + # [6880, 6881] -> ['Ġcalls', 'here'] -> + # [1650, 939, 486] -> ['Ġcall', 'sh', 'ere'] + # To avoid uncontrolled change of the prompt length, + # the encoded sequence is truncated before being decode again. + total_input_len = prefix_len + int(input_lens[i]) + re_encoded_sequence = tokenizer.encode(prompt, add_special_tokens=False)[ + :total_input_len + ] + prompt = tokenizer.decode(re_encoded_sequence) + total_input_len = len(re_encoded_sequence) + requests.append( + SampleRequest( + prompt=prompt, + prompt_len=total_input_len, + expected_output_len=int(output_lens[i]), + ) + ) + return requests + + +# ----------------------------------------------------------------------------- +# ShareGPT Dataset Implementation +# ----------------------------------------------------------------------------- + + +class ShareGPTDataset(BenchmarkDataset): + """ + Implements the ShareGPT dataset. Loads data from a JSON file and generates + sample requests based on conversation turns. + """ + + def __init__(self, **kwargs) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data(self) -> None: + if self.dataset_path is None: + raise ValueError("dataset_path must be provided for loading data.") + + with open(self.dataset_path, encoding="utf-8") as f: + self.data = json.load(f) + # Filter entries with at least two conversation turns. + self.data = [ + entry + for entry in self.data + if "conversations" in entry and len(entry["conversations"]) >= 2 + ] + random.seed(self.random_seed) + random.shuffle(self.data) + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + lora_path: Optional[str] = None, + max_loras: Optional[int] = None, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + samples: list = [] + for entry in self.data: + if len(samples) >= num_requests: + break + prompt, completion = ( + entry["conversations"][0]["value"], + entry["conversations"][1]["value"], + ) + + lora_request, tokenizer = self.get_random_lora_request( + tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path + ) + prompt_ids = tokenizer(prompt).input_ids + completion_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_ids) + new_output_len = len(completion_ids) if output_len is None else output_len + if not is_valid_sequence( + prompt_len, + new_output_len, + skip_min_output_len_check=output_len is not None, + ): + continue + if enable_multimodal_chat: + prompt = self.apply_multimodal_chat_transformation(prompt, None) + samples.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=new_output_len, + lora_request=lora_request, + ) + ) + self.maybe_oversample_requests(samples, num_requests) + return samples + + +# ----------------------------------------------------------------------------- +# Custom Dataset Implementation +# ----------------------------------------------------------------------------- + + +class CustomDataset(BenchmarkDataset): + """ + Implements the Custom dataset. Loads data from a JSONL file and generates + sample requests based on conversation turns. E.g., + ``` + {"prompt": "What is the capital of India?"} + {"prompt": "What is the capital of Iran?"} + {"prompt": "What is the capital of China?"} + ``` + """ + + def __init__(self, **kwargs) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data(self) -> None: + if self.dataset_path is None: + raise ValueError("dataset_path must be provided for loading data.") + + # self.data will be a list of dictionaries + # e.g., [{"prompt": "What is the capital of India?"}, ...] + # This will be the standardized format which load_data() + # has to convert into depending on the filetype of dataset_path. + # sample() will assume this standardized format of self.data + self.data = [] + + # Load the JSONL file + if self.dataset_path.endswith(".jsonl"): + jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True) + + # check if the JSONL file has a 'prompt' column + if "prompt" not in jsonl_data.columns: + raise ValueError("JSONL file must contain a 'prompt' column.") + + # Convert each row to a dictionary and append to self.data + # This will convert the DataFrame to a list of dictionaries + # where each dictionary corresponds to a row in the DataFrame. + # This is the standardized format we want for self.data + for _, row in jsonl_data.iterrows(): + self.data.append(row.to_dict()) + else: + raise NotImplementedError( + "Only JSONL format is supported for CustomDataset." + ) + + random.seed(self.random_seed) + random.shuffle(self.data) + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + lora_path: Optional[str] = None, + max_loras: Optional[int] = None, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + skip_chat_template: bool = False, + **kwargs, + ) -> list: + sampled_requests = [] + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt = item["prompt"] + + # apply template + if not skip_chat_template: + prompt = tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=True, + tokenize=False, + ) + + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Sonnet Dataset Implementation +# ----------------------------------------------------------------------------- + + +class SonnetDataset(BenchmarkDataset): + """ + Simplified implementation of the Sonnet dataset. Loads poem lines from a + text file and generates sample requests. Default values here copied from + `benchmark_serving.py` for the sonnet dataset. + """ + + DEFAULT_PREFIX_LEN = 200 + DEFAULT_INPUT_LEN = 550 + DEFAULT_OUTPUT_LEN = 150 + + def __init__( + self, + **kwargs, + ) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data(self) -> None: + if not self.dataset_path: + raise ValueError("dataset_path must be provided.") + with open(self.dataset_path, encoding="utf-8") as f: + self.data = f.readlines() + + def sample( + self, + tokenizer, + num_requests: int, + prefix_len: int = DEFAULT_PREFIX_LEN, + input_len: int = DEFAULT_INPUT_LEN, + output_len: int = DEFAULT_OUTPUT_LEN, + return_prompt_formatted: bool = False, + **kwargs, + ) -> list: + # Calculate average token length for a poem line. + tokenized_lines = [tokenizer(line).input_ids for line in self.data] + avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines) + + # Build the base prompt. + base_prompt = "Pick as many lines as you can from these poem lines:\n" + base_msg = [{"role": "user", "content": base_prompt}] + base_fmt = tokenizer.apply_chat_template( + base_msg, add_generation_prompt=True, tokenize=False + ) + base_offset = len(tokenizer(base_fmt).input_ids) + if input_len <= base_offset: + raise ValueError( + f"'input_len' must be higher than the base prompt length " + f"({base_offset})." + ) + + # Determine how many poem lines to use. + num_input_lines = round((input_len - base_offset) / avg_len) + num_prefix_lines = max(round((prefix_len - base_offset) / avg_len), 0) + prefix_lines = self.data[:num_prefix_lines] + + samples = [] + while len(samples) < num_requests: + extra_lines = random.choices( + self.data, k=num_input_lines - num_prefix_lines + ) + prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}" + msg = [{"role": "user", "content": prompt}] + prompt_formatted = tokenizer.apply_chat_template( + msg, add_generation_prompt=True, tokenize=False + ) + prompt_len = len(tokenizer(prompt_formatted).input_ids) + if prompt_len <= input_len: + samples.append( + SampleRequest( + prompt=prompt_formatted if return_prompt_formatted else prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + return samples + + +# ----------------------------------------------------------------------------- +# BurstGPT Dataset Implementation +# ----------------------------------------------------------------------------- + + +class BurstGPTDataset(BenchmarkDataset): + """ + Implements the BurstGPT dataset. Loads data from a CSV file and generates + sample requests based on synthetic prompt generation. Only rows with Model + "GPT-4" and positive response tokens are used. + """ + + def __init__(self, **kwargs) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data( + self, + ): + if self.dataset_path is None: + raise ValueError("dataset_path must be provided for loading data.") + + df = pd.read_csv(self.dataset_path) + # Filter to keep only GPT-4 rows. + gpt4_df = df[df["Model"] == "GPT-4"] + # Remove failed requests (where Response tokens is 0 or less). + gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0] + # Sample the desired number of rows. + self.data = gpt4_df + + def _sample_loaded_data(self, num_requests: int) -> list: + if num_requests <= len(self.data): + data = self.data.sample(n=num_requests, random_state=self.random_seed) + else: + data = self.data.sample( + n=num_requests, + random_state=self.random_seed, + replace=True, + ) + # Convert the dataframe to a list of lists. + return data.values.tolist() + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + max_loras: Optional[int] = None, + lora_path: Optional[str] = None, + **kwargs, + ) -> list[SampleRequest]: + samples = [] + data = self._sample_loaded_data(num_requests=num_requests) + for i in range(num_requests): + input_len = int(data[i][2]) + output_len = int(data[i][3]) + lora_req, tokenizer = self.get_random_lora_request( + tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path + ) + vocab_size = tokenizer.vocab_size + # Generate a synthetic prompt: a list of token IDs computed as (i + + # j) modulo vocab_size. + token_ids = [(i + j) % vocab_size for j in range(input_len)] + prompt = tokenizer.decode(token_ids) + samples.append( + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=output_len, + lora_request=lora_req, + ) + ) + return samples + + +# ----------------------------------------------------------------------------- +# HuggingFace Dataset Base Implementation +# ----------------------------------------------------------------------------- +class HuggingFaceDataset(BenchmarkDataset): + """Base class for datasets hosted on HuggingFace.""" + + SUPPORTED_DATASET_PATHS: Union[set[str], dict[str, Callable]] = set() + + def __init__( + self, + dataset_path: str, + dataset_split: str, + dataset_subset: Optional[str] = None, + **kwargs, + ) -> None: + super().__init__(dataset_path=dataset_path, **kwargs) + + self.dataset_split = dataset_split + self.dataset_subset = dataset_subset + self.load_data() + + def load_data(self) -> None: + """Load data from HuggingFace datasets.""" + self.data = load_dataset( + self.dataset_path, + name=self.dataset_subset, + split=self.dataset_split, + streaming=True, + ) + self.data = self.data.shuffle(seed=self.random_seed) + + +# ----------------------------------------------------------------------------- +# Conversation Dataset Implementation +# ----------------------------------------------------------------------------- + + +class ConversationDataset(HuggingFaceDataset): + """Dataset for conversation data with multimodal support.""" + + SUPPORTED_DATASET_PATHS = { + "lmms-lab/LLaVA-OneVision-Data", + "Aeala/ShareGPT_Vicuna_unfiltered", + } + IS_MULTIMODAL = True + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + # Filter examples with at least 2 conversations + filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2) + sampled_requests = [] + dynamic_output = output_len is None + + for item in filtered_data: + if len(sampled_requests) >= num_requests: + break + conv = item["conversations"] + prompt, completion = conv[0]["value"], conv[1]["value"] + + prompt_ids = tokenizer(prompt).input_ids + completion_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_ids) + completion_len = len(completion_ids) + output_len = completion_len if dynamic_output else output_len + assert isinstance(output_len, int) and output_len > 0 + if dynamic_output and not is_valid_sequence(prompt_len, completion_len): + continue + mm_content = process_image(item["image"]) if "image" in item else None + if enable_multimodal_chat: + # Note: when chat is enabled the request prompt_len is no longer + # accurate and we will be using request output to count the + # actual prompt len and output len + prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=mm_content, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Vision Arena Dataset Implementation +# ----------------------------------------------------------------------------- + + +class VisionArenaDataset(HuggingFaceDataset): + """ + Vision Arena Dataset. + """ + + DEFAULT_OUTPUT_LEN = 128 + SUPPORTED_DATASET_PATHS = { + "lmarena-ai/VisionArena-Chat": lambda x: x["conversation"][0][0]["content"], + "lmarena-ai/vision-arena-bench-v0.1": lambda x: x["turns"][0][0]["content"], + } + IS_MULTIMODAL = True + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + sampled_requests = [] + for item in self.data: + if len(sampled_requests) >= num_requests: + break + parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.dataset_path) + if parser_fn is None: + raise ValueError(f"Unsupported dataset path: {self.dataset_path}") + prompt = parser_fn(item) + mm_content = process_image(item["images"][0]) + prompt_len = len(tokenizer(prompt).input_ids) + if enable_multimodal_chat: + # Note: when chat is enabled the request prompt_len is no longer + # accurate and we will be using request output to count the + # actual prompt len + prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=mm_content, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Instruct Coder Dataset Implementation +# ----------------------------------------------------------------------------- + + +class InstructCoderDataset(HuggingFaceDataset): + """ + InstructCoder Dataset. + https://huggingface.co/datasets/likaixin/InstructCoder + + InstructCoder is the dataset designed for general code editing. It consists + of 114,239 instruction-input-output triplets, and covers multiple distinct + code editing scenario. + """ + + DEFAULT_OUTPUT_LEN = 200 # this is the average default output length + SUPPORTED_DATASET_PATHS = { + "likaixin/InstructCoder", + } + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + sampled_requests = [] + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt = f"{item['input']}\n\n{item['instruction']} Just output \ + the code, do not include any explanation." + + # apply template + prompt = tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=True, + tokenize=False, + ) + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# MT-Bench Dataset Implementation +# ----------------------------------------------------------------------------- + + +class MTBenchDataset(HuggingFaceDataset): + """ + MT-Bench Dataset. + https://huggingface.co/datasets/philschmid/mt-bench + + We create a single turn dataset for MT-Bench. + This is similar to Spec decoding benchmark setup in vLLM + https://github.com/vllm-project/vllm/blob/9d98ab5ec/examples/offline_inference/eagle.py#L14-L18 + """ # noqa: E501 + + DEFAULT_OUTPUT_LEN = 256 # avg len used in SD bench in vLLM + SUPPORTED_DATASET_PATHS = { + "philschmid/mt-bench", + } + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + sampled_requests = [] + + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt = item["turns"][0] + + # apply template + prompt = tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=True, + tokenize=False, + ) + + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# AIMO Dataset Implementation +# ----------------------------------------------------------------------------- + + +class AIMODataset(HuggingFaceDataset): + """ + Dataset class for processing a AIMO dataset with reasoning questions. + """ + + SUPPORTED_DATASET_PATHS = { + "AI-MO/aimo-validation-aime", + "AI-MO/NuminaMath-1.5", + "AI-MO/NuminaMath-CoT", + } + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + **kwargs, + ) -> list: + sampled_requests = [] + dynamic_output = output_len is None + + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt, completion = item["problem"], item["solution"] + + prompt_ids = tokenizer(prompt).input_ids + completion_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_ids) + completion_len = len(completion_ids) + output_len = completion_len if dynamic_output else output_len + assert isinstance(output_len, int) and output_len > 0 + if dynamic_output and not is_valid_sequence( + prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000 + ): + continue + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=None, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Next Edit Prediction Dataset Implementation +# ----------------------------------------------------------------------------- + + +zeta_prompt = """### Instruction: +You are a code completion assistant and your task is to analyze user edits and then rewrite an excerpt that the user provides, suggesting the appropriate edits within the excerpt, taking into account the cursor location. + +### User Edits: + +{} + +### User Excerpt: + +{} + +### Response: + +""" # noqa: E501 + + +def _format_zeta_prompt( + sample: dict, original_start_marker: str = "<|editable_region_start|>" +) -> dict: + """Format the zeta prompt for the Next Edit Prediction (NEP) dataset. + + This function formats examples from the NEP dataset + into prompts and expected outputs. It could be + further extended to support more NEP datasets. + + Args: + sample: The dataset sample containing events, + inputs, and outputs. + original_start_marker: The marker indicating the + start of the editable region. Defaults to + "<|editable_region_start|>". + + Returns: + A dictionary with the formatted prompts and expected outputs. + """ + events = sample["events"] + input = sample["input"] + output = sample["output"] + prompt = zeta_prompt.format(events, input) + + # following the original implementation, extract the focused region + # from the raw output + output_start_index = output.find(original_start_marker) + output_focused_region = output[output_start_index:] + expected_output = output_focused_region + + return {"prompt": prompt, "expected_output": expected_output} + + +class NextEditPredictionDataset(HuggingFaceDataset): + """ + Dataset class for processing a Next Edit Prediction dataset. + """ + + SUPPORTED_DATASET_PATHS = { + "zed-industries/zeta", + } + MAPPING_PROMPT_FUNCS = { + "zed-industries/zeta": _format_zeta_prompt, + } + + def sample(self, tokenizer: PreTrainedTokenizerBase, num_requests: int, **kwargs): + formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.dataset_path) + if formatting_prompt_func is None: + raise ValueError(f"Unsupported dataset path: {self.dataset_path}") + samples = [] + for sample in self.data: + sample = formatting_prompt_func(sample) + samples.append( + SampleRequest( + prompt=sample["prompt"], + prompt_len=len(tokenizer(sample["prompt"]).input_ids), + expected_output_len=len( + tokenizer(sample["expected_output"]).input_ids + ), + ) + ) + if len(samples) >= num_requests: + break + self.maybe_oversample_requests(samples, num_requests) + return samples + + +# ----------------------------------------------------------------------------- +# ASR Dataset Implementation +# ----------------------------------------------------------------------------- + + +class ASRDataset(HuggingFaceDataset): + """ + Dataset class for processing a ASR dataset for transcription. + Tested on the following set: + + +----------------+----------------------------------------+--------------------------+-----------------------------+ + | Dataset | Domain | Speaking Style | hf-subset | + +----------------+----------------------------------------+--------------------------+-----------------------------+ + | TED-LIUM | TED talks | Oratory | release1, release2, release3| + | | | | release3-speaker-adaptation | + | VoxPopuli | European Parliament | Oratory | en, de, it, fr, ... | + | LibriSpeech | Audiobook | Narrated | "LIUM/tedlium" | + | GigaSpeech | Audiobook, podcast, YouTube | Narrated, spontaneous | xs, s, m, l, xl, dev, test | + | SPGISpeech | Financial meetings | Oratory, spontaneous | S, M, L, dev, test | + | AMI | Meetings | Spontaneous | ihm, sdm | + +----------------+----------------------------------------+--------------------------+-----------------------------+ + + """ # noqa: E501 + + SUPPORTED_DATASET_PATHS = { + "openslr/librispeech_asr", + "facebook/voxpopuli", + "LIUM/tedlium", + "edinburghcstr/ami", + "speechcolab/gigaspeech", + "kensho/spgispeech", + } + + DEFAULT_OUTPUT_LEN = 128 + IS_MULTIMODAL = True + + # TODO Whisper-specific. Abstract interface when more models are supported. + TRANSCRIPTION_PREAMBLE = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>" + skip_long_audios: bool = True + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + **kwargs, + ) -> list: + import librosa + + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + prompt = ASRDataset.TRANSCRIPTION_PREAMBLE + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests = [] + skipped = 0 + for item in self.data: + if len(sampled_requests) >= num_requests: + break + audio = item["audio"] + y, sr = audio["array"], audio["sampling_rate"] + duration_s = librosa.get_duration(y=y, sr=sr) + # Whisper max supported duration + if self.skip_long_audios and duration_s > 30: + skipped += 1 + continue + + mm_content = {"audio": (y, sr)} + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=mm_content, + ) + ) + if skipped: + logger.warning( + "%d samples discarded from dataset due to" + " their length being greater than" + " what Whisper supports.", + skipped, + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests diff --git a/offline_benchmark_test/benchmarks/benchmark_latency.py b/offline_benchmark_test/benchmarks/benchmark_latency.py new file mode 100644 index 0000000000000000000000000000000000000000..4d2ea126b24a514bde2e7e3808d2120ede353cc7 --- /dev/null +++ b/offline_benchmark_test/benchmarks/benchmark_latency.py @@ -0,0 +1,186 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Benchmark the latency of processing a single batch of requests.""" + +import argparse +import dataclasses +import json +import os +import time +from typing import Any, Optional + +import numpy as np +from tqdm import tqdm + +import vllm.envs as envs +from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json +from vllm import LLM, SamplingParams +from vllm.engine.arg_utils import EngineArgs +from vllm.inputs import PromptType +from vllm.sampling_params import BeamSearchParams +from vllm.utils import FlexibleArgumentParser + + +def save_to_pytorch_benchmark_format( + args: argparse.Namespace, results: dict[str, Any] +) -> None: + pt_records = convert_to_pytorch_benchmark_format( + args=args, + metrics={"latency": results["latencies"]}, + extra_info={k: results[k] for k in ["avg_latency", "percentiles"]}, + ) + if pt_records: + pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json" + write_to_json(pt_file, pt_records) + + +def main(args: argparse.Namespace): + print(args) + + engine_args = EngineArgs.from_cli_args(args) + + # NOTE(woosuk): If the request cannot be processed in a single batch, + # the engine will automatically process the request in multiple batches. + llm = LLM(**dataclasses.asdict(engine_args)) + assert llm.llm_engine.model_config.max_model_len >= ( + args.input_len + args.output_len + ), ( + "Please ensure that max_model_len is greater than" + " the sum of input_len and output_len." + ) + + sampling_params = SamplingParams( + n=args.n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=args.output_len, + detokenize=not args.disable_detokenize, + ) + print(sampling_params) + dummy_prompt_token_ids = np.random.randint( + 10000, size=(args.batch_size, args.input_len) + ) + dummy_prompts: list[PromptType] = [ + {"prompt_token_ids": batch} for batch in dummy_prompt_token_ids.tolist() + ] + + def llm_generate(): + if not args.use_beam_search: + llm.generate(dummy_prompts, sampling_params=sampling_params, use_tqdm=False) + else: + llm.beam_search( + dummy_prompts, + BeamSearchParams( + beam_width=args.n, + max_tokens=args.output_len, + ignore_eos=True, + ), + ) + + def run_to_completion(profile_dir: Optional[str] = None): + if profile_dir: + llm.start_profile() + llm_generate() + llm.stop_profile() + else: + start_time = time.perf_counter() + llm_generate() + end_time = time.perf_counter() + latency = end_time - start_time + return latency + + print("Warming up...") + for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"): + run_to_completion(profile_dir=None) + + if args.profile: + profile_dir = envs.VLLM_TORCH_PROFILER_DIR + print(f"Profiling (results will be saved to '{profile_dir}')...") + run_to_completion(profile_dir=profile_dir) + return + + # Benchmark. + latencies = [] + for _ in tqdm(range(args.num_iters), desc="Profiling iterations"): + latencies.append(run_to_completion(profile_dir=None)) + latencies = np.array(latencies) + percentages = [10, 25, 50, 75, 90, 99] + percentiles = np.percentile(latencies, percentages) + print(f"Avg latency: {np.mean(latencies)} seconds") + for percentage, percentile in zip(percentages, percentiles): + print(f"{percentage}% percentile latency: {percentile} seconds") + + # Output JSON results if specified + if args.output_json: + results = { + "avg_latency": np.mean(latencies), + "latencies": latencies.tolist(), + "percentiles": dict(zip(percentages, percentiles.tolist())), + } + with open(args.output_json, "w") as f: + json.dump(results, f, indent=4) + save_to_pytorch_benchmark_format(args, results) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the latency of processing a single batch of " + "requests till completion." + ) + parser.add_argument("--input-len", type=int, default=32) + parser.add_argument("--output-len", type=int, default=128) + parser.add_argument("--batch-size", type=int, default=8) + parser.add_argument( + "--n", + type=int, + default=1, + help="Number of generated sequences per prompt.", + ) + parser.add_argument("--use-beam-search", action="store_true") + parser.add_argument( + "--num-iters-warmup", + type=int, + default=10, + help="Number of iterations to run for warmup.", + ) + parser.add_argument( + "--num-iters", type=int, default=30, help="Number of iterations to run." + ) + parser.add_argument( + "--profile", + action="store_true", + help="profile the generation process of a single batch", + ) + parser.add_argument( + "--output-json", + type=str, + default=None, + help="Path to save the latency results in JSON format.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize responses (i.e. do not include " + "detokenization time in the latency measurement)" + ), + ) + + parser = EngineArgs.add_cli_args(parser) + # V1 enables prefix caching by default which skews the latency + # numbers. We need to disable prefix caching by default. + parser.set_defaults(enable_prefix_caching=False) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + if args.profile and not envs.VLLM_TORCH_PROFILER_DIR: + raise OSError( + "The environment variable 'VLLM_TORCH_PROFILER_DIR' is not set. " + "Please set it to a valid path to use torch profiler." + ) + main(args) diff --git a/offline_benchmark_test/benchmarks/benchmark_long_document_qa_throughput.py b/offline_benchmark_test/benchmarks/benchmark_long_document_qa_throughput.py new file mode 100644 index 0000000000000000000000000000000000000000..6e0f3b51c9d284d22dd1acdc1b794b9a5c0f31b1 --- /dev/null +++ b/offline_benchmark_test/benchmarks/benchmark_long_document_qa_throughput.py @@ -0,0 +1,202 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +Offline benchmark to test the long document QA throughput. + +Example usage: + # This workload samples 8 different prompts with a default input + # length of 20000 tokens, then replicates each prompt 2 times + # in random order. + python benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --repeat-count 2 + +Commandline arguments: + --num-documents: The number of documents to sample prompts from. + + --document-length: The length of each document in tokens. + (Optional, default: 20000) + + --output-len: The number of tokens to generate for each prompt. + (Optional, default: 10) + + --repeat-count: The number of times to repeat each prompt. + (Optional, default: 2) + + --repeat-mode: The mode to repeat prompts. The supported modes are: + - 'random': shuffle the prompts randomly. (Default) + - 'tile': the entire prompt list is repeated in sequence. (Potentially + lowest cache hit) + - 'interleave': each prompt is repeated consecutively before + moving to the next element. (Highest cache hit) + + --shuffle-seed: Random seed when the repeat mode is "random". + (Optional, default: 0) + +In the meantime, it also supports all the vLLM engine args to initialize the +LLM engine. You can refer to the `vllm.engine.arg_utils.EngineArgs` for more +details. +""" + +import dataclasses +import random +import time + +from vllm import LLM, SamplingParams +from vllm.engine.arg_utils import EngineArgs +from vllm.utils import FlexibleArgumentParser + + +def test_long_document_qa(llm=None, sampling_params=None, prompts=None): + """ + Test long document QA with the given prompts and sampling parameters. + Print the time spent in processing all the prompts. + + Args: + llm: The language model used for generating responses. + sampling_params: Sampling parameter used to generate the response. + prompts: A list of prompt strings to be processed by the LLM. + """ + start_time = time.time() + llm.generate(prompts, sampling_params=sampling_params) + end_time = time.time() + print(f"Time to execute all requests: {end_time - start_time:.4f} secs") + + +def repeat_prompts(prompts, repeat_count, mode: str): + """ + Repeat each prompt in the list for a specified number of times. + The order of prompts in the output list depends on the mode. + + Args: + prompts: A list of prompts to be repeated. + repeat_count: The number of times each prompt is repeated. + mode: The mode of repetition. Supported modes are: + - 'random': Shuffle the prompts randomly after repetition. + - 'tile': Repeat the entire prompt list in sequence. + Example: [1, 2, 3] -> [1, 2, 3, 1, 2, 3]. + - 'interleave': Repeat each prompt consecutively before moving to + the next. Example: [1, 2, 3] -> [1, 1, 2, 2, 3, 3]. + + Returns: + A list of repeated prompts in the specified order. + + Raises: + ValueError: If an invalid mode is provided. + """ + print("Repeat mode: ", mode) + if mode == "random": + repeated_prompts = prompts * repeat_count + random.shuffle(repeated_prompts) + return repeated_prompts + elif mode == "tile": + return prompts * repeat_count + elif mode == "interleave": + repeated_prompts = [] + for prompt in prompts: + repeated_prompts.extend([prompt] * repeat_count) + return repeated_prompts + else: + raise ValueError( + f"Invalid mode: {mode}, only support 'random', 'tile', 'interleave'" + ) + + +def main(args): + random.seed(args.shuffle_seed) + + # Prepare the prompts: + # we append the document id at the beginning to avoid any of the document + # being the prefix of other documents + prompts = [ + str(i) + " ".join(["hi"] * args.document_length) + for i in range(args.num_documents) + ] + + prompts = repeat_prompts(prompts, args.repeat_count, mode=args.repeat_mode) + + warmup_prompts = [ + "This is warm up request " + str(i) + " ".join(["hi"] * args.document_length) + for i in range(args.num_documents) + ] + + # Create the LLM engine + engine_args = EngineArgs.from_cli_args(args) + llm = LLM(**dataclasses.asdict(engine_args)) + sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len) + + print("------warm up------") + test_long_document_qa( + llm=llm, + prompts=warmup_prompts, + sampling_params=sampling_params, + ) + + print("------start generating------") + test_long_document_qa( + llm=llm, + prompts=prompts, + sampling_params=sampling_params, + ) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the performance with or " + "without automatic prefix caching." + ) + + parser.add_argument( + "--document-length", + type=int, + # Roughly the number of tokens for a system paper, + # excluding images + default=20000, + help="Range of input lengths for sampling prompts, " + 'specified as "min:max" (e.g., "128:256").', + ) + + parser.add_argument( + "--num-documents", + type=int, + default=8, + help="Range of input lengths for sampling prompts, " + 'specified as "min:max" (e.g., "128:256").', + ) + + parser.add_argument("--output-len", type=int, default=10) + + parser.add_argument( + "--repeat-count", + type=int, + default=2, + help="Number of times to repeat each prompt", + ) + + parser.add_argument( + "--repeat-mode", + type=str, + default="random", + help="The mode to repeat prompts. The supported " + 'modes are "random", "tile", and "interleave". ' + "See repeat_prompts() in the source code for details.", + ) + + parser.add_argument( + "--shuffle-seed", + type=int, + default=0, + help='Random seed when the repeat mode is "random"', + ) + + parser = EngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/offline_benchmark_test/benchmarks/benchmark_prefix_caching.py b/offline_benchmark_test/benchmarks/benchmark_prefix_caching.py new file mode 100644 index 0000000000000000000000000000000000000000..b5e2613de1cd4a0a2557cf09baa458aa15a6049c --- /dev/null +++ b/offline_benchmark_test/benchmarks/benchmark_prefix_caching.py @@ -0,0 +1,278 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +Benchmark the efficiency of prefix caching. + +This script allows you to benchmark the performance of +a model with and without prefix caching using either fixed prompts +or prompts sampled from the ShareGPT dataset. + +Fixed example usage: + python benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-prompts 1 \ + --repeat-count 100 \ + --input-length-range 128:256 + +ShareGPT example usage: + # This command samples 20 prompts with input lengths + # between 128 and 256 tokens from the ShareGPT dataset, + # then replicates each prompt 5 times. + python benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json \ + --enable-prefix-caching \ + --num-prompts 20 \ + --repeat-count 5 \ + --input-length-range 128:256 +""" + +import dataclasses +import json +import random +import time +from typing import Optional + +from transformers import PreTrainedTokenizerBase + +from vllm import LLM, SamplingParams +from vllm.engine.arg_utils import EngineArgs +from vllm.utils import FlexibleArgumentParser + +try: + from vllm.transformers_utils.tokenizer import get_tokenizer +except ImportError: + from backend_request_func import get_tokenizer + +PROMPT = "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as fellows. You need to answer my question about the table.\n# Table\n|Opening|Opening|Sl. No.|Film|Cast|Director|Music Director|Notes|\n|----|----|----|----|----|----|----|----|\n|J A N|9|1|Agni Pushpam|Jayabharathi, Kamalahasan|Jeassy|M. K. Arjunan||\n|J A N|16|2|Priyamvada|Mohan Sharma, Lakshmi, KPAC Lalitha|K. S. Sethumadhavan|V. Dakshinamoorthy||\n|J A N|23|3|Yakshagaanam|Madhu, Sheela|Sheela|M. S. Viswanathan||\n|J A N|30|4|Paalkkadal|Sheela, Sharada|T. K. Prasad|A. T. Ummer||\n|F E B|5|5|Amma|Madhu, Srividya|M. Krishnan Nair|M. K. Arjunan||\n|F E B|13|6|Appooppan|Thikkurissi Sukumaran Nair, Kamal Haasan|P. Bhaskaran|M. S. Baburaj||\n|F E B|20|7|Srishti|Chowalloor Krishnankutty, Ravi Alummoodu|K. T. Muhammad|M. S. Baburaj||\n|F E B|20|8|Vanadevatha|Prem Nazir, Madhubala|Yusufali Kechery|G. Devarajan||\n|F E B|27|9|Samasya|Madhu, Kamalahaasan|K. Thankappan|Shyam||\n|F E B|27|10|Yudhabhoomi|K. P. Ummer, Vidhubala|Crossbelt Mani|R. K. Shekhar||\n|M A R|5|11|Seemantha Puthran|Prem Nazir, Jayabharathi|A. B. Raj|M. K. Arjunan||\n|M A R|12|12|Swapnadanam|Rani Chandra, Dr. Mohandas|K. G. George|Bhaskar Chandavarkar||\n|M A R|19|13|Thulavarsham|Prem Nazir, sreedevi, Sudheer|N. Sankaran Nair|V. Dakshinamoorthy||\n|M A R|20|14|Aruthu|Kaviyoor Ponnamma, Kamalahasan|Ravi|G. Devarajan||\n|M A R|26|15|Swimming Pool|Kamal Haasan, M. G. Soman|J. Sasikumar|M. K. Arjunan||\n\n# Question\nWhat' s the content in the (1,1) cells\n" # noqa: E501 + + +def test_prefix(llm=None, sampling_params=None, prompts=None): + start_time = time.time() + + llm.generate(prompts, sampling_params=sampling_params) + + end_time = time.time() + print(f"cost time {end_time - start_time}") + + +@dataclasses.dataclass +class Request: + prompt: str + prompt_len: int + output_len: int + + +def sample_tokens(tokenizer: PreTrainedTokenizerBase, length: int) -> list[int]: + vocab = tokenizer.get_vocab() + all_special_ids = set(tokenizer.all_special_ids) + + # Remove the special tokens. + return random.choices( + [v for k, v in vocab.items() if k not in all_special_ids], + k=length, + ) + + +def sample_requests_from_dataset( + dataset_path: str, + num_requests: int, + tokenizer: PreTrainedTokenizerBase, + input_length_range: tuple[int, int], + fixed_output_len: Optional[int], +) -> list[Request]: + if fixed_output_len is not None and fixed_output_len < 4: + raise ValueError("output_len too small") + + # Load the dataset. + with open(dataset_path) as f: + dataset = json.load(f) + # Filter out the conversations with less than 2 turns. + dataset = [data for data in dataset if len(data["conversations"]) >= 2] + # Only keep the first two turns of each conversation. + dataset = [ + (data["conversations"][0]["value"], data["conversations"][1]["value"]) + for data in dataset + ] + + # Shuffle the dataset. + random.shuffle(dataset) + + min_len, max_len = input_length_range + assert min_len >= 0 and max_len >= min_len, "input_length_range too small" + + # Filter out sequences that are too long or too short + filtered_requests: list[Request] = [] + + for i in range(len(dataset)): + if len(filtered_requests) == num_requests: + break + + # Tokenize the prompts and completions. + prompt_token_ids = tokenizer(dataset[i][0]).input_ids + prompt = tokenizer.decode(prompt_token_ids) + completion = dataset[i][1] + completion_token_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_token_ids) + output_len = ( + len(completion_token_ids) if fixed_output_len is None else fixed_output_len + ) + if min_len <= prompt_len <= max_len: + filtered_requests.append(Request(prompt, prompt_len, output_len)) + + return filtered_requests + + +def sample_requests_from_random( + num_requests: int, + tokenizer: PreTrainedTokenizerBase, + input_length_range: tuple[int, int], + fixed_output_len: Optional[int], + prefix_len: int, +) -> list[Request]: + requests = [] + prefix_token_ids = sample_tokens(tokenizer, prefix_len) + min_len, max_len = input_length_range + + for i in range(num_requests): + unique_part_token_ids = sample_tokens( + tokenizer, random.randint(min_len - prefix_len, max_len - prefix_len) + ) + prompt_token_ids = prefix_token_ids + unique_part_token_ids + prompt = tokenizer.decode(prompt_token_ids) + prompt_len = len(prompt_token_ids) + assert min_len <= prompt_len <= max_len, ( + f"prompt_len {prompt_len} out of range {min_len}:{max_len}" + ) + requests.append(Request(prompt, prompt_len, fixed_output_len)) + return requests + + +def repeat_and_sort_requests( + requests: list[Request], repeat_count: int, sort: bool = False +) -> list[str]: + repeated_requests = requests * repeat_count + if sort: + repeated_requests.sort(key=lambda x: x[1]) + else: + random.shuffle(repeated_requests) + return [req.prompt for req in repeated_requests] + + +def main(args): + tokenizer = get_tokenizer(args.model, trust_remote_code=True) + input_length_range = tuple(map(int, args.input_length_range.split(":"))) + random.seed(args.seed) + if args.dataset_path is not None: + if args.prefix_len > 0: + raise ValueError( + "prefix-len is not supported when dataset-path is provided." + ) + print(f"Start to sample {args.num_prompts} prompts from {args.dataset_path}") + filtered_requests = sample_requests_from_dataset( + dataset_path=args.dataset_path, + num_requests=args.num_prompts, + tokenizer=tokenizer, + input_length_range=input_length_range, + fixed_output_len=args.output_len, + ) + else: + print(f"Start to sample {args.num_prompts} prompts from random") + filtered_requests = sample_requests_from_random( + num_requests=args.num_prompts, + tokenizer=tokenizer, + input_length_range=input_length_range, + fixed_output_len=args.output_len, + prefix_len=args.prefix_len, + ) + + # Print some helpful stats of the requests. + print(f"Sampled {len(filtered_requests)} requests.") + prompt_lens = [req.prompt_len for req in filtered_requests] + print(f"Average input length: {sum(prompt_lens) / len(prompt_lens)}") + print(f"P50 input length: {sorted(prompt_lens)[len(prompt_lens) // 2]}") + print(f"Min Prompt Length: {min(prompt_lens)}") + print(f"Max Prompt Length: {max(prompt_lens)}") + + engine_args = EngineArgs.from_cli_args(args) + + llm = LLM(**dataclasses.asdict(engine_args)) + + sampling_params = SamplingParams( + temperature=0, + max_tokens=args.output_len, + detokenize=not args.disable_detokenize, + ) + + print("Testing filtered requests") + prompts = repeat_and_sort_requests( + filtered_requests, repeat_count=args.repeat_count, sort=args.sort + ) + + print("------start generating------") + test_prefix( + llm=llm, + prompts=prompts, + sampling_params=sampling_params, + ) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the performance with or without " + "automatic prefix caching." + ) + parser.add_argument( + "--dataset-path", type=str, default=None, help="Path to the dataset." + ) + parser.add_argument("--output-len", type=int, default=10) + parser.add_argument( + "--num-prompts", + type=int, + required=True, + help="Number of the prompts sampled from dataset", + ) + parser.add_argument( + "--repeat-count", + type=int, + default=1, + help="Number of times to repeat each prompt", + ) + parser.add_argument( + "--sort", action="store_true", help="Sort prompts by input length" + ) + parser.add_argument( + "--input-length-range", + type=str, + required=True, + help="Range of input lengths for sampling prompts," + 'specified as "min:max" (e.g., "128:256").', + ) + parser.add_argument( + "--prefix-len", + type=int, + default=0, + help="Specifies the length of a common prefix to be " + "added to the input prompt. The input-length-range will " + "subtract this length when filtering prompts. Only used " + "when dataset-path is not provided.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize responses (i.e. do not include " + "detokenization time in the latency measurement)" + ), + ) + + parser = EngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/offline_benchmark_test/benchmarks/benchmark_prioritization.py b/offline_benchmark_test/benchmarks/benchmark_prioritization.py new file mode 100644 index 0000000000000000000000000000000000000000..bb453791c1862bdd0883f0a47af6fddeef764f8f --- /dev/null +++ b/offline_benchmark_test/benchmarks/benchmark_prioritization.py @@ -0,0 +1,222 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Benchmark offline prioritization.""" + +import argparse +import dataclasses +import json +import random +import time +from typing import Optional + +from transformers import AutoTokenizer, PreTrainedTokenizerBase + +from vllm.engine.arg_utils import EngineArgs +from vllm.utils import FlexibleArgumentParser + + +# Select a equi-probable random priority +def get_random_flag(): + return 0 if random.random() < 0.5 else 1 + + +def sample_requests( + dataset_path: str, + num_requests: int, + tokenizer: PreTrainedTokenizerBase, + fixed_output_len: Optional[int], +) -> list[tuple[str, int, int, int]]: + if fixed_output_len is not None and fixed_output_len < 4: + raise ValueError("output_len too small") + + # Load the dataset. + with open(dataset_path) as f: + dataset = json.load(f) + # Filter out the conversations with less than 2 turns. + dataset = [data for data in dataset if len(data["conversations"]) >= 2] + # Only keep the first two turns of each conversation. + dataset = [ + (data["conversations"][0]["value"], data["conversations"][1]["value"]) + for data in dataset + ] + + # Shuffle the dataset. + random.shuffle(dataset) + + # Filter out sequences that are too long or too short + filtered_dataset: list[tuple[str, int, int]] = [] + for i in range(len(dataset)): + if len(filtered_dataset) == num_requests: + break + + # Tokenize the prompts and completions. + prompt = dataset[i][0] + prompt_token_ids = tokenizer(prompt).input_ids + completion = dataset[i][1] + completion_token_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_token_ids) + output_len = ( + len(completion_token_ids) if fixed_output_len is None else fixed_output_len + ) + if prompt_len < 4 or output_len < 4: + # Prune too short sequences. + continue + if prompt_len > 1024 or prompt_len + output_len > 2048: + # Prune too long sequences. + continue + + priority = get_random_flag() + + filtered_dataset.append((prompt, prompt_len, output_len, priority)) + + return filtered_dataset + + +def run_vllm( + requests: list[tuple[str, int, int]], + n: int, + engine_args: EngineArgs, + disable_detokenize: bool = False, +) -> float: + from vllm import LLM, SamplingParams + + llm = LLM(**dataclasses.asdict(engine_args)) + + assert all( + llm.llm_engine.model_config.max_model_len >= (request[1] + request[2]) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of" + " input_len and output_len for all requests." + ) + + # Add the requests to the engine. + prompts = [] + sampling_params = [] + priority = [] + for prompt, _, output_len, _priority in requests: + prompts.append(prompt) + priority.append(_priority) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=output_len, + detokenize=not disable_detokenize, + ) + ) + + start = time.perf_counter() + llm.generate(prompts, sampling_params, priority=priority, use_tqdm=True) + end = time.perf_counter() + return end - start + + +def main(args: argparse.Namespace): + print(args) + random.seed(args.seed) + + # Sample the requests. + tokenizer = AutoTokenizer.from_pretrained( + args.tokenizer, trust_remote_code=args.trust_remote_code + ) + if args.dataset is None: + # Synthesize a prompt with the given input length. + prompt = "hi" * (args.input_len - 1) + requests = [ + (prompt, args.input_len, args.output_len, get_random_flag()) + for _ in range(args.num_prompts) + ] + else: + requests = sample_requests( + args.dataset, args.num_prompts, tokenizer, args.output_len + ) + + if args.backend == "vllm": + elapsed_time = run_vllm( + requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize + ) + else: + raise ValueError(f"Unknown backend: {args.backend}") + total_num_tokens = sum( + prompt_len + output_len for _, prompt_len, output_len, priority in requests + ) + print( + f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, " + f"{total_num_tokens / elapsed_time:.2f} tokens/s" + ) + + # Output JSON results if specified + if args.output_json: + results = { + "elapsed_time": elapsed_time, + "num_requests": len(requests), + "total_num_tokens": total_num_tokens, + "requests_per_second": len(requests) / elapsed_time, + "tokens_per_second": total_num_tokens / elapsed_time, + } + with open(args.output_json, "w") as f: + json.dump(results, f, indent=4) + + +def create_argument_parser(): + parser = FlexibleArgumentParser(description="Benchmark the throughput.") + parser.add_argument( + "--backend", type=str, choices=["vllm", "hf", "mii"], default="vllm" + ) + parser.add_argument( + "--dataset", type=str, default=None, help="Path to the dataset." + ) + parser.add_argument( + "--input-len", + type=int, + default=None, + help="Input prompt length for each request", + ) + parser.add_argument( + "--output-len", + type=int, + default=None, + help="Output length for each request. Overrides the " + "output length from the dataset.", + ) + parser.add_argument( + "--n", type=int, default=1, help="Number of generated sequences per prompt." + ) + parser.add_argument( + "--num-prompts", type=int, default=200, help="Number of prompts to process." + ) + parser.add_argument( + "--output-json", + type=str, + default=None, + help="Path to save the throughput results in JSON format.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize responses (i.e. do not include " + "detokenization time in the latency measurement)" + ), + ) + + parser = EngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + if args.tokenizer is None: + args.tokenizer = args.model + if args.dataset is None: + assert args.input_len is not None + assert args.output_len is not None + else: + assert args.input_len is None + + main(args) diff --git a/offline_benchmark_test/benchmarks/benchmark_serving.py b/offline_benchmark_test/benchmarks/benchmark_serving.py new file mode 100644 index 0000000000000000000000000000000000000000..9b235266dff1aac95d6fe0509f4a5f01641579cc --- /dev/null +++ b/offline_benchmark_test/benchmarks/benchmark_serving.py @@ -0,0 +1,1378 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +r"""Benchmark online serving throughput. + +On the server side, run one of the following commands: + vLLM OpenAI API server + vllm serve \ + --swap-space 16 \ + --disable-log-requests + +On the client side, run: + python benchmarks/benchmark_serving.py \ + --backend \ + --model \ + --dataset-name sharegpt \ + --dataset-path \ + --request-rate \ # By default is inf + --num-prompts # By default is 1000 + + when using tgi backend, add + --endpoint /generate_stream + to the end of the command above. +""" + +import argparse +import asyncio +import gc +import json +import os +import random +import time +import warnings +from collections.abc import AsyncGenerator, Iterable +from dataclasses import dataclass +from datetime import datetime +from typing import Any, Literal, Optional + +import numpy as np +from tqdm.asyncio import tqdm +from transformers import PreTrainedTokenizerBase + +from backend_request_func import ( + ASYNC_REQUEST_FUNCS, + OPENAI_COMPATIBLE_BACKENDS, + RequestFuncInput, + RequestFuncOutput, +) + +try: + from vllm.transformers_utils.tokenizer import get_tokenizer +except ImportError: + from backend_request_func import get_tokenizer + +try: + from vllm.utils import FlexibleArgumentParser +except ImportError: + from argparse import ArgumentParser as FlexibleArgumentParser + +from benchmark_dataset import ( + AIMODataset, + ASRDataset, + BurstGPTDataset, + ConversationDataset, + CustomDataset, + HuggingFaceDataset, + InstructCoderDataset, + MTBenchDataset, + NextEditPredictionDataset, + RandomDataset, + SampleRequest, + ShareGPTDataset, + SonnetDataset, + VisionArenaDataset, +) +from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json + +MILLISECONDS_TO_SECONDS_CONVERSION = 1000 + + +@dataclass +class BenchmarkMetrics: + completed: int + total_input: int + total_output: int + request_throughput: float + request_goodput: float + output_throughput: float + total_token_throughput: float + mean_ttft_ms: float + median_ttft_ms: float + std_ttft_ms: float + percentiles_ttft_ms: list[tuple[float, float]] + mean_tpot_ms: float + median_tpot_ms: float + std_tpot_ms: float + percentiles_tpot_ms: list[tuple[float, float]] + mean_itl_ms: float + median_itl_ms: float + std_itl_ms: float + percentiles_itl_ms: list[tuple[float, float]] + # E2EL stands for end-to-end latency per request. + # It is the time taken on the client side from sending + # a request to receiving a complete response. + mean_e2el_ms: float + median_e2el_ms: float + std_e2el_ms: float + percentiles_e2el_ms: list[tuple[float, float]] + + +def _get_current_request_rate( + ramp_up_strategy: Optional[Literal["linear", "exponential"]], + ramp_up_start_rps: Optional[int], + ramp_up_end_rps: Optional[int], + request_index: int, + total_requests: int, + request_rate: float, +) -> float: + if ( + ramp_up_strategy + and ramp_up_start_rps is not None + and ramp_up_end_rps is not None + ): + progress = request_index / max(total_requests - 1, 1) + if ramp_up_strategy == "linear": + increase = (ramp_up_end_rps - ramp_up_start_rps) * progress + return ramp_up_start_rps + increase + elif ramp_up_strategy == "exponential": + ratio = ramp_up_end_rps / ramp_up_start_rps + return ramp_up_start_rps * (ratio**progress) + else: + raise ValueError(f"Unknown ramp-up strategy: {ramp_up_strategy}") + return request_rate + + +async def get_request( + input_requests: list[SampleRequest], + request_rate: float, + burstiness: float = 1.0, + ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None, + ramp_up_start_rps: Optional[int] = None, + ramp_up_end_rps: Optional[int] = None, +) -> AsyncGenerator[tuple[SampleRequest, float], None]: + """ + Asynchronously generates requests at a specified rate + with OPTIONAL burstiness and OPTIONAL ramp-up strategy. + + Args: + input_requests: + A list of input requests, each represented as a SampleRequest. + request_rate: + The rate at which requests are generated (requests/s). + burstiness (optional): + The burstiness factor of the request generation. + Only takes effect when request_rate is not inf. + Default value is 1, which follows a Poisson process. + Otherwise, the request intervals follow a gamma distribution. + A lower burstiness value (0 < burstiness < 1) results + in more bursty requests, while a higher burstiness value + (burstiness > 1) results in a more uniform arrival of requests. + ramp_up_strategy (optional): + The ramp-up strategy. Can be "linear" or "exponential". + If None, uses constant request rate (specified by request_rate). + ramp_up_start_rps (optional): + The starting request rate for ramp-up. + ramp_up_end_rps (optional): + The ending request rate for ramp-up. + """ + assert burstiness > 0, ( + f"A positive burstiness factor is expected, but given {burstiness}." + ) + # Convert to list to get length for ramp-up calculations + if isinstance(input_requests, Iterable) and not isinstance(input_requests, list): + input_requests = list(input_requests) + + total_requests = len(input_requests) + request_index = 0 + + for request in input_requests: + current_request_rate = _get_current_request_rate( + ramp_up_strategy, + ramp_up_start_rps, + ramp_up_end_rps, + request_index, + total_requests, + request_rate, + ) + + yield request, current_request_rate + + request_index += 1 + + if current_request_rate == float("inf"): + # If the request rate is infinity, then we don't need to wait. + continue + + theta = 1.0 / (current_request_rate * burstiness) + + # Sample the request interval from the gamma distribution. + # If burstiness is 1, it follows exponential distribution. + interval = np.random.gamma(shape=burstiness, scale=theta) + # The next request will be sent after the interval. + await asyncio.sleep(interval) + + +def calculate_metrics( + input_requests: list[SampleRequest], + outputs: list[RequestFuncOutput], + dur_s: float, + tokenizer: PreTrainedTokenizerBase, + selected_percentile_metrics: list[str], + selected_percentiles: list[float], + goodput_config_dict: dict[str, float], +) -> tuple[BenchmarkMetrics, list[int]]: + actual_output_lens: list[int] = [] + total_input = 0 + completed = 0 + good_completed = 0 + itls: list[float] = [] + tpots: list[float] = [] + all_tpots: list[float] = [] + ttfts: list[float] = [] + e2els: list[float] = [] + for i in range(len(outputs)): + if outputs[i].success: + output_len = outputs[i].output_tokens + + if not output_len: + # We use the tokenizer to count the number of output tokens + # for some serving backends instead of looking at + # len(outputs[i].itl) since multiple output tokens may be + # bundled together + # Note : this may inflate the output token count slightly + output_len = len( + tokenizer( + outputs[i].generated_text, add_special_tokens=False + ).input_ids + ) + actual_output_lens.append(output_len) + total_input += input_requests[i].prompt_len + tpot = 0 + if output_len > 1: + latency_minus_ttft = outputs[i].latency - outputs[i].ttft + tpot = latency_minus_ttft / (output_len - 1) + tpots.append(tpot) + # Note: if output_len <= 1, we regard tpot as 0 for goodput + all_tpots.append(tpot) + itls += outputs[i].itl + ttfts.append(outputs[i].ttft) + e2els.append(outputs[i].latency) + completed += 1 + else: + actual_output_lens.append(0) + + if goodput_config_dict: + valid_metrics = [] + slo_values = [] + + if "ttft" in goodput_config_dict: + valid_metrics.append(ttfts) + slo_values.append( + goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "tpot" in goodput_config_dict: + valid_metrics.append(all_tpots) + slo_values.append( + goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "e2el" in goodput_config_dict: + valid_metrics.append(e2els) + slo_values.append( + goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + + for req_metric in zip(*valid_metrics): + is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)]) + if is_good_req: + good_completed += 1 + + if completed == 0: + warnings.warn( + "All requests failed. This is likely due to a misconfiguration " + "on the benchmark arguments.", + stacklevel=2, + ) + metrics = BenchmarkMetrics( + completed=completed, + total_input=total_input, + total_output=sum(actual_output_lens), + request_throughput=completed / dur_s, + request_goodput=good_completed / dur_s, + output_throughput=sum(actual_output_lens) / dur_s, + total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s, + mean_ttft_ms=np.mean(ttfts or 0) + * 1000, # ttfts is empty if streaming is not supported by backend + std_ttft_ms=np.std(ttfts or 0) * 1000, + median_ttft_ms=np.median(ttfts or 0) * 1000, + percentiles_ttft_ms=[ + (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles + ], + mean_tpot_ms=np.mean(tpots or 0) * 1000, + std_tpot_ms=np.std(tpots or 0) * 1000, + median_tpot_ms=np.median(tpots or 0) * 1000, + percentiles_tpot_ms=[ + (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles + ], + mean_itl_ms=np.mean(itls or 0) * 1000, + std_itl_ms=np.std(itls or 0) * 1000, + median_itl_ms=np.median(itls or 0) * 1000, + percentiles_itl_ms=[ + (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles + ], + mean_e2el_ms=np.mean(e2els or 0) * 1000, + std_e2el_ms=np.std(e2els or 0) * 1000, + median_e2el_ms=np.median(e2els or 0) * 1000, + percentiles_e2el_ms=[ + (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles + ], + ) + + return metrics, actual_output_lens + + +async def benchmark( + backend: str, + api_url: str, + base_url: str, + model_id: str, + model_name: str, + tokenizer: PreTrainedTokenizerBase, + input_requests: list[SampleRequest], + logprobs: Optional[int], + request_rate: float, + burstiness: float, + disable_tqdm: bool, + profile: bool, + selected_percentile_metrics: list[str], + selected_percentiles: list[float], + ignore_eos: bool, + goodput_config_dict: dict[str, float], + max_concurrency: Optional[int], + lora_modules: Optional[Iterable[str]], + extra_body: Optional[dict], + ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None, + ramp_up_start_rps: Optional[int] = None, + ramp_up_end_rps: Optional[int] = None, +): + if backend in ASYNC_REQUEST_FUNCS: + request_func = ASYNC_REQUEST_FUNCS[backend] + else: + raise ValueError(f"Unknown backend: {backend}") + + print("Starting initial single prompt test run...") + test_prompt, test_prompt_len, test_output_len, test_mm_content = ( + input_requests[0].prompt, + input_requests[0].prompt_len, + input_requests[0].expected_output_len, + input_requests[0].multi_modal_data, + ) + + assert test_mm_content is None or isinstance(test_mm_content, dict) + test_input = RequestFuncInput( + model=model_id, + model_name=model_name, + prompt=test_prompt, + api_url=api_url, + prompt_len=test_prompt_len, + output_len=test_output_len, + logprobs=logprobs, + multi_modal_content=test_mm_content, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + + test_output = await request_func(request_func_input=test_input) + if not test_output.success: + raise ValueError( + "Initial test run failed - Please make sure benchmark arguments " + f"are correctly specified. Error: {test_output.error}" + ) + else: + print("Initial test run completed. Starting main benchmark run...") + + if lora_modules: + # For each input request, choose a LoRA module at random. + lora_modules = iter( + [random.choice(lora_modules) for _ in range(len(input_requests))] + ) + + if profile: + print("Starting profiler...") + profile_input = RequestFuncInput( + model=model_id, + model_name=model_name, + prompt=test_prompt, + api_url=base_url + "/start_profile", + prompt_len=test_prompt_len, + output_len=test_output_len, + logprobs=logprobs, + multi_modal_content=test_mm_content, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler started") + + distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution" + + if ramp_up_strategy is not None: + print( + f"Traffic ramp-up strategy: {ramp_up_strategy}. Will increase " + f"RPS from {ramp_up_start_rps} to {ramp_up_end_rps} RPS over " + "the duration of the benchmark." + ) + else: + print(f"Traffic request rate: {request_rate} RPS.") + + print(f"Burstiness factor: {burstiness} ({distribution})") + print(f"Maximum request concurrency: {max_concurrency}") + + pbar = None if disable_tqdm else tqdm(total=len(input_requests)) + + # This can be used once the minimum Python version is 3.10 or higher, + # and it will simplify the code in limited_request_func. + # semaphore = (asyncio.Semaphore(max_concurrency) + # if max_concurrency else contextlib.nullcontext()) + semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None + + async def limited_request_func(request_func_input, pbar): + if semaphore is None: + return await request_func(request_func_input=request_func_input, pbar=pbar) + async with semaphore: + return await request_func(request_func_input=request_func_input, pbar=pbar) + + benchmark_start_time = time.perf_counter() + tasks: list[asyncio.Task] = [] + + rps_change_events = [] + last_int_rps = -1 + if ramp_up_strategy is not None and ramp_up_start_rps is not None: + last_int_rps = ramp_up_start_rps + rps_change_events.append( + { + "rps": last_int_rps, + "timestamp": datetime.now().isoformat(), + } + ) + + async for request, current_request_rate in get_request( + input_requests, + request_rate, + burstiness, + ramp_up_strategy, + ramp_up_start_rps, + ramp_up_end_rps, + ): + if ramp_up_strategy is not None: + current_int_rps = int(current_request_rate) + if current_int_rps > last_int_rps: + timestamp = datetime.now().isoformat() + for rps_val in range(last_int_rps + 1, current_int_rps + 1): + rps_change_events.append({"rps": rps_val, "timestamp": timestamp}) + last_int_rps = current_int_rps + + prompt, prompt_len, output_len, mm_content = ( + request.prompt, + request.prompt_len, + request.expected_output_len, + request.multi_modal_data, + ) + req_model_id, req_model_name = model_id, model_name + if lora_modules: + req_lora_module = next(lora_modules) + req_model_id, req_model_name = req_lora_module, req_lora_module + + request_func_input = RequestFuncInput( + model=req_model_id, + model_name=req_model_name, + prompt=prompt, + api_url=api_url, + prompt_len=prompt_len, + output_len=output_len, + logprobs=logprobs, + multi_modal_content=mm_content, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + task = limited_request_func(request_func_input=request_func_input, pbar=pbar) + tasks.append(asyncio.create_task(task)) + outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks) + + if profile: + print("Stopping profiler...") + profile_input = RequestFuncInput( + model=model_id, + prompt=test_prompt, + api_url=base_url + "/stop_profile", + prompt_len=test_prompt_len, + output_len=test_output_len, + logprobs=logprobs, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler stopped") + + if pbar is not None: + pbar.close() + + benchmark_duration = time.perf_counter() - benchmark_start_time + + metrics, actual_output_lens = calculate_metrics( + input_requests=input_requests, + outputs=outputs, + dur_s=benchmark_duration, + tokenizer=tokenizer, + selected_percentile_metrics=selected_percentile_metrics, + selected_percentiles=selected_percentiles, + goodput_config_dict=goodput_config_dict, + ) + + print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="=")) + print("{:<40} {:<10}".format("Successful requests:", metrics.completed)) + print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration)) + print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input)) + print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output)) + print( + "{:<40} {:<10.2f}".format( + "Request throughput (req/s):", metrics.request_throughput + ) + ) + if goodput_config_dict: + print( + "{:<40} {:<10.2f}".format( + "Request goodput (req/s):", metrics.request_goodput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Output token throughput (tok/s):", metrics.output_throughput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Total Token throughput (tok/s):", metrics.total_token_throughput + ) + ) + + result = { + "duration": benchmark_duration, + "completed": metrics.completed, + "total_input_tokens": metrics.total_input, + "total_output_tokens": metrics.total_output, + "request_throughput": metrics.request_throughput, + "request_goodput": metrics.request_goodput if goodput_config_dict else None, + "output_throughput": metrics.output_throughput, + "total_token_throughput": metrics.total_token_throughput, + "input_lens": [output.prompt_len for output in outputs], + "output_lens": actual_output_lens, + "ttfts": [output.ttft for output in outputs], + "itls": [output.itl for output in outputs], + "generated_texts": [output.generated_text for output in outputs], + "errors": [output.error for output in outputs], + } + + if rps_change_events: + result["rps_change_events"] = rps_change_events + + def process_one_metric( + # E.g., "ttft" + metric_attribute_name: str, + # E.g., "TTFT" + metric_name: str, + # E.g., "Time to First Token" + metric_header: str, + ): + # This function prints and adds statistics of the specified + # metric. + if metric_attribute_name not in selected_percentile_metrics: + return + print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-")) + print( + "{:<40} {:<10.2f}".format( + f"Mean {metric_name} (ms):", + getattr(metrics, f"mean_{metric_attribute_name}_ms"), + ) + ) + print( + "{:<40} {:<10.2f}".format( + f"Median {metric_name} (ms):", + getattr(metrics, f"median_{metric_attribute_name}_ms"), + ) + ) + result[f"mean_{metric_attribute_name}_ms"] = getattr( + metrics, f"mean_{metric_attribute_name}_ms" + ) + result[f"median_{metric_attribute_name}_ms"] = getattr( + metrics, f"median_{metric_attribute_name}_ms" + ) + result[f"std_{metric_attribute_name}_ms"] = getattr( + metrics, f"std_{metric_attribute_name}_ms" + ) + for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"): + p_word = str(int(p)) if int(p) == p else str(p) + print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value)) + result[f"p{p_word}_{metric_attribute_name}_ms"] = value + + process_one_metric("ttft", "TTFT", "Time to First Token") + process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)") + process_one_metric("itl", "ITL", "Inter-token Latency") + process_one_metric("e2el", "E2EL", "End-to-end Latency") + + print("=" * 50) + + return result + + +def check_goodput_args(args): + # Check and parse goodput arguments + goodput_config_dict = {} + VALID_NAMES = ["ttft", "tpot", "e2el"] + if args.goodput: + goodput_config_dict = parse_goodput(args.goodput) + for slo_name, slo_val in goodput_config_dict.items(): + if slo_name not in VALID_NAMES: + raise ValueError( + f"Invalid metric name found, {slo_name}: {slo_val}. " + "The service level objective name should be one of " + f"{str(VALID_NAMES)}. " + ) + if slo_val < 0: + raise ValueError( + f"Invalid value found, {slo_name}: {slo_val}. " + "The service level objective value should be " + "non-negative." + ) + return goodput_config_dict + + +def parse_goodput(slo_pairs): + goodput_config_dict = {} + try: + for slo_pair in slo_pairs: + slo_name, slo_val = slo_pair.split(":") + goodput_config_dict[slo_name] = float(slo_val) + except ValueError as err: + raise argparse.ArgumentTypeError( + "Invalid format found for service level objectives. " + 'Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is a " + "number in milliseconds." + ) from err + return goodput_config_dict + + +def save_to_pytorch_benchmark_format( + args: argparse.Namespace, results: dict[str, Any], file_name: str +) -> None: + metrics = [ + "median_ttft_ms", + "mean_ttft_ms", + "std_ttft_ms", + "p99_ttft_ms", + "mean_tpot_ms", + "median_tpot_ms", + "std_tpot_ms", + "p99_tpot_ms", + "median_itl_ms", + "mean_itl_ms", + "std_itl_ms", + "p99_itl_ms", + ] + # These raw data might be useful, but they are rather big. They can be added + # later if needed + ignored_metrics = ["ttfts", "itls", "generated_texts", "errors"] + pt_records = convert_to_pytorch_benchmark_format( + args=args, + metrics={k: [results[k]] for k in metrics}, + extra_info={ + k: results[k] + for k in results + if k not in metrics and k not in ignored_metrics + }, + ) + if pt_records: + # Don't use json suffix here as we don't want CI to pick it up + pt_file = f"{os.path.splitext(file_name)[0]}.pytorch.json" + write_to_json(pt_file, pt_records) + + +def main(args: argparse.Namespace): + print(args) + random.seed(args.seed) + np.random.seed(args.seed) + + backend = args.backend + model_id = args.model + model_name = args.served_model_name + tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model + tokenizer_mode = args.tokenizer_mode + + # Validate ramp-up arguments + if args.ramp_up_strategy is not None: + if args.request_rate != float("inf"): + raise ValueError( + "When using ramp-up, do not specify --request-rate. " + "The request rate will be controlled by ramp-up parameters. " + "Please remove the --request-rate argument." + ) + if args.ramp_up_start_rps is None or args.ramp_up_end_rps is None: + raise ValueError( + "When using --ramp-up-strategy, both --ramp-up-start-rps and " + "--ramp-up-end-rps must be specified" + ) + if args.ramp_up_start_rps < 0 or args.ramp_up_end_rps < 0: + raise ValueError("Ramp-up start and end RPS must be non-negative") + if args.ramp_up_start_rps > args.ramp_up_end_rps: + raise ValueError("Ramp-up start RPS must be less than end RPS") + if args.ramp_up_strategy == "exponential" and args.ramp_up_start_rps == 0: + raise ValueError("For exponential ramp-up, the start RPS cannot be 0.") + + if args.base_url is not None: + api_url = f"{args.base_url}{args.endpoint}" + base_url = f"{args.base_url}" + else: + api_url = f"http://{args.host}:{args.port}{args.endpoint}" + base_url = f"http://{args.host}:{args.port}" + + tokenizer = get_tokenizer( + tokenizer_id, + tokenizer_mode=tokenizer_mode, + trust_remote_code=args.trust_remote_code, + ) + + if args.dataset_name is None: + raise ValueError( + "Please specify '--dataset-name' and the corresponding " + "'--dataset-path' if required." + ) + + if args.dataset_name == "custom": + dataset = CustomDataset(dataset_path=args.dataset_path) + input_requests = dataset.sample( + num_requests=args.num_prompts, + tokenizer=tokenizer, + output_len=args.custom_output_len, + skip_chat_template=args.custom_skip_chat_template, + ) + + elif args.dataset_name == "sonnet": + dataset = SonnetDataset(dataset_path=args.dataset_path) + # For the "sonnet" dataset, formatting depends on the backend. + if args.backend == "openai-chat": + input_requests = dataset.sample( + num_requests=args.num_prompts, + input_len=args.sonnet_input_len, + output_len=args.sonnet_output_len, + prefix_len=args.sonnet_prefix_len, + tokenizer=tokenizer, + return_prompt_formatted=False, + ) + else: + assert tokenizer.chat_template or tokenizer.default_chat_template, ( + "Tokenizer/model must have chat template for sonnet dataset." + ) + input_requests = dataset.sample( + num_requests=args.num_prompts, + input_len=args.sonnet_input_len, + output_len=args.sonnet_output_len, + prefix_len=args.sonnet_prefix_len, + tokenizer=tokenizer, + return_prompt_formatted=True, + ) + + elif args.dataset_name == "hf": + # all following datasets are implemented from the + # HuggingFaceDataset base class + if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS: + dataset_class = VisionArenaDataset + args.hf_split = "train" + args.hf_subset = None + elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS: + dataset_class = InstructCoderDataset + args.hf_split = "train" + elif args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS: + dataset_class = MTBenchDataset + args.hf_split = "train" + elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS: + dataset_class = ConversationDataset + elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS: + dataset_class = AIMODataset + args.hf_split = "train" + elif args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS: # noqa: E501 + dataset_class = NextEditPredictionDataset + args.hf_split = "train" + elif args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS: + dataset_class = ASRDataset + args.hf_split = "train" + else: + supported_datasets = set( + [ + dataset_name + for cls in HuggingFaceDataset.__subclasses__() + for dataset_name in cls.SUPPORTED_DATASET_PATHS + ] + ) + raise ValueError( + f"Unsupported dataset path: {args.dataset_path}. " + "Huggingface dataset only supports dataset_path" + f" from one of following: {supported_datasets}. " + "Please consider contributing if you would " + "like to add support for additional dataset formats." + ) + + if dataset_class.IS_MULTIMODAL and backend not in [ + "openai-chat", + "openai-audio", + ]: + # multi-modal benchmark is only available on OpenAI Chat backend. + raise ValueError( + "Multi-modal content is only supported on 'openai-chat' and " + "'openai-audio' backend." + ) + input_requests = dataset_class( + dataset_path=args.dataset_path, + dataset_subset=args.hf_subset, + dataset_split=args.hf_split, + random_seed=args.seed, + ).sample( + num_requests=args.num_prompts, + tokenizer=tokenizer, + output_len=args.hf_output_len, + ) + + else: + # For datasets that follow a similar structure, use a mapping. + dataset_mapping = { + "sharegpt": lambda: ShareGPTDataset( + random_seed=args.seed, dataset_path=args.dataset_path + ).sample( + tokenizer=tokenizer, + num_requests=args.num_prompts, + output_len=args.sharegpt_output_len, + ), + "burstgpt": lambda: BurstGPTDataset( + random_seed=args.seed, dataset_path=args.dataset_path + ).sample(tokenizer=tokenizer, num_requests=args.num_prompts), + "random": lambda: RandomDataset(dataset_path=args.dataset_path).sample( + tokenizer=tokenizer, + num_requests=args.num_prompts, + prefix_len=args.random_prefix_len, + input_len=args.random_input_len, + output_len=args.random_output_len, + range_ratio=args.random_range_ratio, + ), + } + + try: + input_requests = dataset_mapping[args.dataset_name]() + except KeyError as err: + raise ValueError(f"Unknown dataset: {args.dataset_name}") from err + goodput_config_dict = check_goodput_args(args) + + # Collect the sampling parameters. + sampling_params = { + k: v + for k, v in { + "top_p": args.top_p, + "top_k": args.top_k, + "min_p": args.min_p, + "temperature": args.temperature, + }.items() + if v is not None + } + + # Sampling parameters are only supported by openai-compatible backend. + if sampling_params and args.backend not in OPENAI_COMPATIBLE_BACKENDS: + raise ValueError( + "Sampling parameters are only supported by openai-compatible backends." + ) + + if "temperature" not in sampling_params: + sampling_params["temperature"] = 0.0 # Default to greedy decoding. + + if args.backend == "llama.cpp": + # Disable prompt caching in llama.cpp backend + sampling_params["cache_prompt"] = False + + # Avoid GC processing "static" data - reduce pause times. + gc.collect() + gc.freeze() + + benchmark_result = asyncio.run( + benchmark( + backend=backend, + api_url=api_url, + base_url=base_url, + model_id=model_id, + model_name=model_name, + tokenizer=tokenizer, + input_requests=input_requests, + logprobs=args.logprobs, + request_rate=args.request_rate, + burstiness=args.burstiness, + disable_tqdm=args.disable_tqdm, + profile=args.profile, + selected_percentile_metrics=args.percentile_metrics.split(","), + selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")], + ignore_eos=args.ignore_eos, + goodput_config_dict=goodput_config_dict, + max_concurrency=args.max_concurrency, + lora_modules=args.lora_modules, + extra_body=sampling_params, + ramp_up_strategy=args.ramp_up_strategy, + ramp_up_start_rps=args.ramp_up_start_rps, + ramp_up_end_rps=args.ramp_up_end_rps, + ) + ) + + # Save config and results to json + if args.save_result or args.append_result: + result_json: dict[str, Any] = {} + + # Setup + current_dt = datetime.now().strftime("%Y%m%d-%H%M%S") + result_json["date"] = current_dt + result_json["backend"] = backend + result_json["model_id"] = model_id + result_json["tokenizer_id"] = tokenizer_id + result_json["num_prompts"] = args.num_prompts + + # Metadata + if args.metadata: + for item in args.metadata: + if "=" in item: + kvstring = item.split("=") + result_json[kvstring[0].strip()] = kvstring[1].strip() + else: + raise ValueError( + "Invalid metadata format. Please use KEY=VALUE format." + ) + # Traffic + result_json["request_rate"] = ( + args.request_rate if args.request_rate < float("inf") else "inf" + ) + result_json["burstiness"] = args.burstiness + result_json["max_concurrency"] = args.max_concurrency + + if args.ramp_up_strategy is not None: + result_json["ramp_up_strategy"] = args.ramp_up_strategy + result_json["ramp_up_start_rps"] = args.ramp_up_start_rps + result_json["ramp_up_end_rps"] = args.ramp_up_end_rps + + # Merge with benchmark result + result_json = {**result_json, **benchmark_result} + + if not args.save_detailed: + # Remove fields with too many data points + for field in [ + "input_lens", + "output_lens", + "ttfts", + "itls", + "generated_texts", + "errors", + ]: + if field in result_json: + del result_json[field] + if field in benchmark_result: + del benchmark_result[field] + + # Save to file + base_model_id = model_id.split("/")[-1] + max_concurrency_str = ( + f"-concurrency{args.max_concurrency}" + if args.max_concurrency is not None + else "" + ) + if args.ramp_up_strategy is not None: + file_name = f"{backend}-ramp-up-{args.ramp_up_strategy}-{args.ramp_up_start_rps}qps-{args.ramp_up_end_rps}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json" # noqa + else: + file_name = f"{backend}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json" # noqa + if args.result_filename: + file_name = args.result_filename + if args.result_dir: + os.makedirs(args.result_dir, exist_ok=True) + file_name = os.path.join(args.result_dir, file_name) + with open( + file_name, mode="a+" if args.append_result else "w", encoding="utf-8" + ) as outfile: + # Append a newline. + if args.append_result and outfile.tell() != 0: + outfile.write("\n") + json.dump(result_json, outfile) + save_to_pytorch_benchmark_format(args, result_json, file_name) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the online serving throughput." + ) + parser.add_argument( + "--backend", + type=str, + default="vllm", + choices=list(ASYNC_REQUEST_FUNCS.keys()), + ) + parser.add_argument( + "--base-url", + type=str, + default=None, + help="Server or API base url if not using http host and port.", + ) + # Use 127.0.0.1 here instead of localhost to force the use of ipv4 + parser.add_argument("--host", type=str, default="127.0.0.1") + parser.add_argument("--port", type=int, default=8000) + parser.add_argument( + "--endpoint", + type=str, + default="/v1/completions", + help="API endpoint.", + ) + parser.add_argument( + "--dataset-name", + type=str, + default="sharegpt", + choices=["sharegpt", "burstgpt", "sonnet", "random", "hf", "custom"], + help="Name of the dataset to benchmark on.", + ) + parser.add_argument( + "--dataset-path", + type=str, + default=None, + help="Path to the sharegpt/sonnet dataset. " + "Or the huggingface dataset ID if using HF dataset.", + ) + parser.add_argument( + "--max-concurrency", + type=int, + default=None, + help="Maximum number of concurrent requests. This can be used " + "to help simulate an environment where a higher level component " + "is enforcing a maximum number of concurrent requests. While the " + "--request-rate argument controls the rate at which requests are " + "initiated, this argument will control how many are actually allowed " + "to execute at a time. This means that when used in combination, the " + "actual request rate may be lower than specified with --request-rate, " + "if the server is not processing requests fast enough to keep up.", + ) + + parser.add_argument( + "--model", + type=str, + required=True, + help="Name of the model.", + ) + parser.add_argument( + "--tokenizer", + type=str, + help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 + ) + parser.add_argument("--use-beam-search", action="store_true") + parser.add_argument( + "--num-prompts", + type=int, + default=1000, + help="Number of prompts to process.", + ) + parser.add_argument( + "--logprobs", + type=int, + default=None, + help=( + "Number of logprobs-per-token to compute & return as part of " + "the request. If unspecified, then either (1) if beam search " + "is disabled, no logprobs are computed & a single dummy " + "logprob is returned for each token; or (2) if beam search " + "is enabled 1 logprob per token is computed" + ), + ) + parser.add_argument( + "--request-rate", + type=float, + default=float("inf"), + help="Number of requests per second. If this is inf, " + "then all the requests are sent at time 0. " + "Otherwise, we use Poisson process or gamma distribution " + "to synthesize the request arrival times.", + ) + parser.add_argument( + "--burstiness", + type=float, + default=1.0, + help="Burstiness factor of the request generation. " + "Only take effect when request_rate is not inf. " + "Default value is 1, which follows Poisson process. " + "Otherwise, the request intervals follow a gamma distribution. " + "A lower burstiness value (0 < burstiness < 1) results in more " + "bursty requests. A higher burstiness value (burstiness > 1) " + "results in a more uniform arrival of requests.", + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument( + "--trust-remote-code", + action="store_true", + help="Trust remote code from huggingface", + ) + parser.add_argument( + "--disable-tqdm", + action="store_true", + help="Specify to disable tqdm progress bar.", + ) + parser.add_argument( + "--profile", + action="store_true", + help="Use Torch Profiler. The endpoint must be launched with " + "VLLM_TORCH_PROFILER_DIR to enable profiler.", + ) + parser.add_argument( + "--save-result", + action="store_true", + help="Specify to save benchmark results to a json file", + ) + parser.add_argument( + "--save-detailed", + action="store_true", + help="When saving the results, whether to include per request " + "information such as response, error, ttfs, tpots, etc.", + ) + parser.add_argument( + "--append-result", + action="store_true", + help="Append the benchmark result to the existing json file.", + ) + parser.add_argument( + "--metadata", + metavar="KEY=VALUE", + nargs="*", + help="Key-value pairs (e.g, --metadata version=0.3.3 tp=1) " + "for metadata of this run to be saved in the result JSON file " + "for record keeping purposes.", + ) + parser.add_argument( + "--result-dir", + type=str, + default=None, + help="Specify directory to save benchmark json results." + "If not specified, results are saved in the current directory.", + ) + parser.add_argument( + "--result-filename", + type=str, + default=None, + help="Specify the filename to save benchmark json results." + "If not specified, results will be saved in " + "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json" + " format.", + ) + parser.add_argument( + "--ignore-eos", + action="store_true", + help="Set ignore_eos flag when sending the benchmark request." + "Warning: ignore_eos is not supported in deepspeed_mii and tgi.", + ) + parser.add_argument( + "--percentile-metrics", + type=str, + default="ttft,tpot,itl", + help="Comma-separated list of selected metrics to report percentils. " + "This argument specifies the metrics to report percentiles. " + 'Allowed metric names are "ttft", "tpot", "itl", "e2el". ' + 'Default value is "ttft,tpot,itl".', + ) + parser.add_argument( + "--metric-percentiles", + type=str, + default="99", + help="Comma-separated list of percentiles for selected metrics. " + 'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". ' + 'Default value is "99". ' + 'Use "--percentile-metrics" to select metrics.', + ) + parser.add_argument( + "--goodput", + nargs="+", + required=False, + help='Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is in " + 'milliseconds. Multiple "KEY:VALUE" pairs can be provided, ' + "separated by spaces. Allowed request level metric names are " + '"ttft", "tpot", "e2el". For more context on the definition of ' + "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 " + "and the blog: https://hao-ai-lab.github.io/blogs/distserve", + ) + + # group for dataset specific arguments + custom_group = parser.add_argument_group("custom dataset options") + custom_group.add_argument( + "--custom-output-len", + type=int, + default=256, + help="Number of output tokens per request, used only for custom dataset.", + ) + custom_group.add_argument( + "--custom-skip-chat-template", + action="store_true", + help="Skip applying chat template to prompt, used only for custom dataset.", + ) + + sonnet_group = parser.add_argument_group("sonnet dataset options") + sonnet_group.add_argument( + "--sonnet-input-len", + type=int, + default=550, + help="Number of input tokens per request, used only for sonnet dataset.", + ) + sonnet_group.add_argument( + "--sonnet-output-len", + type=int, + default=150, + help="Number of output tokens per request, used only for sonnet dataset.", + ) + sonnet_group.add_argument( + "--sonnet-prefix-len", + type=int, + default=200, + help="Number of prefix tokens per request, used only for sonnet dataset.", + ) + + sharegpt_group = parser.add_argument_group("sharegpt dataset options") + sharegpt_group.add_argument( + "--sharegpt-output-len", + type=int, + default=None, + help="Output length for each request. Overrides the output length " + "from the ShareGPT dataset.", + ) + + random_group = parser.add_argument_group("random dataset options") + random_group.add_argument( + "--random-input-len", + type=int, + default=1024, + help="Number of input tokens per request, used only for random sampling.", + ) + random_group.add_argument( + "--random-output-len", + type=int, + default=128, + help="Number of output tokens per request, used only for random sampling.", + ) + random_group.add_argument( + "--random-range-ratio", + type=float, + default=0.0, + help="Range ratio for sampling input/output length, " + "used only for random sampling. Must be in the range [0, 1) to define " + "a symmetric sampling range" + "[length * (1 - range_ratio), length * (1 + range_ratio)].", + ) + random_group.add_argument( + "--random-prefix-len", + type=int, + default=0, + help=( + "Number of fixed prefix tokens before the random context " + "in a request. " + "The total input length is the sum of `random-prefix-len` and " + "a random " + "context length sampled from [input_len * (1 - range_ratio), " + "input_len * (1 + range_ratio)]." + ), + ) + + hf_group = parser.add_argument_group("hf dataset options") + hf_group.add_argument( + "--hf-subset", type=str, default=None, help="Subset of the HF dataset." + ) + hf_group.add_argument( + "--hf-split", type=str, default=None, help="Split of the HF dataset." + ) + hf_group.add_argument( + "--hf-output-len", + type=int, + default=None, + help="Output length for each request. Overrides the output lengths " + "from the sampled HF dataset.", + ) + + sampling_group = parser.add_argument_group("sampling parameters") + sampling_group.add_argument( + "--top-p", + type=float, + default=None, + help="Top-p sampling parameter. Only has effect on openai-compatible backends.", + ) + sampling_group.add_argument( + "--top-k", + type=int, + default=None, + help="Top-k sampling parameter. Only has effect on openai-compatible backends.", + ) + sampling_group.add_argument( + "--min-p", + type=float, + default=None, + help="Min-p sampling parameter. Only has effect on openai-compatible backends.", + ) + sampling_group.add_argument( + "--temperature", + type=float, + default=None, + help="Temperature sampling parameter. Only has effect on " + "openai-compatible backends. If not specified, default to greedy " + "decoding (i.e. temperature==0.0).", + ) + + parser.add_argument( + "--tokenizer-mode", + type=str, + default="auto", + choices=["auto", "slow", "mistral", "custom"], + help='The tokenizer mode.\n\n* "auto" will use the ' + 'fast tokenizer if available.\n* "slow" will ' + "always use the slow tokenizer. \n* " + '"mistral" will always use the `mistral_common` tokenizer. \n*' + '"custom" will use --tokenizer to select the preregistered tokenizer.', + ) + + parser.add_argument( + "--served-model-name", + type=str, + default=None, + help="The model name used in the API. " + "If not specified, the model name will be the " + "same as the ``--model`` argument. ", + ) + + parser.add_argument( + "--lora-modules", + nargs="+", + default=None, + help="A subset of LoRA module names passed in when " + "launching the server. For each request, the " + "script chooses a LoRA module at random.", + ) + + parser.add_argument( + "--ramp-up-strategy", + type=str, + default=None, + choices=["linear", "exponential"], + help="The ramp-up strategy. This would be used to " + "ramp up the request rate from initial RPS to final " + "RPS rate (specified by --ramp-up-start-rps and --ramp-up-end-rps). " + "over the duration of the benchmark.", + ) + parser.add_argument( + "--ramp-up-start-rps", + type=int, + default=None, + help="The starting request rate for ramp-up (RPS). " + "Needs to be specified when --ramp-up-strategy is used.", + ) + parser.add_argument( + "--ramp-up-end-rps", + type=int, + default=None, + help="The ending request rate for ramp-up (RPS). " + "Needs to be specified when --ramp-up-strategy is used.", + ) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/offline_benchmark_test/benchmarks/benchmark_serving_structured_output.py b/offline_benchmark_test/benchmarks/benchmark_serving_structured_output.py new file mode 100644 index 0000000000000000000000000000000000000000..e23a5a9e2233df685998fbae0ae8cc8910b572cf --- /dev/null +++ b/offline_benchmark_test/benchmarks/benchmark_serving_structured_output.py @@ -0,0 +1,1043 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +r"""Benchmark online serving throughput with structured outputs. + +On the server side, run one of the following commands: + (vLLM OpenAI API server) + vllm serve --disable-log-requests + +On the client side, run: + python benchmarks/benchmark_serving_structured_output.py \ + --backend \ + --model \ + --dataset json \ + --structured-output-ratio 1.0 \ + --request-rate 10 \ + --num-prompts 1000 + + when using tgi backend, add + --endpoint /generate_stream + to the end of the command above. +""" + +import argparse +import asyncio +import copy +import dataclasses +import json +import os +import random +import time +import uuid +import warnings +from collections.abc import AsyncGenerator +from dataclasses import dataclass +from typing import Optional + +import datasets +import numpy as np +import pandas as pd +from tqdm.asyncio import tqdm +from transformers import PreTrainedTokenizerBase + +from backend_request_func import ( + ASYNC_REQUEST_FUNCS, + RequestFuncInput, + RequestFuncOutput, +) + +try: + from vllm.transformers_utils.tokenizer import get_tokenizer +except ImportError: + from backend_request_func import get_tokenizer + +try: + from vllm.utils import FlexibleArgumentParser +except ImportError: + from argparse import ArgumentParser as FlexibleArgumentParser + +from vllm.v1.structured_output.backend_xgrammar import ( + has_xgrammar_unsupported_json_features, +) + +MILLISECONDS_TO_SECONDS_CONVERSION = 1000 + + +@dataclass +class BenchmarkMetrics: + completed: int + total_input: int + total_output: int + request_throughput: float + request_goodput: float + output_throughput: float + total_token_throughput: float + mean_ttft_ms: float + median_ttft_ms: float + std_ttft_ms: float + percentiles_ttft_ms: list[tuple[float, float]] + mean_tpot_ms: float + median_tpot_ms: float + std_tpot_ms: float + percentiles_tpot_ms: list[tuple[float, float]] + mean_itl_ms: float + median_itl_ms: float + std_itl_ms: float + percentiles_itl_ms: list[tuple[float, float]] + # E2EL stands for end-to-end latency per request. + # It is the time taken on the client side from sending + # a request to receiving a complete response. + mean_e2el_ms: float + median_e2el_ms: float + std_e2el_ms: float + percentiles_e2el_ms: list[tuple[float, float]] + + +@dataclasses.dataclass +class SampleRequest: + """A class representing a single inference request for benchmarking. + + Attributes: + prompt: The input text prompt for the model. + multi_modal_data: Optional dictionary containing multi-modal data (e.g. + images). + prompt_len: The length of the prompt in tokens. + expected_output_len: The expected length of the output in tokens. + """ + + prompt: str + prompt_len: int + expected_output_len: int + schema: dict + structure_type: str + completion: str = None + + +def sample_requests( + tokenizer: PreTrainedTokenizerBase, args: argparse.Namespace +) -> list[SampleRequest]: + if args.dataset == "json" or args.dataset == "json-unique": + if args.json_schema_path is None: + dir_path = os.path.dirname(os.path.realpath(__file__)) + args.json_schema_path = os.path.join( + dir_path, "structured_schemas", "structured_schema_1.json" + ) + json_schemas = [] + with open(args.json_schema_path) as f: + schema = json.load(f) + + if args.dataset == "json-unique": + json_schemas = [copy.deepcopy(schema) for _ in range(args.num_prompts)] + for i in range(len(json_schemas)): + if "properties" not in json_schemas[i]: + json_schemas[i]["properties"] = {} + json_schemas[i]["properties"][f"__optional_field_{uuid.uuid4()}"] = { + "type": "string", + "description": "An unique optional field to avoid cached schemas", + } + else: + json_schemas = [schema] * args.num_prompts + + def gen_prompt(index: int): + return f"Generate an example of a brief user profile given the following schema: {json.dumps(get_schema(index))}" # noqa: E501 + + def get_schema(index: int): + return json_schemas[index % len(json_schemas)] + + requests = [ + SampleRequest( + prompt=gen_prompt(i), + prompt_len=len(tokenizer(gen_prompt(i)).input_ids), + expected_output_len=args.output_len, + schema=get_schema(i), + structure_type=args.structure_type, + ) + for i in range(args.num_prompts) + ] + + elif args.dataset == "grammar": + schema = """ + root ::= select_statement + + select_statement ::= "SELECT " column " from " table " where " condition + + column ::= "col_1 " | "col_2 " + + table ::= "table_1 " | "table_2 " + + condition ::= column "= " number + + number ::= "1 " | "2 " + """ + prompt = "Generate an SQL query to show the 'username' \ + and 'email' from the 'users' table." + + input_len = len(tokenizer(prompt).input_ids) + print(f"Input length of the prompt: {input_len} tokens") + requests = [ + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=schema, + structure_type=args.structure_type, + ) + for _ in range(args.num_prompts) + ] + + elif args.dataset == "regex": + regex = r"\w+@\w+\.com\n" + args.regex = regex + prompt = "Generate an email address for Alan Turing, \ + who works in Enigma. End in .com and new line. \ + Example result: alan.turing@enigma.com\n" + + input_len = len(tokenizer(prompt).input_ids) + print(f"Input length of the prompt: {input_len} tokens") + requests = [ + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=regex, + structure_type=args.structure_type, + ) + for _ in range(args.num_prompts) + ] + + elif args.dataset == "choice": + choice = ["Positive", "Negative"] + args.choice = choice + prompt = "Classify this sentiment: vLLM is wonderful!" + input_len = len(tokenizer(prompt).input_ids) + print(f"Input length of the prompt: {input_len} tokens") + requests = [ + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=choice, + structure_type=args.structure_type, + ) + for _ in range(args.num_prompts) + ] + + elif args.dataset == "xgrammar_bench": + requests: list[SampleRequest] = [] + dataset = datasets.load_dataset("NousResearch/json-mode-eval", split="train") + full_dataset_len = len(dataset) + + def _filter_func(item): + import json + + schema = json.loads(item["schema"]) + return not has_xgrammar_unsupported_json_features(schema) + + dataset = dataset.filter(_filter_func) + num_filtered_out = full_dataset_len - len(dataset) + print( + f"dataset has {len(dataset)} entries after filtering " + f"out {num_filtered_out} entries with unsupported features" + ) + len_dataset = len(dataset) + for data_point_idx in range(args.num_prompts): + idx = data_point_idx + while idx >= len_dataset: + idx -= len_dataset + schema = dataset["schema"][idx] + prompt = tokenizer.apply_chat_template( + dataset["prompt"][idx], tokenize=False, add_generation_prompt=True + ) + input_len = len(tokenizer(prompt).input_ids) + completion = dataset["completion"][idx] + + requests.append( + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=schema, + structure_type=args.structure_type, + completion=completion, + ) + ) + + return requests + + +async def get_request( + input_requests: list[SampleRequest], + request_rate: float, + burstiness: float = 1.0, +) -> AsyncGenerator[tuple[int, SampleRequest], None]: + """ + Asynchronously generates requests at a specified rate + with OPTIONAL burstiness. + + Args: + input_requests: + A list of input requests, each represented as a tuple. + request_rate: + The rate at which requests are generated (requests/s). + burstiness (optional): + The burstiness factor of the request generation. + Only takes effect when request_rate is not inf. + Default value is 1, which follows a Poisson process. + Otherwise, the request intervals follow a gamma distribution. + A lower burstiness value (0 < burstiness < 1) results + in more bursty requests, while a higher burstiness value + (burstiness > 1) results in a more uniform arrival of requests. + """ + input_requests = iter(input_requests) + + # Calculate scale parameter theta to maintain the desired request_rate. + assert burstiness > 0, ( + f"A positive burstiness factor is expected, but given {burstiness}." + ) + theta = 1.0 / (request_rate * burstiness) + + for i, request in enumerate(input_requests): + yield i, request + + if request_rate == float("inf"): + # If the request rate is infinity, then we don't need to wait. + continue + + # Sample the request interval from the gamma distribution. + # If burstiness is 1, it follows exponential distribution. + interval = np.random.gamma(shape=burstiness, scale=theta) + # The next request will be sent after the interval. + await asyncio.sleep(interval) + + +def calculate_metrics( + input_requests: list[tuple[str, int, int]], + outputs: list[RequestFuncOutput], + dur_s: float, + tokenizer: PreTrainedTokenizerBase, + selected_percentile_metrics: list[str], + selected_percentiles: list[float], + goodput_config_dict: Optional[dict[str, float]] = None, +) -> tuple[BenchmarkMetrics, list[int]]: + actual_output_lens: list[int] = [] + total_input = 0 + completed = 0 + good_completed = 0 + itls: list[float] = [] + tpots: list[float] = [] + all_tpots: list[float] = [] + ttfts: list[float] = [] + e2els: list[float] = [] + for i in range(len(outputs)): + if outputs[i].success: + # We use the tokenizer to count the number of output tokens for all + # serving backends instead of looking at len(outputs[i].itl) since + # multiple output tokens may be bundled together + # Note : this may inflate the output token count slightly + output_len = len( + tokenizer(outputs[i].generated_text, add_special_tokens=False).input_ids + ) + actual_output_lens.append(output_len) + total_input += input_requests[i].prompt_len + tpot = 0 + if output_len > 1: + latency_minus_ttft = outputs[i].latency - outputs[i].ttft + tpot = latency_minus_ttft / (output_len - 1) + tpots.append(tpot) + outputs[i].tpot = tpot + # Note: if output_len <= 1, we regard tpot as 0 for goodput + all_tpots.append(tpot) + itls += outputs[i].itl + ttfts.append(outputs[i].ttft) + e2els.append(outputs[i].latency) + completed += 1 + else: + actual_output_lens.append(0) + + if goodput_config_dict: + valid_metrics = [] + slo_values = [] + + if "ttft" in goodput_config_dict: + valid_metrics.append(ttfts) + slo_values.append( + goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "tpot" in goodput_config_dict: + valid_metrics.append(all_tpots) + slo_values.append( + goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "e2el" in goodput_config_dict: + valid_metrics.append(e2els) + slo_values.append( + goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + + for req_metric in zip(*valid_metrics): + is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)]) + if is_good_req: + good_completed += 1 + + if completed == 0: + warnings.warn( + "All requests failed. This is likely due to a misconfiguration " + "on the benchmark arguments.", + stacklevel=2, + ) + metrics = BenchmarkMetrics( + completed=completed, + total_input=total_input, + total_output=sum(actual_output_lens), + request_throughput=completed / dur_s, + request_goodput=good_completed / dur_s, + output_throughput=sum(actual_output_lens) / dur_s, + total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s, + mean_ttft_ms=np.mean(ttfts or 0) + * 1000, # ttfts is empty if streaming is not supported by backend + std_ttft_ms=np.std(ttfts or 0) * 1000, + median_ttft_ms=np.median(ttfts or 0) * 1000, + percentiles_ttft_ms=[ + (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles + ], + mean_tpot_ms=np.mean(tpots or 0) * 1000, + std_tpot_ms=np.std(tpots or 0) * 1000, + median_tpot_ms=np.median(tpots or 0) * 1000, + percentiles_tpot_ms=[ + (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles + ], + mean_itl_ms=np.mean(itls or 0) * 1000, + std_itl_ms=np.std(itls or 0) * 1000, + median_itl_ms=np.median(itls or 0) * 1000, + percentiles_itl_ms=[ + (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles + ], + mean_e2el_ms=np.mean(e2els or 0) * 1000, + std_e2el_ms=np.std(e2els or 0) * 1000, + median_e2el_ms=np.median(e2els or 0) * 1000, + percentiles_e2el_ms=[ + (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles + ], + ) + + return metrics, actual_output_lens + + +async def benchmark( + backend: str, + api_url: str, + base_url: str, + model_id: str, + tokenizer: PreTrainedTokenizerBase, + input_requests: list[SampleRequest], + request_rate: float, + burstiness: float, + disable_tqdm: bool, + profile: bool, + selected_percentile_metrics: list[str], + selected_percentiles: list[str], + ignore_eos: bool, + max_concurrency: Optional[int], + structured_output_ratio: float, + goodput_config_dict: Optional[dict[str, float]] = None, +): + if backend in ASYNC_REQUEST_FUNCS: + request_func = ASYNC_REQUEST_FUNCS[backend] + else: + raise ValueError(f"Unknown backend: {backend}") + + def prepare_extra_body(request) -> dict: + extra_body = {} + # Add the schema to the extra_body + extra_body[request.structure_type] = request.schema + return extra_body + + print("Starting initial single prompt test run...") + structured_output_req_idx = random.sample( + range(len(input_requests)), int(len(input_requests) * structured_output_ratio) + ) + + test_request = input_requests[0] + test_req_extra_body = ( + prepare_extra_body(test_request) if 0 in structured_output_req_idx else None + ) + test_input = RequestFuncInput( + model=model_id, + prompt=test_request.prompt, + api_url=api_url, + prompt_len=test_request.prompt_len, + output_len=test_request.expected_output_len, + ignore_eos=ignore_eos, + extra_body=test_req_extra_body, + ) + test_output = await request_func(request_func_input=test_input) + if not test_output.success: + raise ValueError( + "Initial test run failed - Please make sure benchmark arguments " + f"are correctly specified. Error: {test_output.error}" + ) + else: + print("Initial test run completed. Starting main benchmark run...") + + if profile: + print("Starting profiler...") + profile_input = RequestFuncInput( + model=model_id, + prompt=test_request.prompt, + api_url=base_url + "/start_profile", + prompt_len=test_request.prompt_len, + output_len=test_request.expected_output_len, + ignore_eos=ignore_eos, + extra_body=test_req_extra_body, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler started") + + distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution" + + print(f"Traffic request rate: {request_rate}") + print(f"Burstiness factor: {burstiness} ({distribution})") + print(f"Maximum request concurrency: {max_concurrency}") + + pbar = None if disable_tqdm else tqdm(total=len(input_requests)) + + # This can be used once the minimum Python version is 3.10 or higher, + # and it will simplify the code in limited_request_func. + # semaphore = (asyncio.Semaphore(max_concurrency) + # if max_concurrency else contextlib.nullcontext()) + semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None + + async def limited_request_func(request_func_input, pbar): + if semaphore is None: + return await request_func(request_func_input=request_func_input, pbar=pbar) + async with semaphore: + return await request_func(request_func_input=request_func_input, pbar=pbar) + + benchmark_start_time = time.perf_counter() + tasks: list[asyncio.Task] = [] + expected: list[str] = [] + async for i, request in get_request(input_requests, request_rate, burstiness): + extra_body = ( + prepare_extra_body(request) if i in structured_output_req_idx else None + ) + request_func_input = RequestFuncInput( + model=model_id, + prompt=request.prompt, + api_url=api_url, + prompt_len=request.prompt_len, + output_len=request.expected_output_len, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + expected.append(request.completion) + tasks.append( + asyncio.create_task( + limited_request_func(request_func_input=request_func_input, pbar=pbar) + ) + ) + outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks) + + if profile: + print("Stopping profiler...") + profile_input = RequestFuncInput( + model=model_id, + prompt=test_request.prompt, + api_url=base_url + "/stop_profile", + prompt_len=test_request.prompt_len, + output_len=test_request.expected_output_len, + extra_body={test_request.structure_type: test_request.schema}, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler stopped") + + if pbar is not None: + pbar.close() + + benchmark_duration = time.perf_counter() - benchmark_start_time + + metrics, actual_output_lens = calculate_metrics( + input_requests=input_requests, + outputs=outputs, + dur_s=benchmark_duration, + tokenizer=tokenizer, + selected_percentile_metrics=selected_percentile_metrics, + selected_percentiles=selected_percentiles, + goodput_config_dict=goodput_config_dict, + ) + + print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="=")) + print("{:<40} {:<10}".format("Successful requests:", metrics.completed)) + print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration)) + print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input)) + print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output)) + print( + "{:<40} {:<10.2f}".format( + "Request throughput (req/s):", metrics.request_throughput + ) + ) + if goodput_config_dict: + print( + "{:<40} {:<10.2f}".format( + "Request goodput (req/s):", metrics.request_goodput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Output token throughput (tok/s):", metrics.output_throughput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Total Token throughput (tok/s):", metrics.total_token_throughput + ) + ) + + result = { + "duration": benchmark_duration, + "completed": metrics.completed, + "total_input_tokens": metrics.total_input, + "total_output_tokens": metrics.total_output, + "request_throughput": metrics.request_throughput, + "output_throughput": metrics.output_throughput, + "total_token_throughput": metrics.total_token_throughput, + "ttft_description": pd.Series([output.ttft for output in outputs]) + .describe() + .to_dict(), + "tpot_description": pd.Series([output.tpot for output in outputs]) + .describe() + .to_dict(), + "input_lens": [output.prompt_len for output in outputs], + "output_lens": actual_output_lens, + "ttfts": [output.ttft for output in outputs], + "itls": [output.itl for output in outputs], + "errors": [output.error for output in outputs], + } + + ret = [ + {"generated": output.generated_text, "expected": gt} + for output, gt in zip(outputs, expected) + ] + + def process_one_metric( + # E.g., "ttft" + metric_attribute_name: str, + # E.g., "TTFT" + metric_name: str, + # E.g., "Time to First Token" + metric_header: str, + ): + # This function prints and adds statistics of the specified + # metric. + if metric_attribute_name not in selected_percentile_metrics: + return + print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-")) + print( + "{:<40} {:<10.2f}".format( + f"Mean {metric_name} (ms):", + getattr(metrics, f"mean_{metric_attribute_name}_ms"), + ) + ) + print( + "{:<40} {:<10.2f}".format( + f"Median {metric_name} (ms):", + getattr(metrics, f"median_{metric_attribute_name}_ms"), + ) + ) + result[f"mean_{metric_attribute_name}_ms"] = getattr( + metrics, f"mean_{metric_attribute_name}_ms" + ) + result[f"median_{metric_attribute_name}_ms"] = getattr( + metrics, f"median_{metric_attribute_name}_ms" + ) + result[f"std_{metric_attribute_name}_ms"] = getattr( + metrics, f"std_{metric_attribute_name}_ms" + ) + for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"): + p_word = str(int(p)) if int(p) == p else str(p) + print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value)) + result[f"p{p_word}_{metric_attribute_name}_ms"] = value + + process_one_metric("ttft", "TTFT", "Time to First Token") + process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)") + process_one_metric("itl", "ITL", "Inter-token Latency") + process_one_metric("e2el", "E2EL", "End-to-end Latency") + + print("=" * 50) + + return result, ret + + +def evaluate(ret, args): + def _eval_correctness_json(expected, actual): + # extract json string from string using regex + import regex as re + + actual = actual.replace("\n", "").replace(" ", "").strip() + try: + actual = re.search(r"\{.*\}", actual).group() + actual = json.loads(actual) + except Exception: + return False + + return True + + def _eval_correctness_choice(expected, actual): + return actual in args.choice + + def _eval_correctness_regex(expected, actual): + import regex as re + + return re.match(args.regex, actual) is not None + + def _eval_correctness(expected, actual): + if args.structure_type == "guided_json": + return _eval_correctness_json(expected, actual) + elif args.structure_type == "guided_regex": + return _eval_correctness_regex(expected, actual) + elif args.structure_type == "guided_choice": + return _eval_correctness_choice(expected, actual) + else: + return None + + scores = [] + for res in ret: + score = _eval_correctness(res["expected"], res["generated"]) + res["correctness"] = score + scores.append(score) + + not_none_scores = [score for score in scores if score is not None] + + return ( + (sum(not_none_scores) / len(not_none_scores) * 100) + if len(not_none_scores) > 0 + else None + ) + + +def parse_goodput(slo_pairs): + goodput_config_dict = {} + try: + for slo_pair in slo_pairs: + slo_name, slo_val = slo_pair.split(":") + goodput_config_dict[slo_name] = float(slo_val) + except ValueError as err: + raise argparse.ArgumentTypeError( + "Invalid format found for service level objectives. " + 'Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is a " + "number in milliseconds." + ) from err + return goodput_config_dict + + +def check_goodput_args(args): + goodput_config_dict = {} + VALID_NAMES = ["ttft", "tpot", "e2el"] + if args.goodput: + goodput_config_dict = parse_goodput(args.goodput) + for slo_name, slo_val in goodput_config_dict.items(): + if slo_name not in VALID_NAMES: + raise ValueError( + f"Invalid metric name found, {slo_name}: {slo_val}. " + "The service level objective name should be one of " + f"{str(VALID_NAMES)}. " + ) + if slo_val < 0: + raise ValueError( + f"Invalid value found, {slo_name}: {slo_val}. " + "The service level objective value should be " + "non-negative." + ) + return goodput_config_dict + + +def main(args: argparse.Namespace): + print(args) + random.seed(args.seed) + np.random.seed(args.seed) + + backend = args.backend + model_id = args.model + tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model + + if args.base_url is not None: + api_url = f"{args.base_url}{args.endpoint}" + base_url = f"{args.base_url}" + else: + api_url = f"http://{args.host}:{args.port}{args.endpoint}" + base_url = f"http://{args.host}:{args.port}" + + tokenizer = get_tokenizer( + tokenizer_id, + trust_remote_code=args.trust_remote_code, + tokenizer_mode=args.tokenizer_mode, + ) + + if args.dataset == "grammar": + args.structure_type = "guided_grammar" + elif args.dataset == "regex": + args.structure_type = "guided_regex" + elif args.dataset == "choice": + args.structure_type = "guided_choice" + else: + args.structure_type = "guided_json" + + if args.no_structured_output: + args.structured_output_ratio = 0 + if args.save_results: + result_file_name = f"{args.structured_output_ratio}guided" + result_file_name += f"_{backend}" + result_file_name += f"_{args.request_rate}qps" + result_file_name += f"_{args.model.split('/')[-1]}" + result_file_name += f"_{args.dataset}" + result_file_name += f"_{args.num_prompts}" + result_file_name += f"_out{args.output_len}" + result_file_name += ".txt" + else: + result_file_name = None + + input_requests = sample_requests(tokenizer, args) + + goodput_config_dict = check_goodput_args(args) + + benchmark_result, ret = asyncio.run( + benchmark( + backend=backend, + api_url=api_url, + base_url=base_url, + model_id=model_id, + tokenizer=tokenizer, + input_requests=input_requests, + request_rate=args.request_rate, + burstiness=args.burstiness, + disable_tqdm=args.disable_tqdm, + profile=args.profile, + selected_percentile_metrics=args.percentile_metrics.split(","), + selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")], + ignore_eos=args.ignore_eos, + max_concurrency=args.max_concurrency, + structured_output_ratio=args.structured_output_ratio, + goodput_config_dict=goodput_config_dict, + ) + ) + + # Save config and results to json + score = evaluate(ret, args) + print("correct_rate(%)", score, "\n") + if args.save_results: + results = { + "backend": backend, + "model_id": model_id, + "tokenizer_id": tokenizer_id, + "num_prompts": args.num_prompts, + "request_rate": args.request_rate + if args.request_rate < float("inf") + else "inf", + "burstiness": args.burstiness, + "max_concurrency": args.max_concurrency, + "correct_rate(%)": score, + } + results = {"outputs": ret, **results, **benchmark_result} + + # Save to file + if args.result_filename: + result_file_name = args.result_filename + if args.result_dir: + result_file_name = os.path.join(args.result_dir, result_file_name) + with open(result_file_name, "w", encoding="utf-8") as outfile: + json.dump(results, outfile, indent=4) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the online serving throughput." + ) + parser.add_argument( + "--backend", + type=str, + default="vllm", + choices=list(ASYNC_REQUEST_FUNCS.keys()), + ) + parser.add_argument( + "--base-url", + type=str, + default=None, + help="Server or API base url if not using http host and port.", + ) + # Use 127.0.0.1 here instead of localhost to force the use of ipv4 + parser.add_argument("--host", type=str, default="127.0.0.1") + parser.add_argument("--port", type=int, default=8000) + parser.add_argument( + "--endpoint", + type=str, + default="/v1/completions", + help="API endpoint.", + ) + parser.add_argument( + "--dataset", + default="json", + choices=["json", "json-unique", "grammar", "regex", "choice", "xgrammar_bench"], + ) + parser.add_argument( + "--json-schema-path", type=str, default=None, help="Path to json schema." + ) + parser.add_argument( + "--max-concurrency", + type=int, + default=None, + help="Maximum number of concurrent requests. This can be used " + "to help simulate an environment where a higher level component " + "is enforcing a maximum number of concurrent requests. While the " + "--request-rate argument controls the rate at which requests are " + "initiated, this argument will control how many are actually allowed " + "to execute at a time. This means that when used in combination, the " + "actual request rate may be lower than specified with --request-rate, " + "if the server is not processing requests fast enough to keep up.", + ) + parser.add_argument( + "--model", + type=str, + required=True, + help="Name of the model.", + ) + parser.add_argument( + "--tokenizer", + type=str, + help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 + ) + parser.add_argument( + "--tokenizer-mode", + type=str, + default="auto", + help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 + ) + parser.add_argument( + "--num-prompts", + type=int, + default=1000, + help="Number of prompts to process.", + ) + parser.add_argument( + "--output-len", + type=int, + default=128, + help="Number of output tokens.", + ) + parser.add_argument( + "--request-rate", + type=float, + default=float("inf"), + help="Number of requests per second. If this is inf, " + "then all the requests are sent at time 0. " + "Otherwise, we use Poisson process or gamma distribution " + "to synthesize the request arrival times.", + ) + parser.add_argument( + "--burstiness", + type=float, + default=1.0, + help="Burstiness factor of the request generation. " + "Only take effect when request_rate is not inf. " + "Default value is 1, which follows Poisson process. " + "Otherwise, the request intervals follow a gamma distribution. " + "A lower burstiness value (0 < burstiness < 1) results in more " + "bursty requests. A higher burstiness value (burstiness > 1) " + "results in a more uniform arrival of requests.", + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument( + "--trust-remote-code", + action="store_true", + help="Trust remote code from huggingface", + ) + parser.add_argument( + "--disable-tqdm", + action="store_true", + help="Specify to disable tqdm progress bar.", + ) + parser.add_argument( + "--save-results", + action="store_true", + help="Specify to save benchmark results to a json file", + ) + parser.add_argument( + "--profile", + action="store_true", + help="Use Torch Profiler. The endpoint must be launched with " + "VLLM_TORCH_PROFILER_DIR to enable profiler.", + ) + parser.add_argument( + "--result-dir", + type=str, + default=None, + help="Specify directory to save benchmark json results." + "If not specified, results are saved in the current directory.", + ) + parser.add_argument( + "--result-filename", + type=str, + default=None, + help="Specify the filename to save benchmark json results." + "If not specified, results will be saved in " + "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json" + " format.", + ) + parser.add_argument( + "--ignore-eos", + action="store_true", + help="Set ignore_eos flag when sending the benchmark request." + "Warning: ignore_eos is not supported in deepspeed_mii and tgi.", + ) + parser.add_argument( + "--percentile-metrics", + type=str, + default="ttft,tpot,itl", + help="Comma-separated list of selected metrics to report percentils. " + "This argument specifies the metrics to report percentiles. " + 'Allowed metric names are "ttft", "tpot", "itl", "e2el". ' + 'Default value is "ttft,tpot,itl".', + ) + parser.add_argument( + "--metric-percentiles", + type=str, + default="99", + help="Comma-separated list of percentiles for selected metrics. " + 'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". ' + 'Default value is "99". ' + 'Use "--percentile-metrics" to select metrics.', + ) + parser.add_argument( + "--goodput", + nargs="+", + required=False, + help='Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is in " + 'milliseconds. Multiple "KEY:VALUE" pairs can be provided, ' + "separated by spaces. Allowed request level metric names are " + '"ttft", "tpot", "e2el". For more context on the definition of ' + "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 " + "and the blog: https://hao-ai-lab.github.io/blogs/distserve", + ) + + parser.add_argument( + "--no-structured-output", + action="store_true", + default=False, + help="Whether to disable JSON decoding or not.", + ) + parser.add_argument( + "--structured-output-ratio", + type=float, + default=1.0, + help="Ratio of Structured Outputs requests", + ) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/offline_benchmark_test/benchmarks/benchmark_throughput.py b/offline_benchmark_test/benchmarks/benchmark_throughput.py new file mode 100644 index 0000000000000000000000000000000000000000..26c7dcd0bb08c665a98caabd555dc2e665c28027 --- /dev/null +++ b/offline_benchmark_test/benchmarks/benchmark_throughput.py @@ -0,0 +1,757 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Benchmark offline inference throughput.""" + +import argparse +import dataclasses +import json +import os +import random +import time +import warnings +from typing import Any, Optional, Union + +import torch +import uvloop +from tqdm import tqdm +from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase + +from benchmark_dataset import ( + AIMODataset, + BurstGPTDataset, + ConversationDataset, + InstructCoderDataset, + RandomDataset, + SampleRequest, + ShareGPTDataset, + SonnetDataset, + VisionArenaDataset, +) +from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json +from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs +from vllm.entrypoints.openai.api_server import ( + build_async_engine_client_from_engine_args, +) +from vllm.inputs import TextPrompt, TokensPrompt +from vllm.lora.request import LoRARequest +from vllm.outputs import RequestOutput +from vllm.sampling_params import BeamSearchParams +from vllm.utils import FlexibleArgumentParser, merge_async_iterators + + +def run_vllm( + requests: list[SampleRequest], + n: int, + engine_args: EngineArgs, + disable_detokenize: bool = False, +) -> tuple[float, Optional[list[RequestOutput]]]: + from vllm import LLM, SamplingParams + + llm = LLM(**dataclasses.asdict(engine_args)) + assert all( + llm.llm_engine.model_config.max_model_len + >= (request.prompt_len + request.expected_output_len) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of" + " prompt_len and expected_output_len for all requests." + ) + # Add the requests to the engine. + prompts: list[Union[TextPrompt, TokensPrompt]] = [] + sampling_params: list[SamplingParams] = [] + for request in requests: + prompts.append( + TokensPrompt( + prompt_token_ids=request.prompt["prompt_token_ids"], + multi_modal_data=request.multi_modal_data, + ) + if "prompt_token_ids" in request.prompt + else TextPrompt( + prompt=request.prompt, multi_modal_data=request.multi_modal_data + ) + ) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=request.expected_output_len, + detokenize=not disable_detokenize, + ) + ) + lora_requests: Optional[list[LoRARequest]] = None + if engine_args.enable_lora: + lora_requests = [request.lora_request for request in requests] + + use_beam_search = False + + outputs = None + if not use_beam_search: + start = time.perf_counter() + outputs = llm.generate( + prompts, sampling_params, lora_request=lora_requests, use_tqdm=True + ) + end = time.perf_counter() + + total_ttfts = [] + total_tpots = [] + for output in outputs: + if output.metrics.first_token_time == None: + pass + else: + ttft_ = output.metrics.first_token_time - output.metrics.arrival_time + tpot_ = (output.metrics.finished_time - output.metrics.arrival_time - ttft_) / (args.output_len-1) + total_ttfts.append(ttft_) + total_tpots.append(tpot_) + + import numpy as np + ttft_mean = np.mean(total_ttfts) + ttft_max = np.max(total_ttfts) + ttft_min = np.min(total_ttfts) + + tpot_mean = np.mean(total_tpots) + tpot_min = np.min(total_tpots) + tpot_max = np.max(total_tpots) + else: + assert lora_requests is None, "BeamSearch API does not support LoRA" + prompts = [request.prompt for request in requests] + # output_len should be the same for all requests. + output_len = requests[0].expected_output_len + for request in requests: + assert request.expected_output_len == output_len + start = time.perf_counter() + llm.beam_search( + prompts, + BeamSearchParams( + beam_width=n, + max_tokens=output_len, + ignore_eos=True, + ), + ) + end = time.perf_counter() + + return end - start, ttft_mean, tpot_mean, outputs + # return end - start, outputs + + +def run_vllm_chat( + requests: list[SampleRequest], + n: int, + engine_args: EngineArgs, + disable_detokenize: bool = False, +) -> tuple[float, list[RequestOutput]]: + """ + Run vLLM chat benchmark. This function is recommended ONLY for benchmarking + multimodal models as it properly handles multimodal inputs and chat + formatting. For non-multimodal models, use run_vllm() instead. + """ + from vllm import LLM, SamplingParams + + llm = LLM(**dataclasses.asdict(engine_args)) + + assert all( + llm.llm_engine.model_config.max_model_len + >= (request.prompt_len + request.expected_output_len) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of " + "prompt_len and expected_output_len for all requests." + ) + + prompts = [] + sampling_params: list[SamplingParams] = [] + for request in requests: + prompts.append(request.prompt) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=request.expected_output_len, + detokenize=not disable_detokenize, + ) + ) + start = time.perf_counter() + outputs = llm.chat(prompts, sampling_params, use_tqdm=True) + end = time.perf_counter() + return end - start, outputs + + +async def run_vllm_async( + requests: list[SampleRequest], + n: int, + engine_args: AsyncEngineArgs, + disable_frontend_multiprocessing: bool = False, + disable_detokenize: bool = False, +) -> float: + from vllm import SamplingParams + + async with build_async_engine_client_from_engine_args( + engine_args, disable_frontend_multiprocessing + ) as llm: + model_config = await llm.get_model_config() + assert all( + model_config.max_model_len + >= (request.prompt_len + request.expected_output_len) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of" + " prompt_len and expected_output_len for all requests." + ) + + # Add the requests to the engine. + prompts: list[Union[TextPrompt, TokensPrompt]] = [] + sampling_params: list[SamplingParams] = [] + lora_requests: list[Optional[LoRARequest]] = [] + for request in requests: + prompts.append( + TokensPrompt( + prompt_token_ids=request.prompt["prompt_token_ids"], + multi_modal_data=request.multi_modal_data, + ) + if "prompt_token_ids" in request.prompt + else TextPrompt( + prompt=request.prompt, multi_modal_data=request.multi_modal_data + ) + ) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=request.expected_output_len, + detokenize=not disable_detokenize, + ) + ) + lora_requests.append(request.lora_request) + + generators = [] + start = time.perf_counter() + for i, (prompt, sp, lr) in enumerate( + zip(prompts, sampling_params, lora_requests) + ): + generator = llm.generate(prompt, sp, lora_request=lr, request_id=f"test{i}") + generators.append(generator) + all_gens = merge_async_iterators(*generators) + async for i, res in all_gens: + pass + end = time.perf_counter() + return end - start + + +def run_hf( + requests: list[SampleRequest], + model: str, + tokenizer: PreTrainedTokenizerBase, + n: int, + max_batch_size: int, + trust_remote_code: bool, + disable_detokenize: bool = False, +) -> float: + llm = AutoModelForCausalLM.from_pretrained( + model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code + ) + if llm.config.model_type == "llama": + # To enable padding in the HF backend. + tokenizer.pad_token = tokenizer.eos_token + llm = llm.cuda() + + pbar = tqdm(total=len(requests)) + start = time.perf_counter() + batch: list[str] = [] + max_prompt_len = 0 + max_output_len = 0 + for i in range(len(requests)): + prompt = requests[i].prompt + prompt_len = requests[i].prompt_len + output_len = requests[i].expected_output_len + # Add the prompt to the batch. + batch.append(prompt) + max_prompt_len = max(max_prompt_len, prompt_len) + max_output_len = max(max_output_len, output_len) + if len(batch) < max_batch_size and i != len(requests) - 1: + # Check if we can add more requests to the batch. + next_prompt_len = requests[i + 1].prompt_len + next_output_len = requests[i + 1].expected_output_len + if ( + max(max_prompt_len, next_prompt_len) + + max(max_output_len, next_output_len) + ) <= 2048: + # We can add more requests to the batch. + continue + + # Generate the sequences. + input_ids = tokenizer(batch, return_tensors="pt", padding=True).input_ids + llm_outputs = llm.generate( + input_ids=input_ids.cuda(), + do_sample=True, + num_return_sequences=n, + temperature=1.0, + top_p=1.0, + use_cache=True, + max_new_tokens=max_output_len, + ) + if not disable_detokenize: + # Include the decoding time. + tokenizer.batch_decode(llm_outputs, skip_special_tokens=True) + pbar.update(len(batch)) + + # Clear the batch. + batch = [] + max_prompt_len = 0 + max_output_len = 0 + end = time.perf_counter() + return end - start + + +def run_mii( + requests: list[SampleRequest], + model: str, + tensor_parallel_size: int, + output_len: int, +) -> float: + from mii import client, serve + + llm = serve(model, tensor_parallel=tensor_parallel_size) + prompts = [request.prompt for request in requests] + + start = time.perf_counter() + llm.generate(prompts, max_new_tokens=output_len) + end = time.perf_counter() + client = client(model) + client.terminate_server() + return end - start + + +def save_to_pytorch_benchmark_format( + args: argparse.Namespace, results: dict[str, Any] +) -> None: + pt_records = convert_to_pytorch_benchmark_format( + args=args, + metrics={ + "requests_per_second": [results["requests_per_second"]], + "tokens_per_second": [results["tokens_per_second"]], + }, + extra_info={ + k: results[k] for k in ["elapsed_time", "num_requests", "total_num_tokens"] + }, + ) + if pt_records: + # Don't use json suffix here as we don't want CI to pick it up + pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json" + write_to_json(pt_file, pt_records) + + +def get_requests(args, tokenizer): + # Common parameters for all dataset types. + common_kwargs = { + "dataset_path": args.dataset_path, + "random_seed": args.seed, + } + sample_kwargs = { + "tokenizer": tokenizer, + "lora_path": args.lora_path, + "max_loras": args.max_loras, + "num_requests": args.num_prompts, + "input_len": args.input_len, + "output_len": args.output_len, + } + + if args.dataset_path is None or args.dataset_name == "random": + sample_kwargs["range_ratio"] = args.random_range_ratio + sample_kwargs["prefix_len"] = args.prefix_len + dataset_cls = RandomDataset + elif args.dataset_name == "sharegpt": + dataset_cls = ShareGPTDataset + if args.backend == "vllm-chat": + sample_kwargs["enable_multimodal_chat"] = True + elif args.dataset_name == "sonnet": + assert tokenizer.chat_template or tokenizer.default_chat_template, ( + "Tokenizer/model must have chat template for sonnet dataset." + ) + dataset_cls = SonnetDataset + sample_kwargs["prefix_len"] = args.prefix_len + sample_kwargs["return_prompt_formatted"] = True + elif args.dataset_name == "burstgpt": + dataset_cls = BurstGPTDataset + elif args.dataset_name == "hf": + if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS: + dataset_cls = VisionArenaDataset + common_kwargs["dataset_subset"] = None + common_kwargs["dataset_split"] = "train" + sample_kwargs["enable_multimodal_chat"] = True + elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS: + dataset_cls = InstructCoderDataset + common_kwargs["dataset_split"] = "train" + elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS: + dataset_cls = ConversationDataset + common_kwargs["dataset_subset"] = args.hf_subset + common_kwargs["dataset_split"] = args.hf_split + sample_kwargs["enable_multimodal_chat"] = True + elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS: + dataset_cls = AIMODataset + common_kwargs["dataset_subset"] = None + common_kwargs["dataset_split"] = "train" + else: + raise ValueError(f"Unknown dataset name: {args.dataset_name}") + # Remove None values + sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None} + return dataset_cls(**common_kwargs).sample(**sample_kwargs) + + +def main(args: argparse.Namespace): + if args.seed is None: + args.seed = 0 + print(args) + random.seed(args.seed) + # Sample the requests. + tokenizer = AutoTokenizer.from_pretrained( + args.tokenizer, trust_remote_code=args.trust_remote_code + ) + requests = get_requests(args, tokenizer) + is_multi_modal = any(request.multi_modal_data is not None for request in requests) + request_outputs: Optional[list[RequestOutput]] = None + if args.backend == "vllm": + if args.async_engine: + elapsed_time = uvloop.run( + run_vllm_async( + requests, + args.n, + AsyncEngineArgs.from_cli_args(args), + args.disable_frontend_multiprocessing, + args.disable_detokenize, + ) + ) + else: + elapsed_time, ttft_mean, tpop_mean, request_outputs = run_vllm( + requests, + args.n, + EngineArgs.from_cli_args(args), + args.disable_detokenize, + ) + elif args.backend == "hf": + assert args.tensor_parallel_size == 1 + elapsed_time = run_hf( + requests, + args.model, + tokenizer, + args.n, + args.hf_max_batch_size, + args.trust_remote_code, + args.disable_detokenize, + ) + elif args.backend == "mii": + elapsed_time = run_mii( + requests, args.model, args.tensor_parallel_size, args.output_len + ) + elif args.backend == "vllm-chat": + elapsed_time, request_outputs = run_vllm_chat( + requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize + ) + else: + raise ValueError(f"Unknown backend: {args.backend}") + + if request_outputs: + # Note: with the vllm and vllm-chat backends, + # we have request_outputs, which we use to count tokens. + total_prompt_tokens = 0 + total_output_tokens = 0 + for ro in request_outputs: + if not isinstance(ro, RequestOutput): + continue + total_prompt_tokens += ( + len(ro.prompt_token_ids) if ro.prompt_token_ids else 0 + ) + total_output_tokens += sum(len(o.token_ids) for o in ro.outputs if o) + total_num_tokens = total_prompt_tokens + total_output_tokens + else: + total_num_tokens = sum(r.prompt_len + r.expected_output_len for r in requests) + total_output_tokens = sum(r.expected_output_len for r in requests) + total_prompt_tokens = total_num_tokens - total_output_tokens + + if is_multi_modal and args.backend != "vllm-chat": + print( + "\033[91mWARNING\033[0m: Multi-modal request with " + f"{args.backend} backend detected. The " + "following metrics are not accurate because image tokens are not" + " counted. See vllm-project/vllm/issues/9778 for details." + ) + # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length. + # vllm-chat backend counts the image tokens now + + print( + f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, " + f"{total_num_tokens / elapsed_time:.2f} total tokens/s, " + f"{total_output_tokens / elapsed_time:.2f} output tokens/s" + ) + # print(f"Total num prompt tokens: {total_prompt_tokens}") + # print(f"Total num output tokens: {total_output_tokens}") + + print(f"Generate Throughput: {total_output_tokens / elapsed_time:.2f} tokens/s") + print(f"Elapsed_time: {elapsed_time: .2f} s") + print(f"TTFT mean: {ttft_mean: .4f} s") + print(f"TPOT mean: {tpop_mean: .4f} s") + + # Output JSON results if specified + if args.output_json: + results = { + "elapsed_time": elapsed_time, + "num_requests": len(requests), + "total_num_tokens": total_num_tokens, + "requests_per_second": len(requests) / elapsed_time, + "tokens_per_second": total_num_tokens / elapsed_time, + } + with open(args.output_json, "w") as f: + json.dump(results, f, indent=4) + save_to_pytorch_benchmark_format(args, results) + + +def validate_args(args): + """ + Validate command-line arguments. + """ + + # === Deprecation and Defaulting === + if args.dataset is not None: + warnings.warn( + "The '--dataset' argument will be deprecated in the next release. " + "Please use '--dataset-name' and '--dataset-path' instead.", + stacklevel=2, + ) + args.dataset_path = args.dataset + + if not getattr(args, "tokenizer", None): + args.tokenizer = args.model + + # === Backend Validation === + valid_backends = {"vllm", "hf", "mii", "vllm-chat"} + if args.backend not in valid_backends: + raise ValueError(f"Unsupported backend: {args.backend}") + + # === Dataset Configuration === + if not args.dataset and not args.dataset_path: + print("When dataset path is not set, it will default to random dataset") + args.dataset_name = "random" + if args.input_len is None: + raise ValueError("input_len must be provided for a random dataset") + + # === Dataset Name Specific Checks === + # --hf-subset and --hf-split: only used + # when dataset_name is 'hf' + if args.dataset_name != "hf" and ( + getattr(args, "hf_subset", None) is not None + or getattr(args, "hf_split", None) is not None + ): + warnings.warn( + "--hf-subset and --hf-split will be ignored \ + since --dataset-name is not 'hf'.", + stacklevel=2, + ) + elif args.dataset_name == "hf": + if args.dataset_path in ( + VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys() + | ConversationDataset.SUPPORTED_DATASET_PATHS + ): + assert args.backend == "vllm-chat", ( + f"{args.dataset_path} needs to use vllm-chat as the backend." + ) # noqa: E501 + elif args.dataset_path in ( + InstructCoderDataset.SUPPORTED_DATASET_PATHS + | AIMODataset.SUPPORTED_DATASET_PATHS + ): + assert args.backend == "vllm", ( + f"{args.dataset_path} needs to use vllm as the backend." + ) # noqa: E501 + else: + raise ValueError(f"{args.dataset_path} is not supported by hf dataset.") + + # --random-range-ratio: only used when dataset_name is 'random' + if args.dataset_name != "random" and args.random_range_ratio is not None: + warnings.warn( + "--random-range-ratio will be ignored since \ + --dataset-name is not 'random'.", + stacklevel=2, + ) + + # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not + # set. + if ( + args.dataset_name not in {"random", "sonnet", None} + and args.prefix_len is not None + ): + warnings.warn( + "--prefix-len will be ignored since --dataset-name\ + is not 'random', 'sonnet', or not set.", + stacklevel=2, + ) + + # === LoRA Settings === + if getattr(args, "enable_lora", False) and args.backend != "vllm": + raise ValueError("LoRA benchmarking is only supported for vLLM backend") + if getattr(args, "enable_lora", False) and args.lora_path is None: + raise ValueError("LoRA path must be provided when enable_lora is True") + + # === Backend-specific Validations === + if args.backend == "hf" and args.hf_max_batch_size is None: + raise ValueError("HF max batch size is required for HF backend") + if args.backend != "hf" and args.hf_max_batch_size is not None: + raise ValueError("HF max batch size is only for HF backend.") + + if ( + args.backend in {"hf", "mii"} + and getattr(args, "quantization", None) is not None + ): + raise ValueError("Quantization is only for vLLM backend.") + + if args.backend == "mii" and args.dtype != "auto": + raise ValueError("dtype must be auto for MII backend.") + if args.backend == "mii" and args.n != 1: + raise ValueError("n must be 1 for MII backend.") + if args.backend == "mii" and args.tokenizer != args.model: + raise ValueError("Tokenizer must be the same as the model for MII backend.") + + # --data-parallel is not supported currently. + # https://github.com/vllm-project/vllm/issues/16222 + if args.data_parallel_size > 1: + raise ValueError( + "Data parallel is not supported in offline benchmark, \ + please use benchmark serving instead" + ) + + +def create_argument_parser(): + parser = FlexibleArgumentParser(description="Benchmark the throughput.") + parser.add_argument( + "--backend", + type=str, + choices=["vllm", "hf", "mii", "vllm-chat"], + default="vllm", + ) + parser.add_argument( + "--dataset-name", + type=str, + choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"], + help="Name of the dataset to benchmark on.", + default="sharegpt", + ) + parser.add_argument( + "--dataset", + type=str, + default=None, + help="Path to the ShareGPT dataset, will be deprecated in\ + the next release. The dataset is expected to " + "be a json in form of list[dict[..., conversations: " + "list[dict[..., value: ]]]]", + ) + parser.add_argument( + "--dataset-path", type=str, default=None, help="Path to the dataset" + ) + parser.add_argument( + "--input-len", + type=int, + default=None, + help="Input prompt length for each request", + ) + parser.add_argument( + "--output-len", + type=int, + default=None, + help="Output length for each request. Overrides the " + "output length from the dataset.", + ) + parser.add_argument( + "--n", type=int, default=1, help="Number of generated sequences per prompt." + ) + parser.add_argument( + "--num-prompts", type=int, default=1000, help="Number of prompts to process." + ) + parser.add_argument( + "--hf-max-batch-size", + type=int, + default=None, + help="Maximum batch size for HF backend.", + ) + parser.add_argument( + "--output-json", + type=str, + default=None, + help="Path to save the throughput results in JSON format.", + ) + parser.add_argument( + "--async-engine", + action="store_true", + default=False, + help="Use vLLM async engine rather than LLM class.", + ) + parser.add_argument( + "--disable-frontend-multiprocessing", + action="store_true", + default=False, + help="Disable decoupled async engine frontend.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize the response (i.e. do not include " + "detokenization time in the measurement)" + ), + ) + # LoRA + parser.add_argument( + "--lora-path", + type=str, + default=None, + help="Path to the LoRA adapters to use. This can be an absolute path, " + "a relative path, or a Hugging Face model identifier.", + ) + parser.add_argument( + "--prefix-len", + type=int, + default=None, + help=f"Number of prefix tokens to be used in RandomDataset " + "and SonnetDataset. For RandomDataset, the total input " + "length is the sum of prefix-len (default: " + f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length " + "sampled from [input_len * (1 - range_ratio), " + "input_len * (1 + range_ratio)]. For SonnetDataset, " + f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) " + "controls how much of the input is fixed lines versus " + "random lines, but the total input length remains approximately " + "input_len tokens.", + ) + # random dataset + parser.add_argument( + "--random-range-ratio", + type=float, + default=None, + help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) " + "for sampling input/output length, " + "used only for RandomDataset. Must be in the range [0, 1) to " + "define a symmetric sampling range " + "[length * (1 - range_ratio), length * (1 + range_ratio)].", + ) + + # hf dtaset + parser.add_argument( + "--hf-subset", type=str, default=None, help="Subset of the HF dataset." + ) + parser.add_argument( + "--hf-split", type=str, default=None, help="Split of the HF dataset." + ) + + parser = AsyncEngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + if args.tokenizer is None: + args.tokenizer = args.model + validate_args(args) + main(args) diff --git a/offline_benchmark_test/benchmarks/benchmark_utils.py b/offline_benchmark_test/benchmarks/benchmark_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..283f938df50af101fa2ff82f3c46e123b25e5524 --- /dev/null +++ b/offline_benchmark_test/benchmarks/benchmark_utils.py @@ -0,0 +1,74 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import json +import math +import os +from typing import Any + + +def convert_to_pytorch_benchmark_format( + args: argparse.Namespace, metrics: dict[str, list], extra_info: dict[str, Any] +) -> list: + """ + Save the benchmark results in the format used by PyTorch OSS benchmark with + on metric per record + https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database + """ + records = [] + if not os.environ.get("SAVE_TO_PYTORCH_BENCHMARK_FORMAT", False): + return records + + for name, benchmark_values in metrics.items(): + record = { + "benchmark": { + "name": "vLLM benchmark", + "extra_info": { + "args": vars(args), + }, + }, + "model": { + "name": args.model, + }, + "metric": { + "name": name, + "benchmark_values": benchmark_values, + "extra_info": extra_info, + }, + } + + tp = record["benchmark"]["extra_info"]["args"].get("tensor_parallel_size") + # Save tensor_parallel_size parameter if it's part of the metadata + if not tp and "tensor_parallel_size" in extra_info: + record["benchmark"]["extra_info"]["args"]["tensor_parallel_size"] = ( + extra_info["tensor_parallel_size"] + ) + + records.append(record) + + return records + + +class InfEncoder(json.JSONEncoder): + def clear_inf(self, o: Any): + if isinstance(o, dict): + return {k: self.clear_inf(v) for k, v in o.items()} + elif isinstance(o, list): + return [self.clear_inf(v) for v in o] + elif isinstance(o, float) and math.isinf(o): + return "inf" + return o + + def iterencode(self, o: Any, *args, **kwargs) -> Any: + return super().iterencode(self.clear_inf(o), *args, **kwargs) + + +def write_to_json(filename: str, records: list) -> None: + with open(filename, "w") as f: + json.dump( + records, + f, + cls=InfEncoder, + default=lambda o: f"<{type(o).__name__} object is not JSON serializable>", + ) diff --git a/offline_benchmark_test/benchmarks/cutlass_benchmarks/sparse_benchmarks.py b/offline_benchmark_test/benchmarks/cutlass_benchmarks/sparse_benchmarks.py new file mode 100644 index 0000000000000000000000000000000000000000..9ec270bbd2e988b459f99d876215984bf125d0f9 --- /dev/null +++ b/offline_benchmark_test/benchmarks/cutlass_benchmarks/sparse_benchmarks.py @@ -0,0 +1,516 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import itertools +import pickle as pkl +import time +from collections.abc import Iterable +from typing import Callable + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import make_rand_sparse_tensors +from weight_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512] +DEFAULT_TP_SIZES = [1] + + +# bench +def bench_fn( + label: str, sub_label: str, description: str, fn: Callable, *args, **kwargs +) -> TMeasurement: + min_run_time = 1 + + globals = { + "args": args, + "kwargs": kwargs, + "fn": fn, + } + return TBenchmark.Timer( + stmt="fn(*args, **kwargs)", + globals=globals, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + +def bench_int8( + dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str +) -> Iterable[TMeasurement]: + assert dtype == torch.int8 + b_compressed, e, a, b = make_rand_sparse_tensors(torch.int8, m, n, k) + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + + out = ops.cutlass_scaled_sparse_mm( + a, b_compressed, e, scale_a, scale_b, torch.bfloat16 + ) + out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16) + + if not torch.allclose(out, out_ref): + print("Incorrect results") + print(out) + print(out_ref) + else: + print("Correct results") + + timers = [] + # pytorch impl - bfloat16 + timers.append( + bench_fn( + label, + sub_label, + "pytorch_bf16_bf16_bf16_matmul-no-scales", + torch.mm, + a.to(dtype=torch.bfloat16), + b.to(dtype=torch.bfloat16), + ) + ) + + # pytorch impl - float16 + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp16_fp16_fp16_matmul-no-scales", + torch.mm, + a.to(dtype=torch.float16), + b.to(dtype=torch.float16), + ) + ) + + # cutlass impl + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_mm", + ops.cutlass_scaled_mm, + a, + b, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_mm_bias", + ops.cutlass_scaled_mm, + a, + b, + scale_a, + scale_b, + torch.bfloat16, + bias, + ) + ) + + # cutlass sparse impl + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_sparse_mm", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass sparse with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_sparse_mm_bias", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + bias, + ) + ) + + return timers + + +def bench_fp8( + dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str +) -> Iterable[TMeasurement]: + assert dtype == torch.float8_e4m3fn + b_compressed, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n, k) + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + + out = ops.cutlass_scaled_sparse_mm( + a, b_compressed, e, scale_a, scale_b, torch.bfloat16 + ) + out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16) + + if not torch.allclose(out, out_ref): + print("Incorrect results") + print(out) + print(out_ref) + else: + print("Correct results") + + timers = [] + + # pytorch impl w. bf16 + timers.append( + bench_fn( + label, + sub_label, + "pytorch_bf16_bf16_bf16_matmul-no-scales", + torch.mm, + a.to(dtype=torch.bfloat16, device="cuda"), + b.to(dtype=torch.bfloat16, device="cuda"), + ) + ) + + # pytorch impl: bf16 output, without fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_bf16_scaled_mm", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.bfloat16, + ) + ) + + # pytorch impl: bf16 output, with fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.bfloat16, + use_fast_accum=True, + ) + ) + + # pytorch impl: fp16 output, without fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_fp16_scaled_mm", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.float16, + ) + ) + + # pytorch impl: fp16 output, with fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.float16, + use_fast_accum=True, + ) + ) + + # cutlass impl: bf16 output + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_bf16_scaled_mm", + ops.cutlass_scaled_mm, + a, + b, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass impl: bf16 output + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_bf16_scaled_sparse_mm", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass impl: fp16 output + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_fp16_scaled_sparse_mm", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.float16, + ) + ) + + # cutlass impl: bf16 output, with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_bf16_scaled_sparse_mm_bias", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + bias, + ) + ) + + # cutlass impl: fp16 output, with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_fp16_scaled_sparse_mm_bias", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.float16, + bias.to(dtype=torch.float16), + ) + ) + + return timers + + +def bench( + dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str +) -> Iterable[TMeasurement]: + if dtype == torch.int8: + return bench_int8(dtype, m, k, n, label, sub_label) + if dtype == torch.float8_e4m3fn: + return bench_fp8(dtype, m, k, n, label, sub_label) + raise ValueError("unsupported type") + + +# runner +def print_timers(timers: Iterable[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def run( + dtype: torch.dtype, MKNs: Iterable[tuple[int, int, int]] +) -> Iterable[TMeasurement]: + results = [] + for m, k, n in MKNs: + timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm", f"MKN=({m}x{k}x{n})") + print_timers(timers) + results.extend(timers) + + return results + + +# output makers +def make_output( + data: Iterable[TMeasurement], + MKNs: Iterable[tuple[int, int, int]], + base_description: str, + timestamp=None, +): + print(f"== All Results {base_description} ====") + print_timers(data) + + # pickle all the results + timestamp = int(time.time()) if timestamp is None else timestamp + with open(f"{base_description}-{timestamp}.pkl", "wb") as f: + pkl.dump(data, f) + + +# argparse runners + + +def run_square_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment)) + MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes)) + data = run(args.dtype, MKNs) + + make_output(data, MKNs, f"square_bench-{args.dtype}") + + +def run_range_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment)) + n = len(dim_sizes) + Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes + Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes + Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes + MKNs = list(zip(Ms, Ks, Ns)) + data = run(args.dtype, MKNs) + + make_output(data, MKNs, f"range_bench-{args.dtype}") + + +def run_model_bench(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]: + KNs = [] + for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]): + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + KNs.append(KN) + return KNs + + model_bench_data = [] + models_tps = list(itertools.product(args.models, args.tp_sizes)) + for model, tp_size in models_tps: + Ms = args.batch_sizes + KNs = model_shapes(model, tp_size) + MKNs = [] + for m in Ms: + for k, n in KNs: + MKNs.append((m, k, n)) + + data = run(args.dtype, MKNs) + model_bench_data.append(data) + + # Print all results + for data, model_tp in zip(model_bench_data, models_tps): + model, tp_size = model_tp + print(f"== Results {args.dtype} {model}-TP{tp_size} ====") + print_timers(data) + + timestamp = int(time.time()) + + all_data = [] + for d in model_bench_data: + all_data.extend(d) + # pickle all data + with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f: + pkl.dump(all_data, f) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "int8": + return torch.int8 + if dt == "fp8": + return torch.float8_e4m3fn + raise ValueError("unsupported dtype") + + parser = FlexibleArgumentParser( + description=""" +Benchmark Cutlass GEMM. + + To run square GEMMs: + python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64 + + To run constant N and K and sweep M: + python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384 + + To run dimensions from a model: + python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1 + + Output: + - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs. + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + + parser.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['int8', 'fp8']", + ) + subparsers = parser.add_subparsers(dest="cmd") + + square_parser = subparsers.add_parser("square_bench") + square_parser.add_argument("--dim-start", type=int, required=True) + square_parser.add_argument("--dim-end", type=int, required=True) + square_parser.add_argument("--dim-increment", type=int, required=True) + square_parser.set_defaults(func=run_square_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--dim-start", type=int, required=True) + range_parser.add_argument("--dim-end", type=int, required=True) + range_parser.add_argument("--dim-increment", type=int, required=True) + range_parser.add_argument("--m-constant", type=int, default=None) + range_parser.add_argument("--n-constant", type=int, default=None) + range_parser.add_argument("--k-constant", type=int, default=None) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/offline_benchmark_test/benchmarks/cutlass_benchmarks/utils.py b/offline_benchmark_test/benchmarks/cutlass_benchmarks/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b4f3c6bf94eda0e1bf1d253def3b17d273415dcc --- /dev/null +++ b/offline_benchmark_test/benchmarks/cutlass_benchmarks/utils.py @@ -0,0 +1,100 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +# Cutlass bench utils +from collections.abc import Iterable + +import torch + +import vllm._custom_ops as ops + + +def to_fp8(tensor: torch.Tensor) -> torch.Tensor: + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to( + dtype=torch.float8_e4m3fn + ) + + +def to_int8(tensor: torch.Tensor) -> torch.Tensor: + return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8) + + +def to_bf16(tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(dtype=torch.bfloat16) + + +def to_fp16(tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(dtype=torch.float16) + + +def make_rand_tensors( + dtype: torch.dtype, m: int, n: int, k: int +) -> tuple[torch.Tensor, torch.Tensor]: + a = torch.randn((m, k), device="cuda") * 5 + b = torch.randn((n, k), device="cuda").t() * 5 + + if dtype == torch.int8: + return to_int8(a), to_int8(b) + if dtype == torch.float8_e4m3fn: + return to_fp8(a), to_fp8(b) + + raise ValueError("unsupported dtype") + + +def prune_to_2_4(tensor): + # Reshape tensor to [N, 4] where N is number of groups of 4 + original_shape = tensor.shape + reshaped = tensor.reshape(-1, 4) + + # Get indices of top 2 absolute values in each group of 4 + _, indices = torch.topk(torch.abs(reshaped), k=2, dim=1) + + # Create binary mask + mask = torch.zeros_like(reshaped) + mask.scatter_(dim=1, index=indices, src=torch.ones_like(indices, dtype=mask.dtype)) + + # Apply mask and reshape back + pruned = reshaped * mask + + # Turn all -0.0 to 0.0 + pruned[pruned == -0.0] = 0.0 + + return pruned.reshape(original_shape) + + +def make_rand_sparse_tensors( + dtype: torch.dtype, m: int, n: int, k: int +) -> tuple[torch.Tensor, torch.Tensor]: + a = torch.randn((m, k), device="cuda") * 5 + b = torch.randn((n, k), device="cuda").t() * 5 + + b = prune_to_2_4(b.t()).t() + + if dtype == torch.int8: + a, b = to_int8(a), to_int8(b) + elif dtype == torch.float8_e4m3fn: + a, b = to_fp8(a), to_fp8(b) + elif dtype == torch.float16: + a, b = to_fp16(a), to_fp16(b) + elif dtype == torch.bfloat16: + a, b = to_bf16(a), to_bf16(b) + else: + raise ValueError("unsupported dtype") + + b_compressed, e = ops.cutlass_sparse_compress(b.t()) + + # Compressed B, Metadata, Original A, B + return b_compressed, e, a, b + + +def make_n_rand_sparse_tensors( + num_tensors: int, dtype: torch.dtype, m: int, n: int, k: int +) -> tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]: + ABs = [] + for _ in range(num_tensors): + b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k) + if b_comp is not None: + ABs.append(make_rand_sparse_tensors(dtype, m, n, k)) + BComps, Es, As, Bs = zip(*ABs) + return list(BComps), list(Es), list(As), list(Bs) diff --git a/offline_benchmark_test/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py b/offline_benchmark_test/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py new file mode 100644 index 0000000000000000000000000000000000000000..a5a5b52f603977cc0c967e455724bc7c6a2bf585 --- /dev/null +++ b/offline_benchmark_test/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py @@ -0,0 +1,372 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import itertools +import pickle as pkl +import time +from collections.abc import Iterable +from typing import Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import make_rand_tensors +from weight_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + w8a8_block_fp8_matmul, +) +from vllm.utils import FlexibleArgumentParser, cdiv + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512] +DEFAULT_TP_SIZES = [1] + + +# bench +def bench_fn( + label: str, sub_label: str, description: str, fn: Callable, *args, **kwargs +) -> TMeasurement: + min_run_time = 1 + + globals = { + "args": args, + "kwargs": kwargs, + "fn": fn, + } + return TBenchmark.Timer( + stmt="fn(*args, **kwargs)", + globals=globals, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + +def bench_int8( + dtype: torch.dtype, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + """Benchmark INT8-based kernels.""" + assert dtype == torch.int8 + a, b = make_rand_tensors(torch.int8, m, n, k) + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + azp = torch.zeros((m,), device="cuda", dtype=torch.int32) + azp_adj = torch.zeros((n,), device="cuda", dtype=torch.int32) + + bench_fns = { + "pytorch_bf16_bf16_bf16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16) + ), + "pytorch_fp16_fp16_fp16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.float16), b.to(dtype=torch.float16) + ), + "cutlass_i8_i8_bf16_scaled_mm": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16 + ), + "cutlass_i8_i8_bf16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16, bias + ), + "cutlass_i8_i8_bf16_scaled_mm_azp": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj + ), + "cutlass_i8_i8_bf16_scaled_mm_azp_bias": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj, None, bias + ), + "cutlass_i8_i8_bf16_scaled_mm_azp_pt": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj, azp + ), + "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj, azp, bias + ), + } + + timers = [] + for name, fn in bench_fns.items(): + # If bench_kernels is None, run all. Otherwise, run only exact matches. + if bench_kernels is None or name in bench_kernels: + print(f"Running {name}") + timers.append(bench_fn(label, sub_label, name, fn)) + + return timers + + +def bench_fp8( + dtype: torch.dtype, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + """Benchmark FP8-based kernels.""" + assert dtype == torch.float8_e4m3fn + a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k) + a_cont = a.contiguous() + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + + block_scale_a = torch.rand((m, cdiv(k, 128)), device="cuda", dtype=torch.float32) + block_scale_b = torch.rand( + cdiv(k, 128), cdiv(n, 128), device="cuda", dtype=torch.float32 + ) + block_scale_a_M_major = block_scale_a.t().contiguous().t() + block_scale_b_K_major = block_scale_b.t().contiguous().t() + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + + print(m, k, n) + + bench_fns = { + "pytorch_bf16_bf16_bf16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16) + ), + "pytorch_fp16_fp16_fp16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.float16), b.to(dtype=torch.float16) + ), + "pytorch_fp8_fp8_fp16_scaled_mm": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.float16 + ), + "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.float16, use_fast_accum=True + ), + "pytorch_fp8_fp8_bf16_scaled_mm": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.bfloat16 + ), + "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.bfloat16, use_fast_accum=True + ), + "cutlass_fp8_fp8_bf16_scaled_mm": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16 + ), + "cutlass_fp8_fp8_fp16_scaled_mm": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.float16 + ), + "cutlass_fp8_fp8_bf16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16, bias + ), + "cutlass_fp8_fp8_fp16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.float16, bias.to(dtype=torch.float16) + ), + "triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_block_fp8_matmul( + a_cont, b.t(), block_scale_a, block_scale_b.t(), (128, 128) + ), + "cutlass_fp8_fp8_fp16_scaled_mm_blockwise": lambda: ops.cutlass_scaled_mm( + a, b, block_scale_a_M_major, block_scale_b_K_major, torch.float16 + ), + } + + timers = [] + for name, fn in bench_fns.items(): + # If bench_kernels is None, run all. Otherwise, run only exact matches. + if bench_kernels is None or name in bench_kernels: + print(f"Running {name}") + timers.append(bench_fn(label, sub_label, name, fn)) + + return timers + + +def bench( + dtype: torch.dtype, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + if dtype == torch.int8: + return bench_int8(dtype, m, k, n, label, sub_label, bench_kernels) + if dtype == torch.float8_e4m3fn: + return bench_fp8(dtype, m, k, n, label, sub_label, bench_kernels) + raise ValueError("unsupported type") + + +# runner +def print_timers(timers: Iterable[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def run( + dtype: torch.dtype, + MKNs: Iterable[tuple[int, int, int]], + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + results = [] + for m, k, n in MKNs: + timers = bench( + dtype, + m, + k, + n, + f"scaled-{dtype}-gemm", + f"MKN=({m}x{k}x{n})", + bench_kernels=bench_kernels, + ) + print_timers(timers) + results.extend(timers) + return results + + +def make_output( + data: Iterable[TMeasurement], + MKNs: Iterable[tuple[int, int, int]], + base_description: str, + timestamp=None, +): + print(f"== All Results {base_description} ====") + print_timers(data) + + # pickle all the results + timestamp = int(time.time()) if timestamp is None else timestamp + with open(f"{base_description}-{timestamp}.pkl", "wb") as f: + pkl.dump(data, f) + + +def run_square_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment)) + MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes)) + data = run(args.dtype, MKNs, bench_kernels=args.kernels) + make_output(data, MKNs, f"square_bench-{args.dtype}") + + +def run_range_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment)) + n = len(dim_sizes) + Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes + Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes + Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes + MKNs = list(zip(Ms, Ks, Ns)) + data = run(args.dtype, MKNs, bench_kernels=args.kernels) + make_output(data, MKNs, f"range_bench-{args.dtype}") + + +def run_model_bench(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]: + KNs = [] + for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]): + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + KNs.append(KN) + return KNs + + model_bench_data = [] + models_tps = list(itertools.product(args.models, args.tp_sizes)) + for model, tp_size in models_tps: + Ms = args.batch_sizes + KNs = model_shapes(model, tp_size) + MKNs = [] + for m in Ms: + for k, n in KNs: + MKNs.append((m, k, n)) + + data = run(args.dtype, MKNs, bench_kernels=args.kernels) + model_bench_data.append(data) + + # Print all results + for data, model_tp in zip(model_bench_data, models_tps): + model, tp_size = model_tp + print(f"== Results {args.dtype} {model}-TP{tp_size} ====") + print_timers(data) + + timestamp = int(time.time()) + + all_data = [] + for d in model_bench_data: + all_data.extend(d) + # pickle all data + with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f: + pkl.dump(all_data, f) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "int8": + return torch.int8 + if dt == "fp8": + return torch.float8_e4m3fn + raise ValueError("unsupported dtype") + + parser = FlexibleArgumentParser( + description=""" +Benchmark Cutlass GEMM. + + To run square GEMMs: + python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64 + + To run constant N and K and sweep M: + python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384 + + To run dimensions from a model: + python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1 + + Output: + - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs. + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + + parser.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['int8', 'fp8']", + ) + parser.add_argument( + "--kernels", + nargs="+", + type=str, + default=None, + help="Exact names of the kernels to benchmark. If not set, runs all kernels.", + ) + + subparsers = parser.add_subparsers(dest="cmd") + + square_parser = subparsers.add_parser("square_bench") + square_parser.add_argument("--dim-start", type=int, required=True) + square_parser.add_argument("--dim-end", type=int, required=True) + square_parser.add_argument("--dim-increment", type=int, required=True) + square_parser.set_defaults(func=run_square_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--dim-start", type=int, required=True) + range_parser.add_argument("--dim-end", type=int, required=True) + range_parser.add_argument("--dim-increment", type=int, required=True) + range_parser.add_argument("--m-constant", type=int, default=None) + range_parser.add_argument("--n-constant", type=int, default=None) + range_parser.add_argument("--k-constant", type=int, default=None) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/offline_benchmark_test/benchmarks/cutlass_benchmarks/weight_shapes.py b/offline_benchmark_test/benchmarks/cutlass_benchmarks/weight_shapes.py new file mode 100644 index 0000000000000000000000000000000000000000..25b96ef56620ea7dbd97846cdd57ab0e97a6dfd1 --- /dev/null +++ b/offline_benchmark_test/benchmarks/cutlass_benchmarks/weight_shapes.py @@ -0,0 +1,46 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +# Weight Shapes are in the format +# ([K, N], TP_SPLIT_DIM) +# Example: +# A shape of ([14336, 4096], 0) indicates the following GEMM shape, +# - TP1 : K = 14336, N = 4096 +# - TP2 : K = 7168, N = 4096 +# A shape of ([4096, 6144], 1) indicates the following GEMM shape, +# - TP1 : K = 4096, N = 6144 +# - TP4 : K = 4096, N = 1536 + +# TP1 shapes +WEIGHT_SHAPES = { + "mistralai/Mistral-7B-v0.1": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-7b-hf": [ + ([4096, 12288], 1), + ([4096, 4096], 0), + ([4096, 22016], 1), + ([11008, 4096], 0), + ], + "meta-llama/Llama-3-8b": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-13b-hf": [ + ([5120, 15360], 1), + ([5120, 5120], 0), + ([5120, 27648], 1), + ([13824, 5120], 0), + ], + "meta-llama/Llama-2-70b-hf": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 57344], 1), + ([28672, 8192], 0), + ], +} diff --git a/offline_benchmark_test/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh b/offline_benchmark_test/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..94999630bae122fa86d68c77cf9fc3941fbe4b14 --- /dev/null +++ b/offline_benchmark_test/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh @@ -0,0 +1,145 @@ +#!/bin/bash + +# benchmark the overhead of disaggregated prefill. +# methodology: +# - send all request to prefill vLLM instance. It will buffer KV cache. +# - then send all request to decode instance. +# - The TTFT of decode instance is the overhead. + +set -ex + +kill_gpu_processes() { + # kill all processes on GPU. + pgrep pt_main_thread | xargs -r kill -9 + pgrep python3 | xargs -r kill -9 + sleep 10 + + # remove vllm config file + rm -rf ~/.config/vllm + + # Print the GPU memory usage + # so that we know if all GPU processes are killed. + gpu_memory_usage=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0) + # The memory usage should be 0 MB. + echo "GPU 0 Memory Usage: $gpu_memory_usage MB" +} + +wait_for_server() { + # wait for vllm server to start + # return 1 if vllm server crashes + local port=$1 + timeout 1200 bash -c " + until curl -s localhost:${port}/v1/completions > /dev/null; do + sleep 1 + done" && return 0 || return 1 +} + + +benchmark() { + + export VLLM_LOGGING_LEVEL=DEBUG + export VLLM_HOST_IP=$(hostname -I | awk '{print $1}') + + # compare chunked prefill with disaggregated prefill + + results_folder="./results" + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + dataset_name="sonnet" + dataset_path="../sonnet_4x.txt" + num_prompts=10 + qps=$1 + prefix_len=50 + input_len=2048 + output_len=$2 + + + CUDA_VISIBLE_DEVICES=0 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8100 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + + CUDA_VISIBLE_DEVICES=1 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8200 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + wait_for_server 8100 + wait_for_server 8200 + + # let the prefill instance finish prefill + python3 ../benchmark_serving.py \ + --backend vllm \ + --model $model \ + --dataset-name $dataset_name \ + --dataset-path $dataset_path \ + --sonnet-input-len $input_len \ + --sonnet-output-len "$output_len" \ + --sonnet-prefix-len $prefix_len \ + --num-prompts $num_prompts \ + --port 8100 \ + --save-result \ + --result-dir $results_folder \ + --result-filename disagg_prefill_tp1.json \ + --request-rate "inf" + + + # send the request to decode. + # The TTFT of this command will be the overhead of disagg prefill impl. + python3 ../benchmark_serving.py \ + --backend vllm \ + --model $model \ + --dataset-name $dataset_name \ + --dataset-path $dataset_path \ + --sonnet-input-len $input_len \ + --sonnet-output-len "$output_len" \ + --sonnet-prefix-len $prefix_len \ + --num-prompts $num_prompts \ + --port 8200 \ + --save-result \ + --result-dir $results_folder \ + --result-filename disagg_prefill_tp1_overhead.json \ + --request-rate "$qps" + kill_gpu_processes + +} + + +main() { + + (which wget && which curl) || (apt-get update && apt-get install -y wget curl) + (which jq) || (apt-get -y install jq) + (which socat) || (apt-get -y install socat) + + pip install quart httpx datasets + + cd "$(dirname "$0")" + + cd .. + # create sonnet-4x.txt + echo "" > sonnet_4x.txt + for _ in {1..4} + do + cat sonnet.txt >> sonnet_4x.txt + done + cd disagg_benchmarks + + rm -rf results + mkdir results + + default_qps=1 + default_output_len=1 + benchmark $default_qps $default_output_len + +} + + +main "$@" diff --git a/offline_benchmark_test/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh b/offline_benchmark_test/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..eb5d891d0d4a58b75f975ae43655ad539bc6a71e --- /dev/null +++ b/offline_benchmark_test/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh @@ -0,0 +1,163 @@ +#!/bin/bash + +# Requirement: 2x GPUs. + + +# Model: meta-llama/Meta-Llama-3.1-8B-Instruct +# Query: 1024 input tokens, 6 output tokens, QPS 2/4/6/8, 100 requests +# Resource: 2x GPU +# Approaches: +# 2. Chunked prefill: 2 vllm instance with tp=4, equivalent to 1 tp=4 instance with QPS 4 +# 3. Disaggregated prefill: 1 prefilling instance and 1 decoding instance +# Prefilling instance: max_output_token=1 +# Decoding instance: force the input tokens be the same across requests to bypass prefilling + +set -ex + +kill_gpu_processes() { + # kill all processes on GPU. + pgrep pt_main_thread | xargs -r kill -9 + pgrep python3 | xargs -r kill -9 + for port in 8000 8100 8200; do lsof -t -i:$port | xargs -r kill -9; done + sleep 1 +} + +wait_for_server() { + # wait for vllm server to start + # return 1 if vllm server crashes + local port=$1 + timeout 1200 bash -c " + until curl -s localhost:${port}/v1/completions > /dev/null; do + sleep 1 + done" && return 0 || return 1 +} + + +launch_chunked_prefill() { + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + # disagg prefill + CUDA_VISIBLE_DEVICES=0 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8100 \ + --max-model-len 10000 \ + --enable-chunked-prefill \ + --gpu-memory-utilization 0.6 & + CUDA_VISIBLE_DEVICES=1 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8200 \ + --max-model-len 10000 \ + --enable-chunked-prefill \ + --gpu-memory-utilization 0.6 & + wait_for_server 8100 + wait_for_server 8200 + python3 round_robin_proxy.py & + sleep 1 +} + + +launch_disagg_prefill() { + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + # disagg prefill + CUDA_VISIBLE_DEVICES=0 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8100 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + CUDA_VISIBLE_DEVICES=1 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8200 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + wait_for_server 8100 + wait_for_server 8200 + python3 disagg_prefill_proxy_server.py & + sleep 1 +} + + +benchmark() { + results_folder="./results" + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + dataset_name="sonnet" + dataset_path="../sonnet_4x.txt" + num_prompts=100 + qps=$1 + prefix_len=50 + input_len=1024 + output_len=$2 + tag=$3 + + python3 ../benchmark_serving.py \ + --backend vllm \ + --model $model \ + --dataset-name $dataset_name \ + --dataset-path $dataset_path \ + --sonnet-input-len $input_len \ + --sonnet-output-len "$output_len" \ + --sonnet-prefix-len $prefix_len \ + --num-prompts $num_prompts \ + --port 8000 \ + --save-result \ + --result-dir $results_folder \ + --result-filename "$tag"-qps-"$qps".json \ + --request-rate "$qps" + + sleep 2 +} + + +main() { + + (which wget && which curl) || (apt-get update && apt-get install -y wget curl) + (which jq) || (apt-get -y install jq) + (which socat) || (apt-get -y install socat) + (which lsof) || (apt-get -y install lsof) + + pip install quart httpx matplotlib aiohttp datasets + + cd "$(dirname "$0")" + + cd .. + # create sonnet-4x.txt so that we can sample 2048 tokens for input + echo "" > sonnet_4x.txt + for _ in {1..4} + do + cat sonnet.txt >> sonnet_4x.txt + done + cd disagg_benchmarks + + rm -rf results + mkdir results + + default_output_len=6 + + export VLLM_HOST_IP=$(hostname -I | awk '{print $1}') + + launch_chunked_prefill + for qps in 2 4 6 8; do + benchmark $qps $default_output_len chunked_prefill + done + kill_gpu_processes + + launch_disagg_prefill + for qps in 2 4 6 8; do + benchmark $qps $default_output_len disagg_prefill + done + kill_gpu_processes + + python3 visualize_benchmark_results.py + +} + + +main "$@" diff --git a/offline_benchmark_test/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py b/offline_benchmark_test/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py new file mode 100644 index 0000000000000000000000000000000000000000..f62d8102e2d9f29c62f177d62453be5d8496cc1e --- /dev/null +++ b/offline_benchmark_test/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py @@ -0,0 +1,63 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import os + +import aiohttp +from quart import Quart, make_response, request + +AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60) + +app = Quart(__name__) + + +async def forward_request(url, data): + async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session: + headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"} + async with session.post(url=url, json=data, headers=headers) as response: + if response.status == 200: + # if response.headers.get('Transfer-Encoding') == 'chunked': + if True: + async for chunk_bytes in response.content.iter_chunked(1024): + yield chunk_bytes + else: + content = await response.read() + yield content + + +@app.route("/v1/completions", methods=["POST"]) +async def handle_request(): + try: + original_request_data = await request.get_json() + + prefill_request = original_request_data.copy() + # change max_tokens = 1 to let it only do prefill + prefill_request["max_tokens"] = 1 + + # finish prefill + async for _ in forward_request( + "http://localhost:8100/v1/completions", prefill_request + ): + continue + + # return decode + generator = forward_request( + "http://localhost:8200/v1/completions", original_request_data + ) + response = await make_response(generator) + response.timeout = None + + return response + + except Exception as e: + import sys + import traceback + + exc_info = sys.exc_info() + print("Error occurred in disagg prefill proxy server") + print(e) + print("".join(traceback.format_exception(*exc_info))) + + +if __name__ == "__main__": + app.run(port=8000) diff --git a/offline_benchmark_test/benchmarks/disagg_benchmarks/round_robin_proxy.py b/offline_benchmark_test/benchmarks/disagg_benchmarks/round_robin_proxy.py new file mode 100644 index 0000000000000000000000000000000000000000..b1df2f255822dad046f5dfcdc1d6538006463510 --- /dev/null +++ b/offline_benchmark_test/benchmarks/disagg_benchmarks/round_robin_proxy.py @@ -0,0 +1,63 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import asyncio +import itertools + +import aiohttp +from aiohttp import web + + +class RoundRobinProxy: + def __init__(self, target_ports): + self.target_ports = target_ports + self.port_cycle = itertools.cycle(self.target_ports) + + async def handle_request(self, request): + target_port = next(self.port_cycle) + target_url = f"http://localhost:{target_port}{request.path_qs}" + + async with aiohttp.ClientSession() as session: + try: + # Forward the request + async with session.request( + method=request.method, + url=target_url, + headers=request.headers, + data=request.content, + ) as response: + # Start sending the response + resp = web.StreamResponse( + status=response.status, headers=response.headers + ) + await resp.prepare(request) + + # Stream the response content + async for chunk in response.content.iter_any(): + await resp.write(chunk) + + await resp.write_eof() + return resp + + except Exception as e: + return web.Response(text=f"Error: {str(e)}", status=500) + + +async def main(): + proxy = RoundRobinProxy([8100, 8200]) + app = web.Application() + app.router.add_route("*", "/{path:.*}", proxy.handle_request) + + runner = web.AppRunner(app) + await runner.setup() + site = web.TCPSite(runner, "localhost", 8000) + await site.start() + + print("Proxy server started on http://localhost:8000") + + # Keep the server running + await asyncio.Event().wait() + + +if __name__ == "__main__": + asyncio.run(main()) diff --git a/offline_benchmark_test/benchmarks/disagg_benchmarks/visualize_benchmark_results.py b/offline_benchmark_test/benchmarks/disagg_benchmarks/visualize_benchmark_results.py new file mode 100644 index 0000000000000000000000000000000000000000..74fa56d076cf14bc066468be24f6053b45166001 --- /dev/null +++ b/offline_benchmark_test/benchmarks/disagg_benchmarks/visualize_benchmark_results.py @@ -0,0 +1,47 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import json + +import matplotlib.pyplot as plt +import pandas as pd + +if __name__ == "__main__": + data = [] + for name in ["disagg_prefill", "chunked_prefill"]: + for qps in [2, 4, 6, 8]: + with open(f"results/{name}-qps-{qps}.json") as f: + x = json.load(f) + x["name"] = name + x["qps"] = qps + data.append(x) + + df = pd.DataFrame.from_dict(data) + dis_df = df[df["name"] == "disagg_prefill"] + chu_df = df[df["name"] == "chunked_prefill"] + + plt.style.use("bmh") + plt.rcParams["font.size"] = 20 + + for key in [ + "mean_ttft_ms", + "median_ttft_ms", + "p99_ttft_ms", + "mean_itl_ms", + "median_itl_ms", + "p99_itl_ms", + ]: + fig, ax = plt.subplots(figsize=(11, 7)) + plt.plot( + dis_df["qps"], dis_df[key], label="disagg_prefill", marker="o", linewidth=4 + ) + plt.plot( + chu_df["qps"], chu_df[key], label="chunked_prefill", marker="o", linewidth=4 + ) + ax.legend() + + ax.set_xlabel("QPS") + ax.set_ylabel(key) + ax.set_ylim(bottom=0) + fig.savefig(f"results/{key}.png") + plt.close(fig) diff --git a/offline_benchmark_test/benchmarks/fused_kernels/layernorm_rms_benchmarks.py b/offline_benchmark_test/benchmarks/fused_kernels/layernorm_rms_benchmarks.py new file mode 100644 index 0000000000000000000000000000000000000000..901524214469e8677cbb58c3078e8b1d724585bb --- /dev/null +++ b/offline_benchmark_test/benchmarks/fused_kernels/layernorm_rms_benchmarks.py @@ -0,0 +1,228 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import pickle as pkl +import time +from collections.abc import Iterable +from dataclasses import dataclass +from itertools import product +from typing import Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from tqdm import tqdm + +import vllm._custom_ops as ops +from vllm.model_executor.layers.layernorm import RMSNorm + + +@dataclass +class bench_params_t: + num_tokens: int + hidden_size: int + add_residual: bool + dtype: torch.dtype + + def description(self): + return ( + f"N {self.num_tokens} " + f"x D {self.hidden_size} " + f"x R {self.add_residual} " + f"x DT {self.dtype}" + ) + + +def get_bench_params() -> list[bench_params_t]: + ## Test Fixtures + NUM_TOKENS = [2**x for x in range(11)] + HIDDEN_SIZES = list(range(1024, 8129, 1024)) + ADD_RESIDUAL = [True, False] + DTYPES = [torch.bfloat16, torch.float] + + combinations = product(NUM_TOKENS, HIDDEN_SIZES, ADD_RESIDUAL, DTYPES) + bench_params = list( + map(lambda x: bench_params_t(x[0], x[1], x[2], x[3]), combinations) + ) + return bench_params + + +# Reference impls +def unfused_int8_impl( + rms_norm_layer: RMSNorm, + x: torch.Tensor, + residual: Optional[torch.Tensor], + quant_dtype: torch.dtype, +): + # Norm + torch_out = None + if residual is None: + torch_out = rms_norm_layer.forward_cuda(x, residual) + else: + torch_out, _ = rms_norm_layer.forward_cuda(x, residual) + + # Quant + torch_out, _, _ = ops.scaled_int8_quant(torch_out) + + +def unfused_fp8_impl( + rms_norm_layer: RMSNorm, + x: torch.Tensor, + residual: Optional[torch.Tensor], + quant_dtype: torch.dtype, +): + # Norm + torch_out = None + if residual is None: + torch_out = rms_norm_layer.forward_cuda(x, residual) + else: + torch_out, _ = rms_norm_layer.forward_cuda(x, residual) + + # Quant + torch_out, _ = ops.scaled_fp8_quant(torch_out) + + +def fused_impl( + rms_norm_layer: RMSNorm, # this stores the weights + x: torch.Tensor, + residual: Optional[torch.Tensor], + quant_dtype: torch.dtype, +): + out, _ = ops.rms_norm_dynamic_per_token_quant( + x, rms_norm_layer.weight, 1e-6, quant_dtype, residual=residual + ) + + +# Bench functions +def bench_fn( + rms_norm_layer: RMSNorm, + x: torch.Tensor, + residual: torch.Tensor, + quant_dtype: torch.dtype, + label: str, + sub_label: str, + fn: Callable, + description: str, +) -> TMeasurement: + min_run_time = 1 + + globals = { + "rms_norm_layer": rms_norm_layer, + "x": x, + "residual": residual, + "quant_dtype": quant_dtype, + "fn": fn, + } + return TBenchmark.Timer( + stmt="fn(rms_norm_layer, x, residual, quant_dtype)", + globals=globals, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + +def bench(params: bench_params_t, label: str, sub_label: str) -> Iterable[TMeasurement]: + # Make inputs + layer = RMSNorm(params.hidden_size, 1e-6).to(dtype=params.dtype) + # Make weights + layer.weight.data.normal_(mean=1.0, std=0.1) + # Make inputs + scale = 1 / params.hidden_size + x = ( + torch.randn( + params.num_tokens, params.hidden_size, dtype=params.dtype, device="cuda" + ) + * scale + ) + residual = ( + (torch.randn_like(x) * scale).to(device="cuda") if params.add_residual else None + ) + + timers = [] + + # unfused int8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.int8, + label, + sub_label, + unfused_int8_impl, + "unfused_int8_impl", + ) + ) + + # unfused fp8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.float8_e4m3fn, + label, + sub_label, + unfused_fp8_impl, + "unfused_fp8_impl", + ) + ) + + # fused int8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.int8, + label, + sub_label, + fused_impl, + "fused_int8_impl", + ) + ) + + # fused fp8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.float8_e4m3fn, + label, + sub_label, + fused_impl, + "fused_fp8_impl", + ) + ) + + print_timers(timers) + + return timers + + +# launch bench +# runner +def print_timers(timers: Iterable[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def main(): + torch.set_default_device("cuda") + bench_params = get_bench_params() + + timers = [] + for bp in tqdm(bench_params): + timers.extend(bench(bp, "rms-norm-dynamic-per-token-quant", bp.description())) + print_timers(timers) + + # pickle all the results + timestamp = int(time.time()) + with open(f"rms_norm_dpt_quant-{timestamp}.pkl", "wb") as f: + pkl.dump(timers, f) + + +if __name__ == "__main__": + main() diff --git a/offline_benchmark_test/benchmarks/kernels/bench_fp8_gemm.py b/offline_benchmark_test/benchmarks/kernels/bench_fp8_gemm.py new file mode 100644 index 0000000000000000000000000000000000000000..920961899038061c13125bdd37f9843b5e4d548a --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/bench_fp8_gemm.py @@ -0,0 +1,159 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import argparse +import copy +import itertools + +import torch +from weight_shapes import WEIGHT_SHAPES + +from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm +from vllm._custom_ops import scaled_fp8_quant as vllm_scaled_fp8_quant +from vllm.triton_utils import triton + +PROVIDER_CFGS = { + "torch-bf16": dict(enabled=True), + "fp8-tensor-w-token-a": dict( + w="tensor", a="token", no_a_quant=False, enabled=False + ), + "fp8-tensor-w-tensor-a": dict( + w="tensor", a="tensor", no_a_quant=False, enabled=True + ), + "fp8-channel-w-token-a": dict( + w="channel", a="token", no_a_quant=False, enabled=True + ), + "fp8-channel-w-tensor-a": dict( + w="channel", a="tensor", no_a_quant=False, enabled=False + ), + "fp8-tensor-w-token-a-noquant": dict( + w="tensor", a="token", no_a_quant=True, enabled=False + ), + "fp8-tensor-w-tensor-a-noquant": dict( + w="tensor", a="tensor", no_a_quant=True, enabled=True + ), + "fp8-channel-w-token-a-noquant": dict( + w="channel", a="token", no_a_quant=True, enabled=True + ), + "fp8-channel-w-tensor-a-noquant": dict( + w="channel", a="tensor", no_a_quant=True, enabled=False + ), +} + +_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]] + + +def _quant_weight_fp8(b: torch.Tensor, w_type: str, device: str): + if w_type == "tensor": + scale_b = torch.ones(1, device=device, dtype=torch.float32) + b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b) + else: + b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, use_per_token_if_dynamic=True) + return b_fp8.t(), scale_b_fp8 + + +def build_fp8_runner(cfg, a, b, dtype, device): + b_fp8, scale_b_fp8 = _quant_weight_fp8(b, cfg["w"], device) + + scale_a_const = ( + torch.ones(1, device=device, dtype=torch.float32) + if cfg["a"] == "tensor" + else None + ) + + if cfg["no_a_quant"]: + if cfg["a"] == "tensor": + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const) + else: + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True) + + def run(): + return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype) + + return run + + if cfg["a"] == "tensor": + + def run(): + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const) + return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype) + + else: + + def run(): + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True) + return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype) + + return run + + +@triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["batch_size"], + x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384], + x_log=False, + line_arg="provider", + line_vals=_enabled, + line_names=_enabled, + ylabel="TFLOP/s (larger is better)", + plot_name="BF16 vs FP8 GEMMs", + args={}, + ) +) +def benchmark(batch_size, provider, N, K): + M = batch_size + device = "cuda" + dtype = torch.bfloat16 + + a = torch.randn((M, K), device=device, dtype=dtype) + b = torch.randn((N, K), device=device, dtype=dtype) + + quantiles = [0.5, 0.2, 0.8] + + if provider == "torch-bf16": + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: torch.nn.functional.linear(a, b), quantiles=quantiles + ) + else: + cfg = PROVIDER_CFGS[provider] + run_quant = build_fp8_runner(cfg, a, b, dtype, device) + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: run_quant(), quantiles=quantiles + ) + + to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3) + return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms) + + +def prepare_shapes(args): + out = [] + for model, tp_size in itertools.product(args.models, args.tp_sizes): + for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]): + KN[tp_dim] //= tp_size + KN.append(model) + out.append(KN) + return out + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--models", + nargs="+", + type=str, + default=["meta-llama/Llama-3.1-8B-Instruct"], + choices=list(WEIGHT_SHAPES.keys()), + ) + parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1]) + args = parser.parse_args() + + for K, N, model in prepare_shapes(args): + print(f"{model}, N={N} K={K}, BF16 vs FP8 GEMMs TFLOP/s:") + benchmark.run( + print_data=True, + show_plots=True, + save_path=f"bench_fp8_res_n{N}_k{K}", + N=N, + K=K, + ) + + print("Benchmark finished!") diff --git a/offline_benchmark_test/benchmarks/kernels/bench_int8_gemm.py b/offline_benchmark_test/benchmarks/kernels/bench_int8_gemm.py new file mode 100644 index 0000000000000000000000000000000000000000..e9c6d64404d0dc6958aeb675bf6b893623649ffa --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/bench_int8_gemm.py @@ -0,0 +1,169 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import argparse +import copy +import itertools + +import torch +from weight_shapes import WEIGHT_SHAPES + +from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm +from vllm._custom_ops import scaled_int8_quant as vllm_scaled_int8_quant +from vllm.triton_utils import triton + +PROVIDER_CFGS = { + "torch-bf16": dict(enabled=True), + "int8-tensor-w-token-a": dict( + w="tensor", a="token", no_a_quant=False, enabled=False + ), + "int8-tensor-w-tensor-a": dict( + w="tensor", a="tensor", no_a_quant=False, enabled=True + ), + "int8-channel-w-token-a": dict( + w="channel", a="token", no_a_quant=False, enabled=True + ), + "int8-channel-w-tensor-a": dict( + w="channel", a="tensor", no_a_quant=False, enabled=False + ), + "int8-tensor-w-token-a-noquant": dict( + w="tensor", a="token", no_a_quant=True, enabled=False + ), + "int8-tensor-w-tensor-a-noquant": dict( + w="tensor", a="tensor", no_a_quant=True, enabled=True + ), + "int8-channel-w-token-a-noquant": dict( + w="channel", a="token", no_a_quant=True, enabled=True + ), + "int8-channel-w-tensor-a-noquant": dict( + w="channel", a="tensor", no_a_quant=True, enabled=False + ), +} + + +def _quant_weight(b, w_type, device): + if w_type == "tensor": + scale_b = torch.ones(1, device=device, dtype=torch.float32) + b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b, scale_b) + assert scale_b_int8.numel() == 1 + else: # channel + b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b) + assert scale_b_int8.numel() == b.shape[0] + return b_int8.t(), scale_b_int8 + + +def build_int8_runner(cfg, a, b, dtype, device): + # quant before running the kernel + b_int8, scale_b_int8 = _quant_weight(b, cfg["w"], device) + + scale_a_const = None + if cfg["a"] == "tensor": + scale_a_const = torch.ones(1, device=device, dtype=torch.float32) + + # no quant, create activation ahead + if cfg["no_a_quant"]: + if cfg["a"] == "tensor": + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const) + else: # token + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a) + + def run_quant(): + return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype) + + return run_quant + + # dynamic quant, create activation inside + if cfg["a"] == "tensor": + + def run_quant(): + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const) + return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype) + + else: # token + + def run_quant(): + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a) + return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype) + + return run_quant + + +_enabled = [k for k, v in PROVIDER_CFGS.items() if v.get("enabled")] + + +@triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["batch_size"], + x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384], + x_log=False, + line_arg="provider", + line_vals=_enabled, + line_names=[k for k in _enabled], + ylabel="TFLOP/s (larger is better)", + plot_name="BF16 vs INT8 GEMMs", + args={}, + ) +) +def benchmark(batch_size, provider, N, K): + M = batch_size + device = "cuda" + dtype = torch.bfloat16 + a = torch.randn((M, K), device=device, dtype=dtype) + b = torch.randn((N, K), device=device, dtype=dtype) + + quantiles = [0.5, 0.2, 0.8] + + if provider == "torch-bf16": + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: torch.nn.functional.linear(a, b), quantiles=quantiles + ) + else: + cfg = PROVIDER_CFGS[provider] + run_quant = build_int8_runner(cfg, a, b, dtype, device) + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: run_quant(), quantiles=quantiles + ) + + to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3) + return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms) + + +def prepare_shapes(args): + KN_model_names = [] + for model, tp_size in itertools.product(args.models, args.tp_sizes): + for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]): + KN[tp_dim] //= tp_size + KN.append(model) + KN_model_names.append(KN) + return KN_model_names + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--models", + nargs="+", + type=str, + default=["meta-llama/Llama-3.1-8B-Instruct"], + choices=list(WEIGHT_SHAPES.keys()), + help="List of models to benchmark", + ) + parser.add_argument( + "--tp-sizes", + nargs="+", + type=int, + default=[1], + help="List of tensor parallel sizes", + ) + args = parser.parse_args() + + for K, N, model in prepare_shapes(args): + print(f"{model}, N={N} K={K}, BF16 vs INT8 GEMMs TFLOP/s:") + benchmark.run( + print_data=True, + show_plots=True, + save_path=f"bench_int8_res_n{N}_k{K}", + N=N, + K=K, + ) + + print("Benchmark finished!") diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_aqlm.py b/offline_benchmark_test/benchmarks/kernels/benchmark_aqlm.py new file mode 100644 index 0000000000000000000000000000000000000000..42de062b08e424619b010f9170725b6785a35f58 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_aqlm.py @@ -0,0 +1,345 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import os +import sys +from typing import Optional + +import torch +import torch.nn.functional as F + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.aqlm import ( + dequantize_weight, + generic_dequantize_gemm, + get_int_dtype, + optimized_dequantize_gemm, +) +from vllm.utils import FlexibleArgumentParser + +os.environ["CUDA_VISIBLE_DEVICES"] = "0" + + +def torch_mult( + # [..., in_features] + input: torch.Tensor, + weights: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, +) -> torch.Tensor: + output = F.linear(input, weights) + return output + + +def dequant_out_scale( + # [..., in_features] + input: torch.Tensor, + # [num_out_groups, num_in_groups, num_codebooks] + codes: torch.IntTensor, + # [num_codebooks, codebook_size, out_group_size, in_group_size] + codebooks: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, + output_partition_sizes: torch.IntTensor, + bias: Optional[torch.Tensor], +) -> torch.Tensor: + weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes) + + if bias is None: + output = F.linear(input, weights, bias) + orig_shape = output.shape + flattened_output = output.view(-1, output.size(-1)) + f_scales = scales.view(-1, scales.shape[0]) + b_scales = f_scales.expand(flattened_output.shape[0], -1) + flattened_output *= b_scales + return flattened_output.view(orig_shape) + else: + b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1]) + weights *= b_scales + return F.linear(input, weights, bias) + + +def dequant_weight_scale( + # [..., in_features] + input: torch.Tensor, + # [num_out_groups, num_in_groups, num_codebooks] + codes: torch.IntTensor, + # [num_codebooks, codebook_size, out_group_size, in_group_size] + codebooks: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, + output_partition_sizes: torch.IntTensor, + bias: Optional[torch.Tensor], +) -> torch.Tensor: + weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes) + + b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1]) + weights *= b_scales + return F.linear(input, weights, bias) + + +def dequant_no_scale( + # [..., in_features] + input: torch.Tensor, + # [num_out_groups, num_in_groups, num_codebooks] + codes: torch.IntTensor, + # [num_codebooks, codebook_size, out_group_size, in_group_size] + codebooks: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, + output_partition_sizes: torch.IntTensor, + bias: Optional[torch.Tensor], +) -> torch.Tensor: + weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes) + + return F.linear(input, weights, bias) + + +# Compare the optimized 1x16 and 2x8 cuda decompression/dequant kernels against +# the generic pytorch version. +# Just visual comparison. +def dequant_test(k: int, parts: torch.Tensor, nbooks: int, bits: int) -> None: + n = int(parts.sum().item()) + + device = torch.device("cuda:0") + + code_range = (1 << bits) // 2 + ingroups = 8 + + codes = torch.randint( + -code_range, + code_range, + size=(n, k // ingroups, nbooks), + dtype=get_int_dtype(bits), + device=device, + ) + + codebooks = torch.randn( + size=(parts.shape[0] * nbooks, 1 << bits, 1, 8), + dtype=torch.float16, + device=device, + ) + + count = 0 + for index in range(16): + for i in range(8): + for book in range(nbooks): + codebooks[book, index, 0, i] = count * (10**book) + count += 1 + + print("codes shape", codes.shape) + + for i in range(16): + for book in range(nbooks): + codes[0, i, book] = i + codes[0, -i, book] = i + + weights = dequantize_weight(codes, codebooks, None) + weights2 = ops.aqlm_dequant(codes, codebooks, parts) + + print("weights shape:", weights.shape) + print("weights2 shape:", weights2.shape) + + print("weights are:", weights) + print("weights2 are:", weights2) + + print("first 128 weights are", weights[0, 0:128].to(torch.int32)) + print("first 128 weights2 are:", weights2[0, 0:128].to(torch.int32)) + + print("last 128 weights are", weights[0, -128:]) + print("last 128 weights2 are:", weights2[0, -128:]) + + +def main(): + parser = FlexibleArgumentParser(description="Benchmark aqlm performance.") + + # Add arguments + parser.add_argument( + "--nbooks", type=int, default=1, help="Number of codebooks (default: 1)" + ) + parser.add_argument( + "--bits", + type=int, + default=16, + help="Number of bits per code element (default: 16)", + ) + parser.add_argument( + "--test", + type=bool, + default=False, + help="Run the decompression/dequant tester rather than benchmarking " + "(default: False)", + ) + + # Parse the arguments + args = parser.parse_args() + + # Extract values + nbooks = args.nbooks + bits = args.bits + + if args.test: + dequant_test(4096, torch.tensor((4096,)), nbooks, bits) + return + + # Otherwise, benchmark. + methods = [ + ops.aqlm_gemm, + dequant_out_scale, + generic_dequantize_gemm, + optimized_dequantize_gemm, + dequant_weight_scale, + torch_mult, + dequant_no_scale, + ] + + filename = f"./aqlm_benchmark_{nbooks}x{bits}.csv" + print(f"writing benchmarks to file {filename}") + with open(filename, "w") as f: + sys.stdout = f + + print("m | k | n | n parts", end="") + for method in methods: + print(f" | {method.__name__.replace('_', ' ')} (µs)", end="") + print("") + + # These are reasonable prefill sizes. + ksandpartions = ( + (4096, (4096, 4096, 4096)), + (4096, (4096,)), + (4096, (11008, 11008)), + (11008, (4096,)), + ) + + # reasonable ranges for m. + for m in [ + 1, + 2, + 4, + 8, + 10, + 12, + 14, + 16, + 24, + 32, + 48, + 52, + 56, + 64, + 96, + 112, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ]: + print(f"{m}", file=sys.__stdout__) + for ksp in ksandpartions: + run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits, methods) + + sys.stdout = sys.__stdout__ + + +def run_grid(m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, methods): + # I didn't see visible improvements from increasing these, but feel free :) + num_warmup_trials = 1 + num_trials = 1 + + num_calls = 100 + + # warmup. + for method in methods: + for _ in range(num_warmup_trials): + run_timing( + num_calls=num_calls, + m=m, + k=k, + parts=parts, + nbooks=nbooks, + bits=bits, + method=method, + ) + + n = parts.sum().item() + print(f"{m} | {k} | {n} | {parts.tolist()}", end="") + + for method in methods: + best_time_us = 1e20 + for _ in range(num_trials): + kernel_dur_ms = run_timing( + num_calls=num_calls, + m=m, + k=k, + parts=parts, + nbooks=nbooks, + bits=bits, + method=method, + ) + + kernel_dur_us = 1000 * kernel_dur_ms + + if kernel_dur_us < best_time_us: + best_time_us = kernel_dur_us + + print(f" | {kernel_dur_us:.0f}", end="") + + print("") + + +def run_timing( + num_calls: int, m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, method +) -> float: + n = int(parts.sum().item()) + + device = torch.device("cuda:0") + + input = torch.randn((1, m, k), dtype=torch.float16, device=device) + + code_range = (1 << bits) // 2 + ingroups = 8 + + codes = torch.randint( + -code_range, + code_range, + size=(n, k // ingroups, nbooks), + dtype=get_int_dtype(bits), + device=device, + ) + + codebooks = torch.randn( + size=(parts.shape[0] * nbooks, 1 << bits, 1, 8), + dtype=torch.float16, + device=device, + ) + + scales = torch.randn(size=(n, 1, 1, 1), dtype=torch.float16, device=device) + + # for comparison to just a pytorch mult. + weights = torch.randn((n, k), dtype=torch.float16, device=device) + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + start_event.record() + + if method is torch_mult: + for i in range(num_calls): + torch_mult(input, weights, scales) + else: + for i in range(num_calls): + method(input, codes, codebooks, scales, parts, None) + + end_event.record() + end_event.synchronize() + + dur_ms = start_event.elapsed_time(end_event) / num_calls + return dur_ms + + +if __name__ == "__main__": + sys.exit(main()) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_bitblas.py b/offline_benchmark_test/benchmarks/kernels/benchmark_bitblas.py new file mode 100644 index 0000000000000000000000000000000000000000..97ee060341373022b06c6ce62600d5c23c771c37 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_bitblas.py @@ -0,0 +1,242 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +# Copyright (c) Microsoft Corporation. +# Licensed under the MIT License. + +from vllm.model_executor.layers.quantization.utils.bitblas_utils import ( + MINIMUM_BITBLAS_VERSION, +) + +try: + import bitblas + + if bitblas.__version__ < MINIMUM_BITBLAS_VERSION: + raise ImportError( + "bitblas version is wrong. Please " + f"install bitblas>={MINIMUM_BITBLAS_VERSION}" + ) +except ImportError as e: + bitblas_import_exception = e + raise ValueError( + "Trying to use the bitblas backend, but could not import" + f"with the following error: {bitblas_import_exception}. " + "Please install bitblas through the following command: " + f"`pip install bitblas>={MINIMUM_BITBLAS_VERSION}`" + ) from bitblas_import_exception + +from bitblas import Matmul, MatmulConfig, auto_detect_nvidia_target + +from vllm.utils import FlexibleArgumentParser + +parser = FlexibleArgumentParser( + description="Benchmark BitBLAS int4 on a specific target." +) + +# Add arguments to the parser +parser.add_argument( + "--target", + type=str, + default=auto_detect_nvidia_target(), + help="Specify the target device for benchmarking.", +) +parser.add_argument( + "--group_size", type=int, default=None, help="Group size for grouped quantization." +) +parser.add_argument( + "--A_dtype", + type=str, + default="float16", + choices=["float16", "float32", "float64", "int32", "int8"], + help="Data type of activation A.", +) +parser.add_argument( + "--W_dtype", + type=str, + default="int4", + choices=[ + "float16", + "float32", + "float64", + "int32", + "int8", + "int4", + "int2", + "int1", + "nf4", + "fp4_e2m1", + ], + help="Data type of weight W.", +) +parser.add_argument( + "--accum_dtype", + type=str, + default="float16", + choices=["float16", "int32"], + help="Data type for accumulation.", +) +parser.add_argument( + "--out_dtype", + type=str, + default="float16", + choices=["float16", "float32", "int32", "int8"], + help="Data type for output.", +) +parser.add_argument( + "--layout", + type=str, + default="nt", + choices=["nt", "nn"], + help="Matrix layout, 'nt' for non-transpose A and transpose W.", +) +parser.add_argument( + "--with_bias", action="store_true", help="Include bias in the benchmark." +) +parser.add_argument( + "--with_scaling", + action="store_true", + help="Include scaling factor in the quantization.", +) +parser.add_argument( + "--with_zeros", action="store_true", help="Include zeros in the quantization." +) +parser.add_argument( + "--zeros_mode", + type=str, + default=None, + choices=["original", "rescale", "quantized"], + help="Specify the mode for calculating zeros.", +) + +# Parse the arguments +args = parser.parse_args() + +# Assign arguments to variables +target = args.target +A_dtype = args.A_dtype +W_dtype = args.W_dtype +accum_dtype = args.accum_dtype +out_dtype = args.out_dtype +layout = args.layout +with_bias = args.with_bias +group_size = args.group_size +with_scaling = args.with_scaling +with_zeros = args.with_zeros +zeros_mode = args.zeros_mode + +# Define a list of shared arguments that repeat in every config +shared_args = [ + A_dtype, + W_dtype, + out_dtype, + accum_dtype, + layout, + with_bias, + group_size, + with_scaling, + with_zeros, + zeros_mode, +] + +# Define just the (M, K, N) shapes in a more compact list +shapes = [ + # square test + (1, 16384, 16384), + # BLOOM-176B + (1, 43008, 14336), + (1, 14336, 14336), + (1, 57344, 14336), + (1, 14336, 57344), + # OPT-65B + (1, 9216, 9216), + (1, 36864, 9216), + (1, 9216, 36864), + (1, 22016, 8192), + # LLAMA-70B/65B + (1, 8192, 22016), + (1, 8192, 8192), + (1, 28672, 8192), + (1, 8192, 28672), + # square test + (16384, 16384, 16384), + # BLOOM-176B + (8192, 43008, 14336), + (8192, 14336, 14336), + (8192, 57344, 14336), + (8192, 14336, 57344), + # OPT-65B + (8192, 9216, 9216), + (8192, 36864, 9216), + (8192, 9216, 36864), + (8192, 22016, 8192), + # LLAMA-70B/65B + (8192, 8192, 22016), + (8192, 8192, 8192), + (8192, 28672, 8192), + (8192, 8192, 28672), +] + +# Build test shapes with all the shared arguments +test_shapes = [(MatmulConfig, Matmul, (*shape, *shared_args)) for shape in shapes] + +benchmark_sets = [] +benchmark_sets.extend(test_shapes) + +benchmark_results = {} +for config_class, operator, input_args in benchmark_sets: + config = config_class(*input_args) + matmul = operator(config, target=target, enable_tuning=True) + kernel_latency = matmul.profile_latency() + + print("Time cost is: {:.3f} ms".format(kernel_latency)) + + profile_config = { + f"{operator.__name__}-{'-'.join([str(i) for i in input_args])}": { + "BitBLAS_top20_latency": kernel_latency, + } + } + + benchmark_results.update(profile_config) + +# Define headers for the table +headers = [ + "PrimFunc", + "Input Arguments", + "BitBLAS Top20 Latency", +] + +# Calculate column widths for pretty printing +col_widths = [0, 0, 0] +for config_key, values in benchmark_results.items(): + args_split = config_key.split("-") + func_name = args_split[0] + input_args_str = "-".join(args_split[1:]) + col_widths[0] = max(col_widths[0], len(func_name) + 2, len(headers[0]) + 2) + col_widths[1] = max(col_widths[1], len(input_args_str) + 2, len(headers[1]) + 2) + col_widths[2] = max( + col_widths[2], + len(f"{values['BitBLAS_top20_latency']:.3f} ms") + 2, + len(headers[2]) + 2, + ) + # break only if you want to measure widths from a single example; + # otherwise, let it loop over all items. + +# Print header +for i, header in enumerate(headers): + headers[i] = header.ljust(col_widths[i]) +print("".join(headers)) +print("-" * sum(col_widths)) + +# Print rows +for config_key, values in benchmark_results.items(): + args_split = config_key.split("-") + func_name = args_split[0] + input_args_str = "-".join(args_split[1:]) + row = [ + func_name, + input_args_str, + f"{values['BitBLAS_top20_latency']:.3f} ms", + ] + row_str = "".join( + [str(cell).ljust(col_widths[idx]) for idx, cell in enumerate(row)] + ) + print(row_str) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_cutlass_fp4_moe.py b/offline_benchmark_test/benchmarks/kernels/benchmark_cutlass_fp4_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..35c20ee41b9a94c10f1367a45fe391cc15e9ffec --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_cutlass_fp4_moe.py @@ -0,0 +1,490 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +Benchmark the performance of the cutlass_moe_fp4 kernel vs the triton_moe +kernel. The cutlass_moe_fp4 kernel takes in fp4 quantized weights and 16-bit +activations. The triton_moe kernel takes in fp8 weights(tensor scaled to fp8) +and 16-bit activations. +""" + +import nvtx +import torch +import torch.utils.benchmark as benchmark + +from vllm import _custom_ops as ops +from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config +from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4 +from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk +from vllm.scalar_type import scalar_types +from vllm.utils import FlexibleArgumentParser + +WEIGHT_SHAPES_MOE = { + "nvidia/DeepSeek-R1-FP4": [ + [256, 8, 2048, 7168], + ], +} + +DEFAULT_MODELS = [ + "nvidia/DeepSeek-R1-FP4", +] + +DEFAULT_BATCH_SIZES = [4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048] +DEFAULT_TP_SIZES = [1] + +PER_ACT_TOKEN_OPTS = [False] +PER_OUT_CH_OPTS = [False] +FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max() +FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max + + +def to_fp8(tensor: torch.Tensor): + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to( + dtype=torch.float8_e4m3fn + ) + + +def bench_run( + results: list[benchmark.Measurement], + model: str, + num_experts: int, + topk: int, + per_act_token: bool, + per_out_ch: bool, + mkn: tuple[int, int, int], +): + label = "NVFP4 Blockscaled CUTLASS MOE vs FP8 Tensor Scaled Triton" + + sub_label = ( + "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, MKN=({})".format( + model, num_experts, topk, per_act_token, per_out_ch, mkn + ) + ) + + print(f"Testing: {sub_label}") + + (m, k, n) = mkn + + dtype = torch.half + device = "cuda" + a = torch.randn((m, k), device=device, dtype=dtype) / 10 + w1 = torch.randn((num_experts, 2 * n, k), device=device, dtype=dtype) / 10 + w2 = torch.randn((num_experts, k, n), device=device, dtype=dtype) / 10 + + _, a_fp8_scale = ops.scaled_fp8_quant(a) + + w1_fp8q = torch.empty( + (num_experts, 2 * n, k), device=device, dtype=torch.float8_e4m3fn + ) + w2_fp8q = torch.empty((num_experts, k, n), device=device, dtype=torch.float8_e4m3fn) + w1_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32) + w2_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32) + + for expert in range(num_experts): + w1_fp8q[expert], w1_fp8scale[expert] = ops.scaled_fp8_quant(w1[expert]) + w2_fp8q[expert], w2_fp8scale[expert] = ops.scaled_fp8_quant(w2[expert]) + + w1_fp8q_notransp = w1_fp8q.clone() + w2_fp8q_notransp = w2_fp8q.clone() + w1_fp8q = w1_fp8q.transpose(1, 2) + w2_fp8q = w2_fp8q.transpose(1, 2) + + score = torch.randn((m, num_experts), device=device, dtype=dtype) + + topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False) + + quant_blocksize = 16 + w1_blockscale = torch.empty( + (num_experts, 2 * n, k // quant_blocksize), + device=device, + dtype=torch.float8_e4m3fn, + ) + w2_blockscale = torch.empty( + (num_experts, k, n // quant_blocksize), device=device, dtype=torch.float8_e4m3fn + ) + + # n_b_scales = 2 * n if per_out_ch else 1 + # k_b_scales = k if per_out_ch else 1 + w1_fp4 = torch.empty((num_experts, 2 * n, k // 2), device=device, dtype=torch.uint8) + w2_fp4 = torch.empty((num_experts, k, n // 2), device=device, dtype=torch.uint8) + + w1_gs = torch.empty((num_experts,), device=device, dtype=torch.float32) + w2_gs = torch.empty((num_experts,), device=device, dtype=torch.float32) + a1_gs = torch.ones((num_experts,), device=device, dtype=torch.float32) + a2_gs = torch.ones((num_experts,), device=device, dtype=torch.float32) + + for expert in range(num_experts): + w1_e = w1[expert] + w2_e = w2[expert] + w1_amax = torch.abs(w1_e).max().to(torch.float32) + w2_amax = torch.abs(w2_e).max().to(torch.float32) + w1_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w1_amax + w2_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w2_amax + + w1_fp4[expert], w1_blockscale[expert] = ops.scaled_fp4_quant( + w1_e, w1_gs[expert] + ) + + w2_fp4[expert], w2_blockscale[expert] = ops.scaled_fp4_quant( + w2_e, w2_gs[expert] + ) + + def run_triton_moe( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_fp8_scale: torch.Tensor, + num_repeats: int, + ): + for _ in range(num_repeats): + fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_fp8_scale, + ) + + def run_cutlass_moe_fp4( + a: torch.Tensor, + w1_fp4: torch.Tensor, + w2_fp4: torch.Tensor, + w1_blockscale: torch.Tensor, + w2_blockscale: torch.Tensor, + w1_gs: torch.Tensor, + w2_gs: torch.Tensor, + a1_gs: torch.Tensor, + a2_gs: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + m: int, + n: int, + k: int, + e: int, + device: torch.device, + num_repeats: int, + ): + for _ in range(num_repeats): + with nvtx.annotate("cutlass_moe_fp4", color="green"): + cutlass_moe_fp4( + a=a, + a1_gscale=a1_gs, + a2_gscale=a2_gs, + w1_fp4=w1_fp4, + w1_blockscale=w1_blockscale, + w1_alphas=w1_gs, + w2_fp4=w2_fp4, + w2_blockscale=w2_blockscale, + w2_alphas=w2_gs, + topk_weights=topk_weights, + topk_ids=topk_ids, + m=m, + n=n, + k=k, + e=num_experts, + device=device, + ) + + def run_cutlass_from_graph( + a: torch.Tensor, + a1_gscale: torch.Tensor, + w1_fp4: torch.Tensor, + w1_blockscale: torch.Tensor, + w1_alphas: torch.Tensor, + a2_gscale: torch.Tensor, + w2_fp4: torch.Tensor, + w2_blockscale: torch.Tensor, + w2_alphas: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + m: int, + n: int, + k: int, + e: int, + device: torch.device, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return cutlass_moe_fp4( + a=a, + a1_gscale=a1_gs, + w1_fp4=w1_fp4, + w1_blockscale=w1_blockscale, + w1_alphas=w1_alphas, + a2_gscale=a2_gs, + w2_fp4=w2_fp4, + w2_blockscale=w2_blockscale, + w2_alphas=w2_alphas, + topk_weights=topk_weights, + topk_ids=topk_ids, + m=m, + n=n, + k=k, + e=num_experts, + device=device, + ) + + def run_triton_from_graph( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_fp8_scale: torch.Tensor, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_fp8_scale, + ) + + def replay_graph(graph, num_repeats): + for _ in range(num_repeats): + graph.replay() + torch.cuda.synchronize() + + cutlass_stream = torch.cuda.Stream() + cutlass_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(cutlass_graph, stream=cutlass_stream): + run_cutlass_from_graph( + a=a, + a1_gscale=a1_gs, + w1_fp4=w1_fp4, + w1_blockscale=w1_blockscale, + w1_alphas=w1_gs, + a2_gscale=a2_gs, + w2_fp4=w2_fp4, + w2_blockscale=w2_blockscale, + w2_alphas=w2_gs, + topk_weights=topk_weights, + topk_ids=topk_ids, + m=m, + n=n, + k=k, + e=num_experts, + device=device, + ) + torch.cuda.synchronize() + + triton_stream = torch.cuda.Stream() + triton_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(triton_graph, stream=triton_stream): + run_triton_from_graph( + a, + w1_fp8q_notransp, + w2_fp8q_notransp, + topk_weights, + topk_ids, + w1_fp8scale, + w2_fp8scale, + a_fp8_scale, + ) + torch.cuda.synchronize() + + min_run_time = 5 + num_warmup = 5 + num_runs = 25 + + globals = { + # Baseline params + "w1": w1, + "w2": w2, + "score": score, + "topk": topk, + "w1_fp8q_notransp": w1_fp8q_notransp, + "w2_fp8q_notransp": w2_fp8q_notransp, + "w1_fp8scale": w1_fp8scale, + "w2_fp8scale": w2_fp8scale, + "a_fp8_scale": a_fp8_scale, + # Cutlass params + "a": a, + "a1_gscale": a1_gs, + "w1_fp4": w1_fp4, + "w1_blockscale": w1_blockscale, + "w1_alphas": w1_gs, + "a2_gscale": a2_gs, + "w2_fp4": w2_fp4, + "w2_blockscale": w2_blockscale, + "w2_alphas": w2_gs, + "topk_weights": topk_weights, + "topk_ids": topk_ids, + "m": m, + "n": n, + "k": k, + "e": num_experts, + "device": device, + # cuda graph params + "cutlass_graph": cutlass_graph, + "triton_graph": triton_graph, + # Gen params + "num_runs": num_runs, + # Kernels + "run_triton_moe": run_triton_moe, + "run_cutlass_moe_fp4": run_cutlass_moe_fp4, + "replay_graph": replay_graph, + } + + # Warmup + run_triton_moe( + a, + w1_fp8q_notransp, + w2_fp8q_notransp, + topk_weights, + topk_ids, + w1_fp8scale, + w2_fp8scale, + a_fp8_scale, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_triton_moe(a, w1_fp8q_notransp, w2_fp8q_notransp, topk_weights, topk_ids, w1_fp8scale, w2_fp8scale, a_fp8_scale, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(triton_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(triton_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + + run_cutlass_moe_fp4( + a, + w1_fp4, + w2_fp4, + w1_blockscale, + w2_blockscale, + w1_gs, + w2_gs, + a1_gs, + a2_gs, + topk_weights, + topk_ids, + m, + n, + k, + num_experts, + device, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_cutlass_moe_fp4(a, w1_fp4, w2_fp4, w1_blockscale, w2_blockscale, w1_alphas, w2_alphas, a1_gscale, a2_gscale, topk_weights, topk_ids, m, n, k, e, device, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="cutlass_moe_fp4", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(cutlass_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(cutlass_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="cutlass_moe_fp4_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + +def main(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + results: list[benchmark.Measurement] = [] + + for model in args.models: + for tp in args.tp_sizes: + for layer in WEIGHT_SHAPES_MOE[model]: + num_experts = layer[0] + topk = layer[1] + size_k = layer[2] + size_n = layer[3] // tp + + if len(args.limit_k) > 0 and size_k not in args.limit_k: + continue + + if len(args.limit_n) > 0 and size_n not in args.limit_n: + continue + + for per_act_token in PER_ACT_TOKEN_OPTS: + for per_out_ch in PER_OUT_CH_OPTS: + for size_m in args.batch_sizes: + mkn = (size_m, size_k, size_n) + bench_run( + results, + model, + num_experts, + topk, + per_act_token, + per_out_ch, + mkn, + ) + + compare = benchmark.Compare(results) + compare.print() + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark NVFP4 CUTLASS MOE across specified models/shapes/batches" + ) + parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES_MOE.keys(), + ) + parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES) + parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + parser.add_argument("--limit-k", nargs="+", type=int, default=[]) + parser.add_argument("--limit-n", nargs="+", type=int, default=[]) + parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-act-token", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[]) + + args = parser.parse_args() + main(args) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py b/offline_benchmark_test/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py new file mode 100644 index 0000000000000000000000000000000000000000..1d4e730f99ae911535caa4cca17122656db06a4b --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py @@ -0,0 +1,383 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import torch +import torch.utils.benchmark as benchmark +from benchmark_shapes import WEIGHT_SHAPES_MOE + +from vllm import _custom_ops as ops +from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config +from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8 +from vllm.model_executor.layers.fused_moe.fused_moe import ( + fused_experts, + fused_topk, +) +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = [ + "nm-testing/Mixtral-8x7B-Instruct-v0.1", + "nm-testing/deepseekv2-lite", + "ibm-granite/granite-3.0-1b-a400m", + "ibm-granite/granite-3.0-3b-a800m", +] +DEFAULT_BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512] +DEFAULT_TP_SIZES = [1] + +PER_ACT_TOKEN_OPTS = [False] +PER_OUT_CH_OPTS = [False] + + +def to_fp8(tensor: torch.Tensor): + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to( + dtype=torch.float8_e4m3fn + ) + + +def bench_run( + results: list[benchmark.Measurement], + model: str, + num_experts: int, + topk: int, + per_act_token: bool, + per_out_ch: bool, + mkn: tuple[int, int, int], +): + label = "Quant Matmul" + + sub_label = ( + "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, MKN=({})".format( + model, num_experts, topk, per_act_token, per_out_ch, mkn + ) + ) + + print(f"Testing: {sub_label}") + + (m, k, n) = mkn + + dtype = torch.half + + a = torch.randn((m, k), device="cuda", dtype=dtype) / 10 + w1 = torch.randn((num_experts, 2 * n, k), device="cuda", dtype=dtype) / 10 + w2 = torch.randn((num_experts, k, n), device="cuda", dtype=dtype) / 10 + + _, a_scale = ops.scaled_fp8_quant(a) + + w1_q = torch.empty( + (num_experts, 2 * n, k), device="cuda", dtype=torch.float8_e4m3fn + ) + w2_q = torch.empty((num_experts, k, n), device="cuda", dtype=torch.float8_e4m3fn) + w1_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32) + w2_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32) + + for expert in range(num_experts): + w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(w1[expert]) + w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(w2[expert]) + + score = torch.randn((m, num_experts), device="cuda", dtype=dtype) + + topk_weights, topk_ids, token_expert_indices = fused_topk( + a, score, topk, renormalize=False + ) + + def run_triton_moe( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_scale: torch.Tensor, + num_repeats: int, + ): + for _ in range(num_repeats): + fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_scale, + ) + + def run_cutlass_moe( + a: torch.Tensor, + a_scale: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + per_act_token: bool, + num_repeats: int, + ): + for _ in range(num_repeats): + cutlass_moe_fp8( + a, + w1, + w2, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + per_act_token, + a1_scale=None, + ) + + def run_cutlass_from_graph( + a: torch.Tensor, + a_scale: torch.Tensor, + w1_q: torch.Tensor, + w2_q: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return cutlass_moe_fp8( + a, + w1_q, + w2_q, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + per_act_token, + a1_scale=None, + ) + + def run_triton_from_graph( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_scale: torch.Tensor, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_scale, + ) + + def replay_graph(graph, num_repeats): + for _ in range(num_repeats): + graph.replay() + torch.cuda.synchronize() + + cutlass_stream = torch.cuda.Stream() + cutlass_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(cutlass_graph, stream=cutlass_stream): + run_cutlass_from_graph( + a, + a_scale, + w1_q, + w2_q, + w1_scale, + w2_scale, + topk_weights, + topk_ids, + ) + torch.cuda.synchronize() + + triton_stream = torch.cuda.Stream() + triton_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(triton_graph, stream=triton_stream): + run_triton_from_graph( + a, + w1_q, + w2_q, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + a_scale, + ) + torch.cuda.synchronize() + + min_run_time = 5 + num_warmup = 5 + num_runs = 25 + + globals = { + # Baseline params + "w1": w1, + "w2": w2, + "score": score, + "topk": topk, + # Cutlass params + "a_scale": a_scale, + "w1_q": w1_q, + "w2_q": w2_q, + "w1_scale": w1_scale, + "w2_scale": w2_scale, + "per_act_token": per_act_token, + # cuda graph params + "cutlass_graph": cutlass_graph, + "triton_graph": triton_graph, + # Gen params + "a": a, + "topk_weights": topk_weights, + "topk_ids": topk_ids, + "num_runs": num_runs, + # Kernels + "run_triton_moe": run_triton_moe, + "run_cutlass_moe": run_cutlass_moe, + "replay_graph": replay_graph, + } + + # Warmup + run_triton_moe( + a, + w1_q, + w2_q, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + a_scale, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_triton_moe(a, w1_q, w2_q, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(triton_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(triton_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + run_cutlass_moe( + a, + a_scale, + w1_q, + w2_q, + w1_scale, + w2_scale, + topk_weights, + topk_ids, + per_act_token, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, per_act_token, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="grouped_gemm_moe", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(cutlass_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(cutlass_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="grouped_gemm_moe_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + +def main(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + results: list[benchmark.Measurement] = [] + + for model in args.models: + for tp in args.tp_sizes: + for layer in WEIGHT_SHAPES_MOE[model]: + num_experts = layer[0] + topk = layer[1] + size_k = layer[2] + size_n = layer[3] // tp + + if len(args.limit_k) > 0 and size_k not in args.limit_k: + continue + + if len(args.limit_n) > 0 and size_n not in args.limit_n: + continue + + for per_act_token in PER_ACT_TOKEN_OPTS: + for per_out_ch in PER_OUT_CH_OPTS: + for size_m in DEFAULT_BATCH_SIZES: + mkn = (size_m, size_k, size_n) + bench_run( + results, + model, + num_experts, + topk, + per_act_token, + per_out_ch, + mkn, + ) + + compare = benchmark.Compare(results) + compare.print() + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark Marlin across specified models/shapes/batches" + ) + parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES_MOE.keys(), + ) + parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES) + parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + parser.add_argument("--limit-k", nargs="+", type=int, default=[]) + parser.add_argument("--limit-n", nargs="+", type=int, default=[]) + parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-act-token", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[]) + + args = parser.parse_args() + main(args) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_layernorm.py b/offline_benchmark_test/benchmarks/kernels/benchmark_layernorm.py new file mode 100644 index 0000000000000000000000000000000000000000..69978ec6b23e94ce21d8f3dfb633317b3803e6ef --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_layernorm.py @@ -0,0 +1,93 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import time + +import torch + +from vllm.model_executor.layers.layernorm import RMSNorm +from vllm.platforms import current_platform +from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser + + +@torch.inference_mode() +def main( + num_tokens: int, + hidden_size: int, + add_residual: bool, + dtype: torch.dtype, + seed: int = 0, + do_profile: bool = False, + num_warmup_iters: int = 5, + num_iters: int = 100, +) -> None: + current_platform.seed_everything(seed) + torch.set_default_device("cuda") + + layer = RMSNorm(hidden_size).to(dtype=dtype) + layer.weight.data.normal_(mean=1.0, std=0.1) + scale = 1 / (2 * hidden_size) + x = torch.randn(num_tokens, hidden_size, dtype=dtype) + x *= scale + residual = torch.randn_like(x) * scale if add_residual else None + + def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: + torch.cuda.synchronize() + if profile: + torch.cuda.cudart().cudaProfilerStart() + start_time = time.perf_counter() + + for _ in range(num_iters): + layer(x, residual) + torch.cuda.synchronize() + + end_time = time.perf_counter() + if profile: + torch.cuda.cudart().cudaProfilerStop() + return (end_time - start_time) / num_iters + + # Warmup. + print("Warming up...") + run_benchmark = run_cuda_benchmark + run_benchmark(num_iters=num_warmup_iters, profile=False) + + # Benchmark. + if do_profile: + latency = run_benchmark(num_iters=1, profile=True) + else: + latency = run_benchmark(num_iters=num_iters, profile=False) + print(f"Kernel running time: {latency * 1000000:.3f} us") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser(description="Benchmark the layernorm kernel.") + parser.add_argument("--num-tokens", type=int, default=4096) + parser.add_argument("--hidden-size", type=int, default=8192) + parser.add_argument("--add-residual", action="store_true") + parser.add_argument( + "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half" + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--profile", action="store_true") + parser.add_argument("--num-warmup-iters", type=int, default=5) + parser.add_argument( + "--num-iters", + type=int, + default=100, + help="Number of benchmark iterations. " + "If --profile is set, this number is ignored", + ) + + args = parser.parse_args() + print(args) + + main( + num_tokens=args.num_tokens, + hidden_size=args.hidden_size, + add_residual=args.add_residual, + dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype], + seed=args.seed, + do_profile=args.profile, + num_warmup_iters=args.num_warmup_iters, + num_iters=args.num_iters, + ) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_lora.py b/offline_benchmark_test/benchmarks/kernels/benchmark_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..3d38d4b3534e8993abc8aff85712230e7262a740 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_lora.py @@ -0,0 +1,1065 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import json +import pickle +import time +from dataclasses import dataclass +from enum import Enum, auto +from itertools import product +from pathlib import Path +from typing import Any, Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import ArgPool, Bench, CudaGraphBenchParams +from weight_shapes import WEIGHT_SHAPES + +from vllm.triton_utils import HAS_TRITON + +if HAS_TRITON: + from vllm.lora.ops.triton_ops import LoRAKernelMeta, lora_expand, lora_shrink + from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT + +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_TP_SIZES = [1] +DEFAULT_BATCH_SIZES = [ + 1, + 16, + 32, + 64, + 128, + 192, + 256, + 320, + 384, + 448, + 512, + 640, + 768, + 896, + 1024, + 2048, + 3072, + 4096, + 5120, + 6144, + 7168, + 8192, +] +DEFAULT_HIDDEN_SIZES = [1024, 2048, 4096, 8192, 16384] +DEFAULT_LORA_RANKS = [16] +DEFAULT_NUM_LORAS = [1, 2, 3, 4] +DEFAULT_SORT_BY_LORA_IDS = [False, True] +DEFAULT_SEQ_LENGTHS = [1] +DEFAULT_EXPAND_FN_ADD_INPUTS = [True, False] + + +# Utilities +def dtype_to_str(dtype: torch.dtype): + if dtype == torch.float16: + return "f16" + if dtype == torch.bfloat16: + return "bf16" + if dtype == torch.float32: + return "f32" + raise ValueError(f"Unsupported dtype {dtype}") + + +def make_rand_lora_weight_tensor( + k: int, n: int, num_loras: int, dtype: torch.dtype, device: str = "cuda" +) -> torch.Tensor: + # LoRA weights column major + return torch.rand((num_loras, n, k), dtype=dtype).to(device) + + +def make_rand_tensors( + a_shape: tuple[int], + b_shape: tuple[int], + c_shape: tuple[int], + a_dtype: torch.dtype, + b_dtype: torch.dtype, + c_dtype: torch.dtype, + num_slices: int, + device: str = "cuda", +) -> tuple[torch.Tensor, list[torch.Tensor], torch.Tensor]: + """ + Make LoRA input/output matrices. + """ + A = torch.rand(a_shape, dtype=a_dtype).to(device) + + # LoRA weights column major + Bs = [torch.rand(b_shape, dtype=b_dtype).to(device) for _ in range(num_slices)] + + C = torch.zeros(c_shape, dtype=c_dtype).to(device) + return A, Bs, C + + +def make_prompt_lora_mapping( + num_prompts: int, num_active_loras: int, sort_by_lora_id: bool, device: str +) -> torch.Tensor: + """ + All prompts are mapped to a LoRA ID in range [0, num_active_loras). + where 0 refers to first lora, 1 refers to second lora and so on. + """ + assert num_active_loras > 0 + + if not sort_by_lora_id: + return torch.randint(0, num_active_loras, (num_prompts,), dtype=torch.long) + + # Divide LoRAs equally and in order. + part_size = num_prompts // num_active_loras + part_size = max(part_size, 1) + + lora_id = 0 + prompt_lora_mapping = [] + while len(prompt_lora_mapping) < num_prompts: + prompt_lora_mapping.extend([lora_id] * part_size) + lora_id = lora_id + 1 if lora_id + 1 < num_active_loras else lora_id + return torch.tensor( + prompt_lora_mapping[:num_prompts], dtype=torch.long, device=device + ) + + +def make_token_lora_mapping( + num_tokens: int, + num_prompts: int, + prompt_lora_mapping: torch.Tensor, + seq_len_tensor: torch.Tensor, + device: str, +): + """ + Make token_lora_mapping from prompt_lora_mapping and seq_lens_tensor + """ + assert prompt_lora_mapping.shape[0] == num_prompts + + # token to lora index mapping + token_lora_mapping = [0] * num_tokens + current_offset = 0 + for b_id in range(num_prompts): + lora_index = prompt_lora_mapping[b_id].item() + s = current_offset + e = s + seq_len_tensor[b_id].item() + token_lora_mapping[s:e] = [lora_index] * (e - s) + current_offset += seq_len_tensor[b_id].item() + + return torch.tensor(token_lora_mapping, dtype=torch.long, device=device) + + +def ref_group_gemm( + ref_out: torch.Tensor, + input: torch.Tensor, + lora_weights: list[torch.Tensor], + seq_lens_cpu: torch.Tensor, + prompt_lora_mapping_cpu: torch.Tensor, + scaling: float, + add_inputs: Optional[bool], +): + """ + Torch group gemm reference implementation to test correctness of + benchmarking operations. + """ + batches = seq_lens_cpu.size(0) + out_list = [] + current_offset = 0 + for lora_index, b_length in zip(range(batches), seq_lens_cpu): + x = input[current_offset : b_length + current_offset, :] + current_offset += b_length + w = lora_weights[prompt_lora_mapping_cpu[lora_index]] + result = torch.nn.functional.linear(x, w) + result *= scaling + out_list.append(result) + + cat_result = torch.cat(out_list, dim=0) + + if add_inputs: + ref_out += cat_result + else: + ref_out.copy_(cat_result) + + +class OpType(Enum): + """ + LoRA Ops to benchmark and its properties. + """ + + LORA_SHRINK = auto() + LORA_EXPAND = auto() + + @staticmethod + def from_str(s: str) -> "OpType": + if s.lower() == "lora_shrink": + return OpType.LORA_SHRINK + if s.lower() == "lora_expand": + return OpType.LORA_EXPAND + raise ValueError(f"Unrecognized str {s} to convert to OpType") + + def is_shrink_fn(self) -> bool: + return self in [OpType.LORA_SHRINK] + + def is_expand_fn(self) -> bool: + return self in [OpType.LORA_EXPAND] + + def num_slices(self) -> list[int]: + return [1, 2, 3] + + def mkn( + self, batch_size: int, seq_length: int, hidden_size: int, lora_rank: int + ) -> tuple[int, int, int]: + num_tokens = batch_size * seq_length + if self.is_shrink_fn(): + m = num_tokens + k = hidden_size + n = lora_rank + else: + assert self.is_expand_fn() + m = num_tokens + k = lora_rank + n = hidden_size + return m, k, n + + def matmul_dtypes( + self, op_dtype: torch.dtype + ) -> tuple[torch.dtype, torch.dtype, torch.dtype]: + """ + return a type, b type and c type for A x B = C + """ + if self.is_shrink_fn(): + return op_dtype, op_dtype, torch.float32 + else: + assert self.is_expand_fn() + return torch.float32, op_dtype, op_dtype + + def matmul_shapes( + self, + batch_size: int, + seq_length: int, + hidden_size: int, + lora_rank: int, + num_loras: int, + num_slices: int, + ) -> tuple[tuple[int], tuple[int], tuple[int]]: + """ + Given num_slices, return the shapes of the A, B, and C matrices + in A x B = C, for the op_type + """ + m, k, n = self.mkn(batch_size, seq_length, hidden_size, lora_rank) + + b_shape = (num_loras, n, k) # col-major + if self in [OpType.LORA_SHRINK]: + # LoRA shrink kernels support num_slices inherently in the kernel. + return ((m, k), b_shape, (num_slices, m, n)) + if self in [OpType.LORA_EXPAND]: + # LoRA expand kernels support num_slices inherently in the kernel + return ((num_slices, m, k), b_shape, (m, n * num_slices)) + raise ValueError(f"Unrecognized op_type {self}") + + def bench_fn(self) -> Callable: + if self == OpType.LORA_SHRINK: + return lora_shrink + if self == OpType.LORA_EXPAND: + return lora_expand + + raise ValueError(f"Unrecognized optype {self}") + + def run_ref_group_gemm( + self, + output: torch.Tensor, + input: torch.Tensor, + lora_weights: list[torch.Tensor], + **kwargs, + ) -> Callable: + """Each benchmark operation expects the input, lora_weights and outputs + in a slightly different format. Refer to self.matmul_shapes(). + run_ref_group_gemm accounts for those differences in executing a + reference group gemm for correctness testing. + """ + w_dtype = lora_weights[0].dtype + num_slices = len(lora_weights) + if self in [OpType.LORA_SHRINK]: + for slice_idx in range(num_slices): + ref_group_gemm( + ref_out=output[slice_idx, :], + input=input, + lora_weights=lora_weights[slice_idx], + **kwargs, + ) + elif self in [OpType.LORA_EXPAND]: + hidden_size = lora_weights[0].shape[1] + for slice_idx in range(num_slices): + slice_offset = slice_idx * hidden_size + ref_group_gemm( + ref_out=output[:, slice_offset : slice_offset + hidden_size], + input=input[slice_idx].clone().to(dtype=w_dtype), + lora_weights=lora_weights[slice_idx], + **kwargs, + ) + else: + raise ValueError(f"Unrecognized optype {self}") + + +@dataclass +class BenchmarkContext: + """ + LoRA benchmark context + """ + + batch_size: int + hidden_size: int + num_loras: int + num_active_loras: int + lora_rank: int + sort_by_lora_id: bool + dtype: torch.dtype + seq_length: Optional[int] = None + num_slices: Optional[int] = None # num_slices for slice based ops + + def with_seq_length(self, seq_length: int) -> "BenchmarkContext": + ctx = copy.copy(self) + ctx.seq_length = seq_length + return ctx + + def with_num_slices(self, num_slices: int) -> "BenchmarkContext": + ctx = copy.copy(self) + ctx.num_slices = num_slices + return ctx + + def bench_label(self) -> str: + return f"lora-{self.dtype}" + + def bench_sublabel(self, op_type: OpType) -> str: + m, k, n = op_type.mkn( + self.batch_size, self.seq_length, self.hidden_size, self.lora_rank + ) + desc = { + "bs": self.batch_size, + "sl": self.seq_length, + "m": m, + "k": k, + "n": n, + "num_loras": self.num_loras, + "sort_by_lora": self.sort_by_lora_id, + "num_slices": self.num_slices, + } + return json.dumps(desc) + + +@dataclass +class BenchmarkTensors: + """ + Input/Output tensors used for benchmarks + """ + + # matmul tensors + input: torch.Tensor + lora_weights_lst: list[torch.Tensor] + output: torch.Tensor + # LoRA kernel metadata + lora_kernel_meta: LoRAKernelMeta + # Metadata tensors used in testing correctness + seq_lens: torch.Tensor + prompt_lora_mapping: torch.Tensor + + def io_types(self) -> str: + return ( + f"{dtype_to_str(self.input.dtype)}x" + f"{dtype_to_str(self.lora_weights_lst[0].dtype)}=>" + f"{dtype_to_str(self.output.dtype)}" + ) + + @staticmethod + def make( + ctx: BenchmarkContext, op_type: OpType, device: str = "cuda" + ) -> "BenchmarkTensors": + # Make input / output matmul tensors. + a_shape, b_shape, c_shape = op_type.matmul_shapes( + ctx.batch_size, + ctx.seq_length, + ctx.hidden_size, + ctx.lora_rank, + ctx.num_loras, + ctx.num_slices, + ) + a_type, b_type, c_type = op_type.matmul_dtypes(ctx.dtype) + input_tensor, lora_weights, output_tensor = make_rand_tensors( + a_shape, b_shape, c_shape, a_type, b_type, c_type, num_slices=ctx.num_slices + ) + + # Make metadata tensors. + # Keep the metadata tensors in the CPU for further processing if needed. + # The tensors get moved to the GPU before benchmarking. + assert ctx.num_active_loras <= ctx.num_loras + total_tokens = ctx.batch_size * ctx.seq_length + + # Make metadata tensors involved in correctness testing. + # Prepare seq lens tensor + seq_len_tensor = torch.randint( + ctx.seq_length, ctx.seq_length + 1, (ctx.batch_size,) + ) + assert total_tokens == seq_len_tensor.sum() + # Prepare prompt lora indices tensor + prompt_lora_indices_tensor = make_prompt_lora_mapping( + ctx.batch_size, ctx.num_active_loras, ctx.sort_by_lora_id, "cpu" + ) + + # Make LoRAKernelMeta + token_lora_indices_tensor = make_token_lora_mapping( + total_tokens, + ctx.batch_size, + prompt_lora_indices_tensor, + seq_len_tensor, + "cpu", + ) + lora_kernel_meta = LoRAKernelMeta.make( + max_loras=ctx.num_loras, + max_num_tokens=token_lora_indices_tensor.size(0), + device="cpu", + ) + lora_kernel_meta.prepare_tensors(token_lora_mapping=token_lora_indices_tensor) + + return BenchmarkTensors( + input_tensor, + lora_weights, + output_tensor, + lora_kernel_meta, + seq_len_tensor, + prompt_lora_indices_tensor, + ) + + def sanity_check(self) -> None: + """ + Fails asserts when non-conformality is detected. + """ + num_tokens = self.input.shape[-2] + # check metadata tensors + assert torch.sum(self.seq_lens) == num_tokens + num_seqs = self.seq_lens.shape[0] + # assert self.seq_start_loc.shape[0] == num_seqs + assert self.prompt_lora_mapping.shape[0] == num_seqs + assert self.lora_kernel_meta.token_lora_mapping.shape[0] == num_tokens + + def to_device(self, device: str): + """ + Transfer tensors to device if the tensors aren't already on the device + """ + + def to_device(tensor: torch.Tensor): + if tensor.device != device: + tensor = tensor.to(device=device) + return tensor + + self.input = to_device(self.input) + self.output = to_device(self.output) + self.seq_lens = to_device(self.seq_lens) + self.prompt_lora_mapping = to_device(self.prompt_lora_mapping) + for i in range(len(self.lora_weights_lst)): + self.lora_weights_lst[i] = to_device(self.lora_weights_lst[i]) + + # LoRA meta + for field_name in LoRAKernelMeta.__dataclass_fields__: + field = getattr(self.lora_kernel_meta, field_name) + assert isinstance(field, torch.Tensor) + setattr(self.lora_kernel_meta, field_name, to_device(field)) + + def metadata(self) -> tuple[int, int, int]: + """ + Return num_seqs, num_tokens and max_seq_len + """ + num_seqs = self.seq_lens.shape[0] + num_tokens = self.lora_kernel_meta.token_lora_mapping.shape[0] + max_seq_len = torch.max(self.seq_lens).item() + num_slices = len(self.lora_weights_lst) + return num_seqs, num_tokens, max_seq_len, num_slices + + def as_lora_shrink_kwargs(self) -> dict[str, Any]: + self.sanity_check() + self.to_device(self.input.device) + + _, num_tokens, _, num_slices = self.metadata() + + # Sanity check matrix shapes. + i_shape, lw_shape, o_shape = ( + self.input.shape, + self.lora_weights_lst[0].shape, + self.output.shape, + ) + # Expected input shape [num_tokens, hidden_size] + assert len(i_shape) == 2 + assert i_shape[0] == num_tokens + hidden_size = i_shape[1] + # Expected lora weight shape [num_loras, lora_rank, hidden_size] + assert len(lw_shape) == 3 + assert lw_shape[2] == hidden_size + lora_rank = lw_shape[1] + # Expected output shape [num_slices, num_tokens, lora_rank] + assert len(o_shape) == 3 + assert o_shape == (num_slices, num_tokens, lora_rank) + + return { + "inputs": self.input, + "lora_a_weights": self.lora_weights_lst, + "output_tensor": self.output, + "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, + "token_indices_sorted_by_lora_ids": ( + self.lora_kernel_meta.token_indices_sorted_by_lora_ids + ), + "num_tokens_per_lora": self.lora_kernel_meta.num_tokens_per_lora, + "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc, + "lora_ids": self.lora_kernel_meta.active_lora_ids, + "scaling": 1.0, + } + + def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]: + self.sanity_check() + self.to_device(self.input.device) + + _, num_tokens, _, num_slices = self.metadata() + + # Sanity check matrix shapes. + i_shape, lw_shape, o_shape = ( + self.input.shape, + self.lora_weights_lst[0].shape, + self.output.shape, + ) + # Expected input shape : [num_slices, num_tokens, lora_rank] + assert len(i_shape) == 3 + assert i_shape[0] == num_slices + assert i_shape[1] == num_tokens + lora_rank = i_shape[2] + # Expected lora weight shape : [num_lora, hidden_size, lora_rank] + assert len(lw_shape) == 3 + assert lw_shape[2] == lora_rank + hidden_size = lw_shape[1] + # Expected output shape : [num_tokens, hidden_size * num_slices] + assert len(o_shape) == 2 + assert o_shape == (num_tokens, hidden_size * num_slices) + + return { + "inputs": self.input, + "lora_b_weights": self.lora_weights_lst, + "output_tensor": self.output, + "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, + "token_indices_sorted_by_lora_ids": ( + self.lora_kernel_meta.token_indices_sorted_by_lora_ids + ), + "num_tokens_per_lora": self.lora_kernel_meta.num_tokens_per_lora, + "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc, + "lora_ids": self.lora_kernel_meta.active_lora_ids, + "offset_start": 0, + "add_inputs": add_inputs, + } + + def bench_fn_kwargs( + self, op_type: OpType, add_inputs: Optional[bool] = None + ) -> dict[str, Any]: + if op_type.is_shrink_fn(): + assert add_inputs is None + else: + assert add_inputs is not None + + if op_type == OpType.LORA_SHRINK: + return self.as_lora_shrink_kwargs() + if op_type == OpType.LORA_EXPAND: + return self.as_lora_expand_kwargs(add_inputs) + raise ValueError(f"Unrecognized optype {self}") + + def test_correctness( + self, op_type: OpType, expand_fn_add_inputs: Optional[bool] + ) -> bool: + """ + Test correctness of op_type implementation against a grouped gemm + reference implementation. + """ + seq_lens_cpu = self.seq_lens.to(device="cpu") + prompt_lora_mapping_cpu = self.prompt_lora_mapping.to(device="cpu") + ref_output = self.output.clone() + + self.output.zero_() + op_type.bench_fn()(**self.bench_fn_kwargs(op_type, expand_fn_add_inputs)) + + op_type.run_ref_group_gemm( + ref_output, + self.input, + self.lora_weights_lst, + seq_lens_cpu=seq_lens_cpu, + prompt_lora_mapping_cpu=prompt_lora_mapping_cpu, + scaling=1.0, + add_inputs=expand_fn_add_inputs, + ) + + rtol, atol = { + torch.float16: (6e-2, 6e-2), + torch.bfloat16: (6e-2, 6e-2), + torch.float32: (1e-2, 1e-2), + }[self.output.dtype] + + return torch.allclose(ref_output, self.output, rtol=rtol, atol=atol) + + +def bench_optype( + ctx: BenchmarkContext, + arg_pool_size: int, + op_type: OpType, + cuda_graph_nops: Optional[int] = None, + expand_fn_add_inputs: Optional[bool] = None, + test_correctness: bool = False, +) -> TMeasurement: + assert arg_pool_size >= 1 + if op_type.is_shrink_fn(): + assert expand_fn_add_inputs is None + else: + assert expand_fn_add_inputs is not None + + # BenchmarkContext -> BenchmarkTensors + bench_tensors: list[BenchmarkTensors] = [ + BenchmarkTensors.make(ctx, op_type) for _ in range(arg_pool_size) + ] + for bt in bench_tensors: + bt.sanity_check() + + # Test correctness of our implementation. + if test_correctness: + assert all( + [bt.test_correctness(op_type, expand_fn_add_inputs) for bt in bench_tensors] + ) + + # BenchmarkTensors -> dict (kwargs) + kwargs_list = [ + bt.bench_fn_kwargs(op_type, add_inputs=expand_fn_add_inputs) + for bt in bench_tensors + ] + + # Clear LoRA optimization hash-maps. + _LORA_A_PTR_DICT.clear() + _LORA_B_PTR_DICT.clear() + # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are setup + for kwargs in kwargs_list: + op_type.bench_fn()(**kwargs) + torch.cuda.synchronize() + + # Merge into a single kwargs and qualify arguments as ArgPool + kwargs = {k: ArgPool([]) for k in kwargs_list[0]} + for _kwargs in kwargs_list: + for k, v in _kwargs.items(): + kwargs[k].values.append(v) + + describe_args = ( + f"add_inputs={expand_fn_add_inputs}" if expand_fn_add_inputs is not None else "" + ) + description = f"{op_type.name}({describe_args}) ({bench_tensors[0].io_types()})" + + cuda_graph_params = None + if cuda_graph_nops: + cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) + timer = None + with Bench( + cuda_graph_params, + ctx.bench_label(), + ctx.bench_sublabel(op_type), + description, + op_type.bench_fn(), + **kwargs, + ) as bench: + timer = bench.run() + return timer + + +def bench_torch_mm( + ctx: BenchmarkContext, + arg_pool_size: int, + op_type: OpType, + cuda_graph_nops: Optional[int] = None, +) -> TMeasurement: + """ + Benchmark basic torch.mm as a roofline. + + When all the input tokens have the same LoRA ID, the LoRA kernels are just + a matmul. This torch.mm benchmark serves as a roofline for that case. + + input op_type is used in determining the m, k, n dimensions for the matmul. + """ + + batch_size, hidden_size, lora_rank, seq_length, dtype = ( + ctx.batch_size, + ctx.hidden_size, + ctx.lora_rank, + ctx.seq_length, + ctx.dtype, + ) + + m, k, n = op_type.mkn(batch_size, seq_length, hidden_size, lora_rank) + # For a fairer comparison. + n = n * ctx.num_slices + + # Get matmul input and output tensors for A x B = C + As, Bs, Cs = [], [], [] + for _ in range(arg_pool_size): + As.append(torch.rand((m, k), dtype=dtype).to("cuda")) + Bs.append(torch.rand((n, k), dtype=dtype).to("cuda").t()) + Cs.append(torch.rand((m, n), dtype=dtype).to("cuda")) + + # Make torch.mm kwargs + mm_kwargs = {"input": ArgPool(As), "mat2": ArgPool(Bs), "out": ArgPool(Cs)} + + description = ( + f"single-lora roofline using torch.mm ({dtype_to_str(dtype)}" + f"x{dtype_to_str(dtype)}" + f"=>{dtype_to_str(dtype)})" + ) + cuda_graph_params = None + if cuda_graph_nops: + cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) + with Bench( + cuda_graph_params, + ctx.bench_label(), + ctx.bench_sublabel(op_type), + description, + torch.mm, + **mm_kwargs, + ) as bench: + return bench.run() + + +# runner +def use_cuda_graph_recommendation() -> str: + return """ + Triton kernels have a significant launch overhead with + launched directly via python. This overhead is more noticeable + for small the problem sizes. For these cases, it is recommended + to use the script with `--cuda-graph-nops N` to benchmark N + consecutive invocations of the benchmarking operations from + inside a CUDA Graph. Note that the returned measurement is for N + invocations of the operation. + """ + + +def print_timers(timers: list[TMeasurement], args: Optional[argparse.Namespace] = None): + compare = TBenchmark.Compare(timers) + compare.print() + + if args and args.cuda_graph_nops: + print( + f"Note : The timings reported above is for {args.cuda_graph_nops} " + "consecutive invocations of the benchmarking functions. " + f"Please divide by {args.cuda_graph_nops} for single invocation " + "timings." + ) + + print( + "Note on Comparison with torch.mm : The torch.mm numbers are " + "benchmark numbers of a simple matmul emulating the single lora " + "case. It is provided as a roofline for comparing our LoRA Kernel " + "implementations. It is expected that the LoRA kernels will be " + "slower than torch.mm in cases where num_loras is big. But for " + "small num_loras the goal should be to match the torch.mm numbers." + ) + + +def run(args: argparse.Namespace, bench_ctxs: list[BenchmarkContext]): + if args.cuda_graph_nops is not None: + assert args.cuda_graph_nops > 0 + print(f"Benchmarking {args.cuda_graph_nops} invocations inside a CUDA Graph") + else: + print(f"CUDA Graphs not enabled.\n{use_cuda_graph_recommendation()}") + + timers = [] + for bench_ctx in bench_ctxs: + for seq_len in args.seq_lengths: + bench_ops: list[OpType] = args.op_types + seq_len_timers = [] + for bench_op in bench_ops: + for num_slices in bench_op.num_slices(): + _ctx = bench_ctx.with_seq_length(seq_len).with_num_slices( + num_slices + ) + # Benchmark torch.mm as a roofline + seq_len_timers.append( + bench_torch_mm( + _ctx, args.arg_pool_size, bench_op, args.cuda_graph_nops + ) + ) + + # Benchmark bench_op + expand_fn_add_inputs = ( + [None] if bench_op.is_shrink_fn() else args.expand_fn_add_inputs + ) + for add_input_arg in expand_fn_add_inputs: + seq_len_timers.append( + bench_optype( + _ctx, + args.arg_pool_size, + bench_op, + args.cuda_graph_nops, + add_input_arg, + args.test_correctness, + ) + ) + + print_timers(seq_len_timers) + timers.extend(seq_len_timers) + + # Result stdout dump + print("== All Results ====") + print_timers(timers, args) + + if args.output_directory: + # Result file dump + od = Path(args.output_directory) + if not od.exists(): + od.mkdir() + + timestamp = int(time.time()) + pkl_file = od / f"lora_bench-{timestamp}.pkl" + print(f"Writing benchmarks to {pkl_file}") + with open(pkl_file, "wb") as f: + pickle.dump(timers, f) + + +def as_benchmark_contexts( + hidden_sizes: list[int], lora_ranks: list[int], args: argparse.Namespace +) -> list[BenchmarkContext]: + ctxs: list[BenchmarkContext] = [] + for batch_size, hidden_size, lora_rank, num_loras, sort_by_lora_id in product( # noqa + args.batch_sizes, + list(hidden_sizes), + lora_ranks, + args.num_loras, + args.sort_by_lora_id, + ): + ctxs.append( + BenchmarkContext( + batch_size=batch_size, + hidden_size=hidden_size, + lora_rank=lora_rank, + num_loras=num_loras, + num_active_loras=args.num_active_loras + if args.num_active_loras + else num_loras, + # To be filled based on the OpType to benchmark + seq_length=None, + sort_by_lora_id=sort_by_lora_id, + dtype=args.dtype, + # To be filled based on the OpType to benchmark + num_slices=None, + ) + ) + + return ctxs + + +def run_list_bench(args: argparse.Namespace): + print(args) + + print( + "List bench :\n" + f" Hidden Sizes {args.hidden_sizes}" + f" LoRA Ranks {args.lora_ranks}" + ) + + # Get all benchmarking contexts + bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=args.hidden_sizes, lora_ranks=args.lora_ranks, args=args + ) + + run(args, bench_contexts) + + +def run_range_bench(args: argparse.Namespace): + print(args) + + hidden_sizes = list( + range( + args.hidden_sizes_start, + args.hidden_sizes_end + 1, + args.hidden_sizes_increment, + ) + ) + lora_ranks = list( + range(args.lora_ranks_start, args.lora_ranks_end + 1, args.lora_ranks_increment) + ) + + print(f"Range bench :\n Hidden Sizes {hidden_sizes} LoRA Ranks {lora_ranks}") + + # Get all benchmarking contexts + bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=hidden_sizes, lora_ranks=lora_ranks, args=args + ) + + run(args, bench_contexts) + + +def run_model_bench(args: argparse.Namespace): + print(args) + + def hidden_sizes_from_model(model: str, tp_size: int) -> set[int]: + hidden_sizes = set() + for KN, tp_split_dim in WEIGHT_SHAPES[model]: + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + hidden_sizes.add(KN[1]) + return hidden_sizes + + # Get all hidden sizes + hidden_sizes: set[int] = set() + for model_name, tp_size in product(args.models, args.tp_sizes): + hidden_sizes = hidden_sizes.union(hidden_sizes_from_model(model_name, tp_size)) + + print(f"Model bench :\n Hidden Sizes {hidden_sizes} LoRA Ranks {args.lora_ranks}") + + # Get all benchmarking contexts + bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=hidden_sizes, lora_ranks=args.lora_ranks, args=args + ) + + run(args, bench_contexts) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "torch.float16": + return torch.float16 + if dt == "torch.bfloat16": + return torch.bfloat16 + raise ValueError("unsupported dtype") + + def get_bool(s: str) -> bool: + return s.lower() in ["true", "1"] + + def add_common_command_args(p: argparse.ArgumentParser): + p.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['torch.float16', 'torch.bfloat16']", + ) + + p.add_argument( + "--arg-pool-size", + type=int, + default=32, + help="Run profiles with a pool of input/output/meta tensors instead" + "of simply reusing the same tensors for all runs. A bigger arg-pool" + "mitigates hardware caching effects during benchmarking.", + ) + + p.add_argument( + "--cuda-graph-nops", + type=int, + help=( + "when set profiling is done using cudagraph, " + "with the given number of operations in a graph." + "Note that the measurement returned is the time " + "taken for N consecutive executions of the benchmarking " + "functions, where N is the value of this argument." + ), + ) + p.add_argument("--num-loras", nargs="+", type=int, default=DEFAULT_NUM_LORAS) + p.add_argument( + "--num-active-loras", + type=int, + default=None, + help="Active LoRAs. When None, all LoRAs are active", + ) + p.add_argument( + "--sort-by-lora-id", + nargs="+", + type=get_bool, + default=DEFAULT_SORT_BY_LORA_IDS, + ) + p.add_argument( + "--op-types", nargs="+", type=OpType.from_str, default=list(OpType) + ) + p.add_argument( + "--seq-lengths", nargs="+", type=int, default=DEFAULT_SEQ_LENGTHS + ) + p.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + p.add_argument( + "--expand-fn-add-inputs", + nargs="+", + type=get_bool, + default=DEFAULT_EXPAND_FN_ADD_INPUTS, + ) + p.add_argument( + "-o", + "--output-directory", + type=str, + help=( + "Output directory to store a the list of benchmarking" + "TMeasurement objects as a pickle file" + ), + ) + + p.add_argument( + "--test-correctness", + action="store_true", + help=( + "When enabled, the benchmarking functions are tested" + "for correctness before the actual benchmarking" + ), + ) + + parser = FlexibleArgumentParser( + description=f""" +Benchmark LoRA kernels: + {use_cuda_graph_recommendation()} + + list_bench example: + python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 + + model_bench example: + python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 + + range_bench example: + python3 benchmarks/kernels/benchmark_lora.py range_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + + subparsers = parser.add_subparsers(dest="cmd", required=True) + + list_parser = subparsers.add_parser("list_bench") + list_parser.add_argument( + "--hidden-sizes", nargs="+", type=int, default=DEFAULT_HIDDEN_SIZES + ) + list_parser.add_argument( + "--lora-ranks", nargs="+", type=int, default=DEFAULT_LORA_RANKS + ) + add_common_command_args(list_parser) + list_parser.set_defaults(func=run_list_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--hidden-sizes-start", type=int, required=True) + range_parser.add_argument("--hidden-sizes-end", type=int, required=True) + range_parser.add_argument("--hidden-sizes-increment", type=int, required=True) + range_parser.add_argument("--lora-ranks-start", type=int, required=True) + range_parser.add_argument("--lora-ranks-end", type=int, required=True) + range_parser.add_argument("--lora-ranks-increment", type=int, required=True) + add_common_command_args(range_parser) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--lora-ranks", nargs="+", type=int, default=DEFAULT_LORA_RANKS + ) + add_common_command_args(model_parser) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_machete.py b/offline_benchmark_test/benchmarks/kernels/benchmark_machete.py new file mode 100644 index 0000000000000000000000000000000000000000..f73d0511e01fc3851847278f8604a2baed8c8f33 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_machete.py @@ -0,0 +1,732 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import itertools +import math +import os +import pickle as pkl +import time +from collections.abc import Iterable +from dataclasses import dataclass +from itertools import product +from typing import Callable, Optional + +import pandas as pd +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from weight_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.marlin_utils import ( + GPTQ_MARLIN_MAX_PARALLEL, + GPTQ_MARLIN_MIN_THREAD_N, + marlin_permute_scales, + marlin_zero_points, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_test import ( + MarlinWorkspace, +) +from vllm.model_executor.layers.quantization.utils.quant_utils import ( + pack_rows, + quantize_weights, +) +from vllm.scalar_type import ScalarType, scalar_types +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = ["meta-llama/Llama-3-8b", "meta-llama/Llama-2-70b-hf"] +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024] +DEFAULT_TP_SIZES = [1] + +NVTX_PROFILE = os.environ.get("NVTX_PROFILE", False) + +if NVTX_PROFILE: + import nvtx + + +def terse_type_name(dt): + return { + torch.bfloat16: "bf16", + torch.float16: "fp16", + torch.int8: "int8", + torch.float8_e4m3fn: "fp8", + torch.float: "float", + torch.int: "int", + }[dt] + + +@dataclass +class BenchmarkTensors: + w_ref: torch.Tensor + a: torch.Tensor + + w_q: torch.Tensor + group_size: Optional[int] + wtype: ScalarType + w_g_s: torch.Tensor + w_g_zp: Optional[torch.Tensor] + w_ch_s: Optional[torch.Tensor] + w_tok_s: Optional[torch.Tensor] + + +@dataclass +class TypeConfig: + act_type: torch.dtype + weight_type: ScalarType + output_type: Optional[torch.dtype] + group_scale_type: Optional[torch.dtype] + group_zero_type: Optional[torch.dtype] + channel_scale_type: Optional[torch.dtype] + token_scale_type: Optional[torch.dtype] + + +def rand_data(shape, dtype=torch.float16, scale=1): + if dtype.is_floating_point: + return (scale * torch.rand(shape, device="cuda") - 0.3).to(dtype) + else: + return torch.randint(-15, 15, shape, dtype=dtype, device="cuda") + + +def quantize_and_pack( + atype: torch.dtype, + w: torch.Tensor, + wtype: ScalarType, + stype: Optional[torch.dtype], + group_size: Optional[int], + zero_points: bool = False, +): + assert wtype.is_integer(), "TODO: support floating point weights" + + w_ref, w_q, w_s, w_zp = quantize_weights( + w, + wtype, + group_size=group_size, + zero_points=zero_points, + # to match how the kernel applies zps + ref_zero_points_after_scales=True, + ) + + w_q = pack_rows(w_q, wtype.size_bits, *w_q.shape) + return w_ref, w_q, w_s, w_zp + + +def create_bench_tensors( + shape: tuple[int, int, int], types: TypeConfig, group_size: Optional[int] +) -> list[BenchmarkTensors]: + m, n, k = shape + + # we want to make sure that weights don't fit into L2 cache between runs so + # we construct enough weights to exceed L2 cache, which is 50mb on a H100 + # so we target total weight size > 2*50mb + num_weights = math.ceil( + 2 * 50 * 1024**2 * 8 / (k * n * types.weight_type.size_bits) + ) + + a = rand_data((m, k), types.act_type, scale=5) + + benchmark_tensors: list[BenchmarkTensors] = [] + for _ in range(num_weights): + w = rand_data((k, n), types.act_type, scale=5) + + if types.group_scale_type is not None: + w = w.to(types.group_scale_type) + if w.dtype.itemsize == 1: + w = w.to(torch.float16) + + w_ref, w_q_packed, w_s, w_zp = quantize_and_pack( + a.dtype, + w, + types.weight_type, + types.group_scale_type, + group_size, + types.group_zero_type is not None, + ) + + if not a.dtype.is_floating_point: + aiinfo = torch.iinfo(a.dtype) + w_ref = w_ref.round().clamp(aiinfo.min, aiinfo.max) + + w_ref = w_ref.to(torch.float32) + + w_ch_s = ( + None + if types.channel_scale_type is None + else rand_data((n,), types.channel_scale_type) + ) + w_tok_s = ( + None + if types.token_scale_type is None + else rand_data((m,), types.token_scale_type) + ) + + benchmark_tensors.append( + BenchmarkTensors( + w_ref=w_ref, + a=a, + w_q=w_q_packed, + wtype=types.weight_type, + w_g_s=w_s, + w_g_zp=w_zp, + group_size=group_size, + w_ch_s=w_ch_s, + w_tok_s=w_tok_s, + ) + ) + + return benchmark_tensors + + +def torch_matmul_f16_create_bench_fn(bt: BenchmarkTensors) -> Callable: + a = bt.a + w = bt.w_ref.to(bt.a.dtype) # use float reference tensor + if a.dtype not in [torch.float16, torch.bfloat16]: + a = a.to(torch.float16) + w = w.to(torch.float16) + return lambda: torch.matmul(a, w) + + +def cutlass_scaled_mm_create_bench_fn(bt: BenchmarkTensors) -> Callable: + if bt.w_ch_s is not None and bt.w_tok_s is not None: + scale_a = bt.w_tok_s.to(torch.float32) + scale_b = bt.w_ch_s.to(torch.float32) + else: + scale_a = torch.tensor(1.0, dtype=torch.float32, device=bt.a.device) + scale_b = torch.tensor(1.0, dtype=torch.float32, device=bt.a.device) + w_col_major = bt.w_ref.to(bt.a.dtype).t().contiguous().t() + return lambda: ops.cutlass_scaled_mm( + bt.a, w_col_major, scale_a, scale_b, out_dtype=torch.float16 + ) + + +def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable: + device = bt.a.device + + workspace = MarlinWorkspace( + bt.w_ref.shape[1], GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL + ) + + if bt.w_g_zp is None: + w_zp = torch.empty(0, dtype=torch.int, device=device) + else: + w_zp = marlin_zero_points( + bt.w_g_zp, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits + ) + + if bt.group_size is None: + w_s = torch.tensor([], device="cuda", dtype=torch.half) + else: + w_s = marlin_permute_scales( + bt.w_g_s, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.group_size + ) + + sort_indices = torch.empty(0, dtype=torch.int, device=device) + g_idx = torch.empty(0, dtype=torch.int, device=device) + w_q = ops.gptq_marlin_repack( + bt.w_q, sort_indices, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits + ) + + if bt.a.dtype.is_floating_point: + assert bt.w_ch_s is None + assert bt.w_tok_s is None + assert bt.group_size is not None + + fn = lambda: ops.gptq_marlin_gemm( + a=bt.a, + c=None, + b_q_weight=w_q, + b_scales=w_s, + global_scale=None, + b_zeros=w_zp, + g_idx=g_idx, + perm=sort_indices, + workspace=workspace.scratch, + b_q_type=bt.wtype, + size_m=bt.a.shape[0], + size_n=bt.w_ref.shape[1], + size_k=bt.w_ref.shape[0], + is_k_full=True, + is_zp_float=False, + ) + else: + assert bt.a.dtype == torch.int8 + assert bt.wtype == scalar_types.uint4b8 + + if bt.w_ch_s is not None: + s_ch = bt.w_ch_s.to(torch.float32) + else: + s_ch = torch.ones(bt.w_ref.shape[1], dtype=torch.float32, device=device) + + if bt.w_tok_s is not None: + s_tok = bt.w_tok_s.to(torch.float32) + else: + s_tok = torch.ones(bt.a.shape[0], dtype=torch.float32, device=device) + + fn = lambda: ops.marlin_qqq_gemm( + a=bt.a, + b_q_weight=w_q, + s_group=w_s, + s_tok=s_tok, + s_ch=s_ch, + workspace=workspace.scratch, + size_m=bt.a.shape[0], + size_n=bt.w_ref.shape[1], + size_k=bt.w_ref.shape[0], + ) + + return fn + + +def machete_create_bench_fn( + bt: BenchmarkTensors, out_type=torch.dtype, schedule=None +) -> Callable: + w_q = bt.w_q.t().contiguous().t() # make col major + w_q = ops.machete_prepack_B( + w_q, bt.a.dtype, bt.wtype, None if bt.w_g_s is None else bt.w_g_s.dtype + ) + + w_g_zp = bt.w_g_zp + if w_g_zp is not None: + w_g_zp = -1 * bt.w_g_s * (w_g_zp.to(bt.w_g_s.dtype)) + + return lambda: ops.machete_mm( + a=bt.a, + b_q=w_q, + b_type=bt.wtype, + b_group_scales=bt.w_g_s, + b_group_zeros=w_g_zp, + b_group_size=bt.group_size, + b_channel_scales=bt.w_ch_s, + a_token_scales=bt.w_tok_s, + out_type=out_type, + schedule=schedule, + ) + + +# impl + +# bench + + +def bench_fns(label: str, sub_label: str, description: str, fns: list[Callable]): + min_run_time = 1 if not NVTX_PROFILE else 0.1 + res = TBenchmark.Timer( + stmt=""" + for fn in fns: + fn() + """, + globals={"fns": fns}, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + if NVTX_PROFILE: + with ( + nvtx.annotate("mm-bench"), + nvtx.annotate(f"{label}|{sub_label}|{description}"), + ): + fns[0]() + + return res + + +_SWEEP_SCHEDULES_RESULTS: Optional[pd.DataFrame] = None +_SWEEP_SCHEDULES_RESULTS_CSV: Optional[str] = None + + +def bench( + types: TypeConfig, + group_size: int, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + sweep_schedules: bool = True, +) -> list[TMeasurement]: + benchmark_tensors = create_bench_tensors((m, n, k), types, group_size) + sub_label += f", L={len(benchmark_tensors)}" + + name_type_string = f"W{types.weight_type}" + f"-A{terse_type_name(types.act_type)}" + if types.group_scale_type is not None: + name_type_string += f"-GS{terse_type_name(types.group_scale_type)}" + if types.group_zero_type is not None: + name_type_string += f"-GZ{terse_type_name(types.group_zero_type)}" + if group_size is not None: + name_type_string += f"-G{group_size}" + if types.channel_scale_type is not None: + name_type_string += f"-CS{terse_type_name(types.channel_scale_type)}" + if types.token_scale_type is not None: + name_type_string += f"-TS{terse_type_name(types.token_scale_type)}" + + timers = [] + # pytorch impl + timers.append( + bench_fns( + label, + sub_label, + "torch.matmul (fp16)", + [torch_matmul_f16_create_bench_fn(bt) for bt in benchmark_tensors], + ) + ) + + if types.act_type == torch.int8 or types.act_type == torch.float8_e4m3fn: + timers.append( + bench_fns( + label, + sub_label, + f"cutlass_scaled_mm ({terse_type_name(types.act_type)})", + [cutlass_scaled_mm_create_bench_fn(bt) for bt in benchmark_tensors], + ) + ) + + if types.act_type != torch.float8_e4m3fn: + timers.append( + bench_fns( + label, + sub_label, + f"marlin ({name_type_string})", + [marlin_create_bench_fn(bt) for bt in benchmark_tensors], + ) + ) + + # machete + timers.append( + bench_fns( + label, + sub_label, + f"machete ({name_type_string})", + [ + machete_create_bench_fn(bt, out_type=types.output_type) + for bt in benchmark_tensors + ], + ) + ) + + if sweep_schedules: + global _SWEEP_SCHEDULES_RESULTS + + print("Finding best schedule for machete") + best = None + best_schedule = None + schedules = ops.machete_supported_schedules( + a_type=types.act_type, + b_type=types.weight_type, + group_scales_type=types.group_scale_type, + group_zeros_type=types.group_zero_type, + token_scales_type=types.token_scale_type, + channel_scales_type=types.channel_scale_type, + out_type=types.output_type, + ) + + if schedules is None or len(schedules) == 0: + raise ValueError("No schedules found to sweep") + + for schedule in reversed(schedules): + schedule_M = int(schedule.split("_")[0].split("x")[1]) + + # Prune known bad schedules + if schedule_M >= 2 * max(m, 16) or schedule_M < m // 4: + continue + + res = bench_fns( + label, + sub_label, + "machete_best", + [ + machete_create_bench_fn( + bt, out_type=types.output_type, schedule=schedule + ) + for bt in benchmark_tensors + ], + ) + + results_row = { + "M": m, + "K": k, + "N": n, + "group_size": group_size, + "schedule": schedule, + "median": res.median, + } + if _SWEEP_SCHEDULES_RESULTS is None: + _SWEEP_SCHEDULES_RESULTS = pd.DataFrame(columns=results_row.keys()) + _SWEEP_SCHEDULES_RESULTS.loc[len(_SWEEP_SCHEDULES_RESULTS)] = results_row + + print(f" {res.median:5.5} ", schedule) + if not best or res.median < best.median: + best = res + best_schedule = schedule + print("Best schedule:", best_schedule) + timers.append(best) + + return timers + + +# runner +def print_timers(timers: list[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def run(args, MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]: + types = TypeConfig( + act_type=args.act_type, + weight_type=scalar_types.uint4b8 + if args.group_zero_type is None + else scalar_types.uint4, + output_type=args.out_type, + group_scale_type=args.group_scale_type, + group_zero_type=args.group_zero_type, + channel_scale_type=args.channel_scale_type, + token_scale_type=args.token_scale_type, + ) + + results: list[TMeasurement] = [] + for m, k, n in MKNs: + timers = bench( + types, + args.group_size, + m, + k, + n, + f"{args.act_type}-gemm", + f"MKN=({m}x{k}x{n})", + sweep_schedules=args.sweep_schedules, + ) + print_timers(timers) + results.extend(timers) + + return results + + +# output makers +def make_output( + data: list[TMeasurement], + MKNs: Iterable[tuple[int, int, int]], + base_description: str, + timestamp=None, +): + print(f"== All Results {base_description} ====") + print_timers(data) + + # pickle all the results + timestamp = int(time.time()) if timestamp is None else timestamp + with open(f"{base_description}-{timestamp}.pkl", "wb") as f: + pkl.dump(data, f) + + +# argparse runners + + +def run_square_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment)) + MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes)) + data = run(args.dtype, args.sweep_schedules, MKNs) + + make_output(data, MKNs, f"square_bench-{args.dtype}") + + +def run_range_bench(args): + m_start, k_start, n_start = (int(x) for x in args.dim_start.split(",")) + m_end, k_end, n_end = (int(x) for x in args.dim_end.split(",")) + m_increment, k_increment, n_increment = ( + int(x) for x in args.dim_increment.split(",") + ) + Ms = list(range(m_start, m_end + 1, m_increment)) + Ks = list(range(k_start, k_end + 1, k_increment)) + Ns = list(range(n_start, n_end + 1, n_increment)) + MKNs = list(product(Ms, Ks, Ns)) + + data = run(args.dtype, args.sweep_schedules, MKNs) + + make_output(data, MKNs, f"range_bench-{args.dtype}") + + +def run_model_bench(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]: + KNs = [] + for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]): + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + KNs.append(KN) + return KNs + + model_bench_data = [] + models_tps = list(itertools.product(args.models, args.tp_sizes)) + for model, tp_size in models_tps: + Ms = args.batch_sizes + KNs = model_shapes(model, tp_size) + MKNs = [] + for m in Ms: + for k, n in KNs: + MKNs.append((m, k, n)) + + data = run(args, MKNs) + model_bench_data.append(data) + + type_string = f"{args.act_type}" + + # Print all results + for data, model_tp in zip(model_bench_data, models_tps): + model, tp_size = model_tp + print(f"== Results {type_string} {model}-TP{tp_size} ====") + print_timers(data) + + timestr = time.strftime("%Y%m%d-%H%M%S") + + all_results = [] + for d in model_bench_data: + all_results.extend(d) + + # pickle all data + with open(f"model_bench-{type_string}-{timestr}.pkl", "wb") as f: + args_dict = vars(args) + args_dict.pop("func") + pkl.dump( + { + "args": args_dict, + "results": all_results, + }, + f, + ) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + return { + "bfloat16": torch.bfloat16, + "float16": torch.float16, + "int8": torch.int8, + "float8_e4m3fn": torch.float8_e4m3fn, + "int": torch.int, + "float": torch.float, + }[dt] + + class ToTorchDtype(argparse.Action): + def __call__(self, parser, namespace, values, option_string=None): + setattr(namespace, self.dest, to_torch_dtype(values)) + + parser = FlexibleArgumentParser( + description=""" +Benchmark Machete GEMM. + + To run square GEMMs: + python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 square_bench --dim-start 128 --dim-end 512 --dim-increment 64 + + To run constant N and K and sweep M: + python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384 + + To run dimensions from a model: + python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1 + + Output: + - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs. + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + parser.add_argument( + "--act-type", + action=ToTorchDtype, + required=True, + choices=["bfloat16", "float16", "int8", "float8_e4m3fn"], + ) + parser.add_argument( + "--group-scale-type", + action=ToTorchDtype, + choices=["bfloat16", "float16"], + ) + parser.add_argument( + "--group-zero-type", + type=to_torch_dtype, + choices=["bfloat16", "float16"], + ) + parser.add_argument( + "--channel-scale-type", + action=ToTorchDtype, + choices=["float"], + ) + parser.add_argument( + "--token-scale-type", + action=ToTorchDtype, + choices=["float"], + ) + parser.add_argument( + "--out-type", + action=ToTorchDtype, + choices=["bfloat16", "float16"], + ) + parser.add_argument( + "--group-size", + type=int, + help="Available options are ['None', '-1', '128'], default=128", + default=128, + ) + parser.add_argument( + "--sweep-schedules", + action="store_true", + help="Run a sweep over all supported schedules", + ) + parser.add_argument( + "--sweep-csv-out", + help="CSV to store sweep results", + default="sch_sweep_results.csv", + ) + subparsers = parser.add_subparsers(dest="cmd", required=True) + + square_parser = subparsers.add_parser("square_bench") + square_parser.add_argument("--dim-start", type=int, required=True) + square_parser.add_argument("--dim-end", type=int, required=True) + square_parser.add_argument("--dim-increment", type=int, required=True) + square_parser.set_defaults(func=run_square_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument( + "--dim-start", + type=str, + required=True, + help="Start value for M,K,N as common separated list", + ) + range_parser.add_argument( + "--dim-end", + type=str, + required=True, + help="End value (inclusive) for M,K,N as common separated list", + ) + range_parser.add_argument( + "--dim-increment", + type=str, + required=True, + help="Increment value for M,K,N as common separated list", + ) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + + _SWEEP_SCHEDULES_RESULTS_CSV = args.sweep_csv_out + args.func(args) + + if _SWEEP_SCHEDULES_RESULTS is not None: + _SWEEP_SCHEDULES_RESULTS.to_csv(_SWEEP_SCHEDULES_RESULTS_CSV) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_marlin.py b/offline_benchmark_test/benchmarks/kernels/benchmark_marlin.py new file mode 100644 index 0000000000000000000000000000000000000000..34cc45e94d76d930013c9d60ca8c5f78dd0b7254 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_marlin.py @@ -0,0 +1,413 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import torch +import torch.utils.benchmark as benchmark +from benchmark_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.gptq_marlin_24 import ( + GPTQ_MARLIN_24_MAX_PARALLEL, + GPTQ_MARLIN_24_MIN_THREAD_N, + GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, + GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES, +) +from vllm.model_executor.layers.quantization.utils.allspark_utils import ( + ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD, + ALLSPARK_SUPPORTED_QUANT_TYPES, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils import ( + GPTQ_MARLIN_MAX_PARALLEL, + GPTQ_MARLIN_MIN_THREAD_N, + MARLIN_SUPPORTED_GROUP_SIZES, + query_marlin_supported_quant_types, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import ( + FP4_MARLIN_SUPPORTED_GROUP_SIZES, + rand_marlin_weight_fp4_like, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import ( + marlin_quant_fp8_torch, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_test import ( + MarlinWorkspace, + awq_marlin_quantize, + marlin_quantize, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import ( + marlin_24_quantize, +) +from vllm.model_executor.layers.quantization.utils.quant_utils import ( + gptq_pack, + gptq_quantize_weights, + quantize_weights, + sort_weights, +) +from vllm.scalar_type import ScalarType, scalar_types +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"] +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192] + +ACT_ORDER_OPTS = [False, True] +K_FULL_OPTS = [False, True] + + +def bench_run( + results: list[benchmark.Measurement], + model: str, + act_order: bool, + is_k_full: bool, + quant_type: ScalarType, + group_size: int, + size_m: int, + size_k: int, + size_n: int, +): + label = "Quant Matmul" + sub_label = "{}, act={} k_full={}, q={}, g={}, MKN=({}x{}x{})".format( + model, act_order, is_k_full, str(quant_type), group_size, size_m, size_k, size_n + ) + print(f"Testing: {sub_label}") + + a = torch.randn(size_m, size_k).to(torch.half).cuda() + b = torch.rand(size_k, size_n).to(torch.half).cuda() + has_zp = quant_type in [scalar_types.uint4, scalar_types.uint8] + if act_order and (group_size == -1 or group_size == size_k or has_zp): + return + if size_k % group_size != 0: + return + + marlin_24_supported = ( + quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES + and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES + ) + repack_supported = ( + quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES + and group_size in MARLIN_SUPPORTED_GROUP_SIZES + ) + allspark_supported = ( + quant_type in ALLSPARK_SUPPORTED_QUANT_TYPES + and group_size == -1 + and not act_order + and is_k_full + ) + + def gen_marlin_params(): + # Marlin quant + marlin_g_idx = marlin_sort_indices = marlin_zp = marlin_s2 = None + if quant_type == scalar_types.float4_e2m1f: + if group_size != 16 or act_order: + return + marlin_w_ref, marlin_q_w, marlin_s, marlin_s2 = rand_marlin_weight_fp4_like( + b.T, group_size + ) + elif quant_type == scalar_types.float8_e4m3fn: + if group_size not in [-1, 128] or act_order: + return + marlin_w_ref, marlin_q_w, marlin_s = marlin_quant_fp8_torch(b.T, group_size) + elif group_size == 16: + return + elif has_zp: + marlin_w_ref, marlin_q_w, marlin_s, marlin_zp = awq_marlin_quantize( + b, quant_type, group_size + ) + else: + marlin_w_ref, marlin_q_w, marlin_s, marlin_g_idx, marlin_sort_indices, _ = ( + marlin_quantize(b, quant_type, group_size, act_order) + ) + return ( + marlin_w_ref, + marlin_q_w, + marlin_s, + marlin_s2, + marlin_zp, + marlin_g_idx, + marlin_sort_indices, + ) + + def gen_marlin_24_params(): + marlin_24_w_ref = marlin_24_q_w_comp = marlin_24_meta = marlin_24_s = None + if marlin_24_supported: + (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) = ( + marlin_24_quantize(b, quant_type, group_size) + ) + return (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) + + def gen_repack_params(): + q_w_gptq = None + repack_sort_indices = None + if repack_supported: + (w_ref, q_w, s, g_idx, rand_perm) = gptq_quantize_weights( + b, quant_type, group_size, act_order + ) + q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n) + + # For act_order, sort the "weights" and "g_idx" + # so that group ids are increasing + repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device) + if act_order: + (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx) + return q_w_gptq, repack_sort_indices + + def gen_allspark_params(): + qw_reorder = s_reorder = zp_reorder = sm_count = sm_version = ( + CUBLAS_M_THRESHOLD + ) = None + nonlocal allspark_supported + if allspark_supported: + properties = torch.cuda.get_device_properties(b.device.index) + sm_count = properties.multi_processor_count + sm_version = properties.major * 10 + properties.minor + + supported_arch = sm_version >= 80 and sm_version < 90 + allspark_supported = allspark_supported and supported_arch + if supported_arch: + w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size, has_zp) + qw = qw.to(torch.uint8) + + qw_reorder, s_reorder, zp_reorder = ops.allspark_repack_weight( + qw, s, zp, has_zp + ) + CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD + return ( + qw_reorder, + s_reorder, + zp_reorder, + sm_count, + sm_version, + CUBLAS_M_THRESHOLD, + ) + + ( + marlin_w_ref, + marlin_q_w, + marlin_s, + marlin_s2, + marlin_zp, + marlin_g_idx, + marlin_sort_indices, + ) = gen_marlin_params() + marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s = ( + gen_marlin_24_params() + ) + q_w_gptq, repack_sort_indices = gen_repack_params() + qw_reorder, s_reorder, zp_reorder, sm_count, sm_version, CUBLAS_M_THRESHOLD = ( + gen_allspark_params() + ) + + # Prepare + marlin_workspace = MarlinWorkspace( + size_n, GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL + ) + marlin_24_workspace = MarlinWorkspace( + size_n, GPTQ_MARLIN_24_MIN_THREAD_N, GPTQ_MARLIN_24_MAX_PARALLEL + ) + + globals = { + # Gen params + "quant_type": quant_type, + "group_size": group_size, + "size_m": size_m, + "size_n": size_n, + "size_k": size_k, + "a": a, + # Marlin params + "marlin_w_ref": marlin_w_ref, + "marlin_q_w": marlin_q_w, + "marlin_s": marlin_s, + "marlin_s2": marlin_s2, + "marlin_zp": marlin_zp, + "marlin_g_idx": marlin_g_idx, + "marlin_sort_indices": marlin_sort_indices, + "marlin_workspace": marlin_workspace, + "is_k_full": is_k_full, + # Marlin_24 params + "marlin_24_w_ref": marlin_24_w_ref, + "marlin_24_q_w_comp": marlin_24_q_w_comp, + "marlin_24_meta": marlin_24_meta, + "marlin_24_s": marlin_24_s, + "marlin_24_workspace": marlin_24_workspace, + # GPTQ params + "q_w_gptq": q_w_gptq, + "repack_sort_indices": repack_sort_indices, + # AllSpark W8A16 params + "qw_reorder": qw_reorder, + "s_reorder": s_reorder, + "zp_reorder": zp_reorder, + "sm_count": sm_count, + "sm_version": sm_version, + "CUBLAS_M_THRESHOLD": CUBLAS_M_THRESHOLD, + # Kernels + "gptq_marlin_gemm": ops.gptq_marlin_gemm, + "gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm, + "gptq_marlin_repack": ops.gptq_marlin_repack, + "allspark_w8a16_gemm": ops.allspark_w8a16_gemm, + } + + min_run_time = 1 + + # Warmup pytorch + for _ in range(5): + torch.matmul(a, marlin_w_ref) + + results.append( + benchmark.Timer( + stmt="torch.matmul(a, marlin_w_ref)", + globals=globals, + label=label, + sub_label=sub_label, + description="pytorch_gemm", + ).blocked_autorange(min_run_time=min_run_time) + ) + + results.append( + benchmark.Timer( + stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False, False)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_gemm", + ).blocked_autorange(min_run_time=min_run_time) + ) + + results.append( + benchmark.Timer( + stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True, False)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_gemm_fp32", + ).blocked_autorange(min_run_time=min_run_time) + ) + + if marlin_24_supported: + results.append( + benchmark.Timer( + stmt="output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, quant_type, size_m, size_n, size_k)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_24_gemm", + ).blocked_autorange(min_run_time=min_run_time) + ) + + if repack_supported: + results.append( + benchmark.Timer( + stmt="q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_repack", + ).blocked_autorange(min_run_time=min_run_time) + ) + + if allspark_supported: + results.append( + benchmark.Timer( + stmt="output = allspark_w8a16_gemm(a, qw_reorder, s_reorder, zp_reorder, size_n, group_size, sm_count, sm_version, CUBLAS_M_THRESHOLD, False, True)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="allspark_w8a16_gemm_fp32", + ).blocked_autorange(min_run_time=min_run_time) + ) + + +def main(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + results: list[benchmark.Measurement] = [] + + for model in args.models: + for layer in WEIGHT_SHAPES[model]: + size_k = layer[0] + size_n = layer[1] + + if len(args.limit_k) > 0 and size_k not in args.limit_k: + continue + + if len(args.limit_n) > 0 and size_n not in args.limit_n: + continue + + for act_order in ACT_ORDER_OPTS: + if ( + len(args.limit_act_order) > 0 + and act_order not in args.limit_act_order + ): + continue + + for is_k_full in K_FULL_OPTS: + if ( + len(args.limit_k_full) > 0 + and is_k_full not in args.limit_k_full + ): + continue + + for quant_type in query_marlin_supported_quant_types(): + if ( + len(args.limit_num_bits) > 0 + and quant_type.size_bits not in args.limit_num_bits + ): + continue + + for group_size in ( + MARLIN_SUPPORTED_GROUP_SIZES + + FP4_MARLIN_SUPPORTED_GROUP_SIZES + ): + if ( + len(args.limit_group_size) > 0 + and group_size not in args.limit_group_size + ): + continue + + # For act_order, the group_size must be less than + # size_k + if act_order and (group_size == size_k or group_size == -1): + continue + + for size_m in args.batch_sizes: + bench_run( + results, + model, + act_order, + is_k_full, + quant_type, + group_size, + size_m, + size_k, + size_n, + ) + + compare = benchmark.Compare(results) + compare.print() + + +# For quick benchmarking use: +# python benchmark_marlin.py --batch-sizes 1 16 32 --limit-k 4096 --limit-n 4096 --limit-group-size 128 --limit-num-bits 4 --limit-act-order 0 --limit-k-full 1 # noqa E501 +# +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark Marlin across specified models/shapes/batches" + ) + parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + parser.add_argument("--limit-k", nargs="+", type=int, default=[]) + parser.add_argument("--limit-n", nargs="+", type=int, default=[]) + parser.add_argument("--limit-group-size", nargs="+", type=int, default=[]) + parser.add_argument("--limit-num-bits", nargs="+", type=int, default=[]) + parser.add_argument("--limit-act-order", nargs="+", type=int, default=[]) + parser.add_argument("--limit-k-full", nargs="+", type=int, default=[]) + + args = parser.parse_args() + main(args) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_moe.py b/offline_benchmark_test/benchmarks/kernels/benchmark_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..07af58d81c68331597f1e63c0c62dc814b41025c --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_moe.py @@ -0,0 +1,737 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import json +import time +from contextlib import nullcontext +from datetime import datetime +from itertools import product +from typing import Any, TypedDict + +import ray +import torch +from ray.experimental.tqdm_ray import tqdm + +from vllm.model_executor.layers.fused_moe.fused_moe import * +from vllm.platforms import current_platform +from vllm.transformers_utils.config import get_config +from vllm.triton_utils import triton +from vllm.utils import FlexibleArgumentParser + +FP8_DTYPE = current_platform.fp8_dtype() + + +class BenchmarkConfig(TypedDict): + BLOCK_SIZE_M: int + BLOCK_SIZE_N: int + BLOCK_SIZE_K: int + GROUP_SIZE_M: int + num_warps: int + num_stages: int + + +def benchmark_config( + config: BenchmarkConfig, + num_tokens: int, + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + num_iters: int = 100, + block_quant_shape: list[int] = None, + use_deep_gemm: bool = False, +) -> float: + init_dtype = torch.float16 if use_fp8_w8a8 else dtype + x = torch.randn(num_tokens, hidden_size, dtype=dtype) + if use_int8_w8a16: + w1 = torch.randint( + -127, + 127, + ( + num_experts, + shard_intermediate_size, + hidden_size, + ), + dtype=torch.int8, + ) + w2 = torch.randint( + -127, + 127, + ( + num_experts, + hidden_size, + shard_intermediate_size // 2, + ), + dtype=torch.int8, + ) + else: + w1 = torch.randn( + num_experts, shard_intermediate_size, hidden_size, dtype=init_dtype + ) + w2 = torch.randn( + num_experts, hidden_size, shard_intermediate_size // 2, dtype=init_dtype + ) + gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32) + + w1_scale = None + w2_scale = None + a1_scale = None + a2_scale = None + if use_int8_w8a16: + w1_scale = torch.randn( + (num_experts, 2 * shard_intermediate_size), dtype=torch.float32 + ) + w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32) + if use_fp8_w8a8: + if block_quant_shape: + block_n, block_k = block_quant_shape[0], block_quant_shape[1] + E = num_experts + N = shard_intermediate_size // 2 + K = hidden_size + factor_for_scale = 1e-2 + n_tiles_w1 = (2 * N + block_n - 1) // block_n + n_tiles_w2 = (K + block_n - 1) // block_n + k_tiles_w1 = (K + block_k - 1) // block_k + k_tiles_w2 = (N + block_k - 1) // block_k + w1_scale = ( + torch.rand((E, n_tiles_w1, k_tiles_w1), dtype=torch.float32) + * factor_for_scale + ) + w2_scale = ( + torch.rand((E, n_tiles_w2, k_tiles_w2), dtype=torch.float32) + * factor_for_scale + ) + else: + w1_scale = torch.randn(num_experts, dtype=torch.float32) + w2_scale = torch.randn(num_experts, dtype=torch.float32) + + a1_scale = torch.randn(1, dtype=torch.float32) + a2_scale = torch.randn(1, dtype=torch.float32) + + w1 = w1.to(FP8_DTYPE) + w2 = w2.to(FP8_DTYPE) + + input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32) + + def prepare(i: int): + input_gating.copy_(gating_output[i]) + + def run(): + from vllm.model_executor.layers.fused_moe import override_config + + with override_config(config): + if use_deep_gemm: + topk_weights, topk_ids, token_expert_indices = fused_topk( + x, input_gating, topk, False + ) + return fused_experts( + x, + w1, + w2, + topk_weights, + topk_ids, + inplace=True, + use_fp8_w8a8=use_fp8_w8a8, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_quant_shape, + allow_deep_gemm=True, + ) + else: + fused_moe( + x, + w1, + w2, + input_gating, + topk, + renormalize=True, + inplace=True, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_quant_shape, + ) + + # JIT compilation & warmup + run() + torch.cuda.synchronize() + + # Capture 10 invocations with CUDA graph + graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(graph): + for _ in range(10): + run() + torch.cuda.synchronize() + + # Warmup + for _ in range(5): + graph.replay() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + prepare(i) + torch.cuda.synchronize() + + start_event.record() + graph.replay() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + graph.reset() + return avg + + +def get_rocm_tuning_space(use_fp16): + block_mn_range = [16, 32, 64, 128, 256] + block_k_range = [16, 32, 64, 128, 256] + if not use_fp16: + block_k_range.remove(16) # BLOCK_K=16 not supported for fp8 + num_warps_range = [1, 2, 4, 8] + group_m_range = [1, 4, 8, 16, 32] + num_stage_range = [2] + waves_per_eu_range = [0] + matrix_instr_nonkdim_range = [16, 32] if use_fp16 else [] + kpack_range = [1, 2] if use_fp16 else [] + + param_ranges = { + "BLOCK_SIZE_M": block_mn_range, + "BLOCK_SIZE_N": block_mn_range, + "BLOCK_SIZE_K": block_k_range, + "GROUP_SIZE_M": group_m_range, + "num_warps": num_warps_range, + "num_stages": num_stage_range, + "waves_per_eu": waves_per_eu_range, + } + if use_fp16: + param_ranges["matrix_instr_nonkdim"] = matrix_instr_nonkdim_range + param_ranges["kpack"] = kpack_range + + return param_ranges + + +def get_configs_compute_bound(use_fp16, block_quant_shape) -> list[dict[str, int]]: + configs: list[BenchmarkConfig] = [] + + if current_platform.is_rocm(): + param_ranges = get_rocm_tuning_space(use_fp16) + else: + # Reduced search space for faster tuning. + # TODO(woosuk): Increase the search space and use a performance model to + # prune the search space. + block_m_range = [16, 32, 64, 128, 256] + block_n_range = [32, 64, 128, 256] + block_k_range = [64, 128, 256] + num_warps_range = [4, 8] + group_m_range = [1, 16, 32, 64] + num_stage_range = [2, 3, 4, 5] + + param_ranges = { + "BLOCK_SIZE_M": block_m_range, + "BLOCK_SIZE_N": block_n_range, + "BLOCK_SIZE_K": block_k_range, + "GROUP_SIZE_M": group_m_range, + "num_warps": num_warps_range, + "num_stages": num_stage_range, + } + + keys, values = zip(*param_ranges.items()) + for config_values in product(*values): + config = dict(zip(keys, config_values)) + configs.append(config) + + # Remove configs that are not compatible with fp8 block quantization + # BLOCK_SIZE_K must be a multiple of block_k + # BLOCK_SIZE_N must be a multiple of block_n + if block_quant_shape is not None and not use_fp16: + block_n, block_k = block_quant_shape[0], block_quant_shape[1] + for config in configs[:]: + if ( + config["BLOCK_SIZE_K"] % block_k != 0 + or config["BLOCK_SIZE_N"] % block_n != 0 + ): + configs.remove(config) + return configs + + +def prune_rocm_search_space( + num_tokens, shard_intermediate_size, hidden_size, search_space, is_fp16, topk +): + N1, K1 = shard_intermediate_size, hidden_size + N2, K2 = hidden_size, shard_intermediate_size // 2 + pruned_space_1 = prune_rocm_configs( + num_tokens * topk, N1, K1, search_space, is_fp16 + ) + pruned_space_2 = prune_rocm_configs( + num_tokens * topk, N2, K2, search_space, is_fp16 + ) + search_space = merge_unique_dicts(pruned_space_1, pruned_space_2) + return search_space + + +# The following code is inspired by ROCm/Triton GEMM tuning script: +# https://github.com/ROCm/triton/blob/triton-mlir/scripts/amd/gemm/tune_gemm.py#L89 +def prune_rocm_configs(M, N, K, configs, is_fp16=True): + pruned_configs = [] + elemBytes_a = 2 if is_fp16 else 1 + elemBytes_b = 2 if is_fp16 else 1 + + mfma = 16 if M < 32 or N < 32 else 32 + + # TODO (zhanglx): figure out the boundary between large and small gemms + large_gemm = False + if M >= 2048 and N >= 2048: + large_gemm = True + + for config in configs: + BLOCK_SIZE_M = config.get("BLOCK_SIZE_M") + BLOCK_SIZE_N = config.get("BLOCK_SIZE_N") + BLOCK_SIZE_K = config.get("BLOCK_SIZE_K") + num_warps = config.get("num_warps") + + if is_fp16: + matrix_instr_nonkdim = config.get("matrix_instr_nonkdim") + if matrix_instr_nonkdim > mfma: + continue + if mfma == 4 and BLOCK_SIZE_K < 64: + continue + # some layouts could not work properly in case + # number elements per thread is less 1 + if BLOCK_SIZE_M * BLOCK_SIZE_N < 64: + continue + SPLIT_K = config.get("SPLIT_K", 1) + GROUP_M = config.get("GROUP_SIZE_M") + if is_fp16: + if ( + matrix_instr_nonkdim > BLOCK_SIZE_M + or matrix_instr_nonkdim > BLOCK_SIZE_N + ): + continue + if matrix_instr_nonkdim >= M and matrix_instr_nonkdim != BLOCK_SIZE_M: + continue + if matrix_instr_nonkdim >= N and matrix_instr_nonkdim != BLOCK_SIZE_N: + continue + # Skip BLOCK_SIZE that is too large compare to M/N + # unless BLOCK_SIZE is already small enough + if M * 2 < BLOCK_SIZE_M and BLOCK_SIZE_M != 16: + continue + if N * 2 < BLOCK_SIZE_N and BLOCK_SIZE_N != 16: + continue + # skip large split_k when not necessary + if SPLIT_K != 1 and not need_split_k(M, N, K): + continue + # skip split_k that leads to EVEN_K = false + leap = SPLIT_K * BLOCK_SIZE_K + modv = K % leap + if modv != 0: + continue + # skip large GROUP_M + if GROUP_M * BLOCK_SIZE_M > M and GROUP_M != 1: + continue + # out of shared memory resource + # TODO (zhanglx): This does not consider the LDS usage in the epilogue + LDS = ( + BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a + + BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b + ) + if LDS > 65536: + continue + # Skip small block sizes and num_warps for large gemm + # For fp16 and f8, we want to only use BLOCK_SIZE >= 64 + if large_gemm: + if BLOCK_SIZE_M < 64 or BLOCK_SIZE_N < 64: + continue + if BLOCK_SIZE_K < 64: + continue + if num_warps < 4: + continue + + pruned_configs.append(config) + + return pruned_configs + + +def need_split_k(SIZE_M, SIZE_N, SIZE_K): + return (SIZE_M < 64 or SIZE_N < 64) and SIZE_K > 1024 + + +def merge_unique_dicts(list1, list2): + result = [] + combined_list = list1.copy() + combined_list.extend(list2) + for dictionary in combined_list: + if dictionary not in result: + result.append(dictionary) + return result + + +@ray.remote(num_gpus=1) +class BenchmarkWorker: + def __init__(self, seed: int) -> None: + torch.set_default_device("cuda") + current_platform.seed_everything(seed) + self.seed = seed + # Get the device ID to allocate tensors and kernels + # on the respective GPU. This is required for Ray to work + # correctly with multi-GPU tuning on the ROCm platform. + self.device_id = int(ray.get_gpu_ids()[0]) + + def benchmark( + self, + num_tokens: int, + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + block_quant_shape: list[int] = None, + use_deep_gemm: bool = False, + ) -> tuple[dict[str, int], float]: + current_platform.seed_everything(self.seed) + dtype_str = get_config_dtype_str( + dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8 + ) + # NOTE(woosuk): The current naming convention uses w2.shape[2], which + # is the intermediate size after silu_and_mul. + op_config = get_moe_configs( + num_experts, shard_intermediate_size // 2, dtype_str + ) + if op_config is None: + config = get_default_config( + num_tokens, + num_experts, + shard_intermediate_size, + hidden_size, + topk, + dtype_str, + is_marlin=False, + ) + else: + config = op_config[min(op_config.keys(), key=lambda x: abs(x - num_tokens))] + kernel_time = benchmark_config( + config, + num_tokens, + num_experts, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=100, + block_quant_shape=block_quant_shape, + use_deep_gemm=use_deep_gemm, + ) + return config, kernel_time + + def tune( + self, + num_tokens: int, + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + search_space: list[dict[str, int]], + block_quant_shape: list[int], + use_deep_gemm: bool, + ) -> dict[str, int]: + best_config = None + best_time = float("inf") + if current_platform.is_rocm(): + is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16) + search_space = prune_rocm_search_space( + num_tokens, + shard_intermediate_size, + hidden_size, + search_space, + is_fp16, + topk, + ) + + need_device_guard = False + if current_platform.is_rocm(): + visible_device = os.environ.get("ROCR_VISIBLE_DEVICES", None) + if visible_device != f"{self.device_id}": + need_device_guard = True + + with torch.cuda.device(self.device_id) if need_device_guard else nullcontext(): + for config in tqdm(search_space): + try: + kernel_time = benchmark_config( + config, + num_tokens, + num_experts, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=20, + block_quant_shape=block_quant_shape, + use_deep_gemm=use_deep_gemm, + ) + except triton.runtime.autotuner.OutOfResources: + # Some configurations may be invalid and fail to compile. + continue + + if kernel_time < best_time: + best_time = kernel_time + best_config = config + now = datetime.now() + print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}") + assert best_config is not None + return best_config + + +def sort_config(config: BenchmarkConfig) -> BenchmarkConfig: + return { + "BLOCK_SIZE_M": config["BLOCK_SIZE_M"], + "BLOCK_SIZE_N": config["BLOCK_SIZE_N"], + "BLOCK_SIZE_K": config["BLOCK_SIZE_K"], + "GROUP_SIZE_M": config["GROUP_SIZE_M"], + "num_warps": config["num_warps"], + "num_stages": config["num_stages"], + **( + {"waves_per_eu": config["waves_per_eu"]} if "waves_per_eu" in config else {} + ), + **( + {"matrix_instr_nonkdim": config["matrix_instr_nonkdim"]} + if "matrix_instr_nonkdim" in config + else {} + ), + **({"kpack": config["kpack"]} if "kpack" in config else {}), + } + + +def save_configs( + configs: dict[int, BenchmarkConfig], + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + block_quant_shape: list[int], +) -> None: + dtype_str = get_config_dtype_str( + dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8 + ) + + # NOTE(woosuk): The current naming convention uses w2.shape[2], which + # is the intermediate size after silu_and_mul. + filename = get_config_file_name( + num_experts, shard_intermediate_size // 2, dtype_str, block_quant_shape + ) + + print(f"Writing best config to {filename}...") + with open(filename, "w") as f: + json.dump(configs, f, indent=4) + f.write("\n") + + +def get_weight_block_size_safety(config, default_value=None): + quantization_config = getattr(config, "quantization_config", {}) + if isinstance(quantization_config, dict): + return quantization_config.get("weight_block_size", default_value) + return default_value + + +def main(args: argparse.Namespace): + print(args) + + config = get_config(model=args.model, trust_remote_code=args.trust_remote_code) + if args.model_prefix: + config = getattr(config, args.model_prefix) + + if config.architectures[0] == "DbrxForCausalLM": + E = config.ffn_config.moe_num_experts + topk = config.ffn_config.moe_top_k + intermediate_size = config.ffn_config.ffn_hidden_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + elif config.architectures[0] == "JambaForCausalLM": + E = config.num_experts + topk = config.num_experts_per_tok + intermediate_size = config.intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + elif config.architectures[0] in ("DeepseekV3ForCausalLM", "DeepseekV2ForCausalLM"): + E = config.n_routed_experts + topk = config.num_experts_per_tok + intermediate_size = config.moe_intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + elif config.architectures[0] in ("Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"): + E = config.num_experts + topk = config.num_experts_per_tok + intermediate_size = config.moe_intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + else: + # Support for llama4 + config = config.get_text_config() + # Default: Mixtral. + E = config.num_local_experts + topk = config.num_experts_per_tok + intermediate_size = config.intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + + hidden_size = config.hidden_size + dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype + use_fp8_w8a8 = args.dtype == "fp8_w8a8" + use_int8_w8a16 = args.dtype == "int8_w8a16" + block_quant_shape = get_weight_block_size_safety(config) + + if args.batch_size is None: + batch_sizes = [ + 1, + 2, + 4, + 8, + 16, + 24, + 32, + 48, + 64, + 96, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ] + else: + batch_sizes = args.batch_size + + use_deep_gemm = bool(args.use_deep_gemm) + + if current_platform.is_rocm() and "HIP_VISIBLE_DEVICES" in os.environ: + # Ray will set ROCR_VISIBLE_DEVICES for device visibility + logger.warning( + "Ray uses ROCR_VISIBLE_DEVICES to control device accessibility." + "Replacing HIP_VISIBLE_DEVICES with ROCR_VISIBLE_DEVICES." + ) + val = os.environ["HIP_VISIBLE_DEVICES"] + os.environ["ROCR_VISIBLE_DEVICES"] = val + del os.environ["HIP_VISIBLE_DEVICES"] + + ray.init() + num_gpus = int(ray.available_resources()["GPU"]) + workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)] + + def _distribute(method: str, inputs: list[Any]) -> list[Any]: + outputs = [] + worker_idx = 0 + for input_args in inputs: + worker = workers[worker_idx] + worker_method = getattr(worker, method) + output = worker_method.remote(*input_args) + outputs.append(output) + worker_idx = (worker_idx + 1) % num_gpus + return ray.get(outputs) + + if args.tune: + is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16) + search_space = get_configs_compute_bound(is_fp16, block_quant_shape) + print(f"Start tuning over {len(search_space)} configurations...") + + start = time.time() + configs = _distribute( + "tune", + [ + ( + batch_size, + E, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + search_space, + block_quant_shape, + use_deep_gemm, + ) + for batch_size in batch_sizes + ], + ) + best_configs = { + M: sort_config(config) for M, config in zip(batch_sizes, configs) + } + save_configs( + best_configs, + E, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + block_quant_shape, + ) + end = time.time() + print(f"Tuning took {end - start:.2f} seconds") + else: + outputs = _distribute( + "benchmark", + [ + ( + batch_size, + E, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + block_quant_shape, + use_deep_gemm, + ) + for batch_size in batch_sizes + ], + ) + + for batch_size, (config, kernel_time) in zip(batch_sizes, outputs): + print(f"Batch size: {batch_size}, config: {config}") + print(f"Kernel time: {kernel_time:.2f} us") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser() + parser.add_argument( + "--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1" + ) + parser.add_argument( + "--tp-size", "-tp", "--tensor-parallel-size", type=int, default=2 + ) + parser.add_argument( + "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto" + ) + parser.add_argument("--use-deep-gemm", action="store_true") + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--batch-size", type=int, nargs="+", required=False) + parser.add_argument("--tune", action="store_true") + parser.add_argument("--trust-remote-code", action="store_true") + parser.add_argument("--model-prefix", type=str, required=False) + args = parser.parse_args() + + main(args) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_moe_align_block_size.py b/offline_benchmark_test/benchmarks/kernels/benchmark_moe_align_block_size.py new file mode 100644 index 0000000000000000000000000000000000000000..5170ac09dc42af8fac0f15e48dc5d348b728f04a --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_moe_align_block_size.py @@ -0,0 +1,159 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import argparse +import itertools + +import torch + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.fused_moe.moe_align_block_size import ( + moe_align_block_size_triton, +) +from vllm.triton_utils import triton + + +def get_topk_ids(num_tokens: int, num_experts: int, topk: int) -> torch.Tensor: + return torch.stack( + [ + torch.randperm(num_experts, dtype=torch.int32, device="cuda")[:topk] + for _ in range(num_tokens) + ] + ) + + +def check_correctness(num_tokens, num_experts=256, block_size=256, topk=8): + """ + Verifies vllm vs. Triton + """ + topk_ids = get_topk_ids(num_tokens, num_experts, topk) + + # 1. malloc space for triton and vllm + # malloc enough space (max_num_tokens_padded) for the sorted ids + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + sorted_ids_triton = torch.empty( + (max_num_tokens_padded,), dtype=torch.int32, device="cuda" + ) + sorted_ids_triton.fill_(topk_ids.numel()) # fill with sentinel value + expert_ids_triton = torch.zeros( + (max_num_tokens_padded // block_size,), dtype=torch.int32, device="cuda" + ) + num_tokens_post_pad_triton = torch.empty((1,), dtype=torch.int32, device="cuda") + + sorted_ids_vllm = torch.empty_like(sorted_ids_triton) + sorted_ids_vllm.fill_(topk_ids.numel()) + expert_ids_vllm = torch.zeros_like(expert_ids_triton) + num_tokens_post_pad_vllm = torch.empty_like(num_tokens_post_pad_triton) + + # 2. run implementations + moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids_triton, + expert_ids_triton, + num_tokens_post_pad_triton, + ) + + ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids_vllm, + expert_ids_vllm, + num_tokens_post_pad_vllm, + ) + print(f"✅ VLLM implementation works with {num_experts} experts!") + + # 3. compare results + if torch.allclose(expert_ids_triton, expert_ids_vllm) and torch.allclose( + num_tokens_post_pad_triton, num_tokens_post_pad_vllm + ): + print("✅ Triton and VLLM implementations match.") + else: + print("❌ Triton and VLLM implementations DO NOT match.") + print("Triton expert_ids:", expert_ids_triton) + print("VLLM expert_ids:", expert_ids_vllm) + print("Triton num_tokens_post_pad:", num_tokens_post_pad_triton) + print("VLLM num_tokens_post_pad:", num_tokens_post_pad_vllm) + + +# test configurations +num_tokens_range = [1, 16, 256, 4096] +num_experts_range = [16, 64, 224, 256, 280, 512] +topk_range = [1, 2, 8] +configs = list(itertools.product(num_tokens_range, num_experts_range, topk_range)) + + +@triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["num_tokens", "num_experts", "topk"], + x_vals=configs, + line_arg="provider", + line_vals=["vllm", "triton"], # "triton" + line_names=["VLLM", "Triton"], # "Triton" + plot_name="moe-align-block-size-performance", + args={}, + ) +) +def benchmark(num_tokens, num_experts, topk, provider): + """Benchmark function for Triton.""" + block_size = 256 + topk_ids = get_topk_ids(num_tokens, num_experts, topk) + + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + sorted_ids = torch.empty((max_num_tokens_padded,), dtype=torch.int32, device="cuda") + sorted_ids.fill_(topk_ids.numel()) + max_num_m_blocks = max_num_tokens_padded // block_size + expert_ids = torch.empty((max_num_m_blocks,), dtype=torch.int32, device="cuda") + num_tokens_post_pad = torch.empty((1,), dtype=torch.int32, device="cuda") + + quantiles = [0.5, 0.2, 0.8] + + if provider == "vllm": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids.clone(), + expert_ids.clone(), + num_tokens_post_pad.clone(), + ), + quantiles=quantiles, + ) + elif provider == "triton": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids.clone(), + expert_ids.clone(), + num_tokens_post_pad.clone(), + ), + quantiles=quantiles, + ) + + return 1000 * ms, 1000 * max_ms, 1000 * min_ms + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--num_experts", + type=int, + default=64, + choices=[8, 16, 32, 64, 128, 256], + ) + parser.add_argument( + "--topk", + type=int, + default=8, + choices=[2, 4, 8], + help="Top-k value for correctness check.", + ) + args = parser.parse_args() + + print("Running correctness check...") + check_correctness(num_tokens=1024, num_experts=args.num_experts, topk=args.topk) + benchmark.run(print_data=True, show_plots=True) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_moe_permute_unpermute.py b/offline_benchmark_test/benchmarks/kernels/benchmark_moe_permute_unpermute.py new file mode 100644 index 0000000000000000000000000000000000000000..dba1f3943b96c370acf1262273695dbdb4097711 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_moe_permute_unpermute.py @@ -0,0 +1,417 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +from typing import Any, TypedDict + +import ray +import torch +from transformers import AutoConfig + +from vllm.model_executor.layers.fused_moe.deep_gemm_moe import ( + _moe_permute, + _moe_unpermute_and_reduce, +) +from vllm.model_executor.layers.fused_moe.fused_moe import * +from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import * +from vllm.model_executor.layers.fused_moe.utils import _fp8_quantize +from vllm.platforms import current_platform +from vllm.utils import FlexibleArgumentParser + +FP8_DTYPE = current_platform.fp8_dtype() + + +class BenchmarkConfig(TypedDict): + BLOCK_SIZE_M: int + BLOCK_SIZE_N: int + BLOCK_SIZE_K: int + GROUP_SIZE_M: int + num_warps: int + num_stages: int + + +def benchmark_permute( + num_tokens: int, + num_experts: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + num_iters: int = 100, + use_customized_permute: bool = False, +) -> float: + # init_dtype = torch.float16 if use_fp8_w8a8 else dtype + hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype) + # output_hidden_states = torch.empty_like(hidden_states) + if use_fp8_w8a8: + align_block_size = 128 # deepgemm needs 128 m aligned block + qhidden_states, scale = _fp8_quantize(hidden_states, None, None) + else: + align_block_size = None + qhidden_states = hidden_states + + gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32) + + input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32) + topk_weights, topk_ids, token_expert_indices = fused_topk( + qhidden_states, input_gating, topk, False + ) + + def prepare(i: int): + input_gating.copy_(gating_output[i]) + + def run(): + if use_customized_permute: + (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = ( + moe_permute( + qhidden_states, + topk_weights=topk_weights, + topk_ids=topk_ids, + token_expert_indices=token_expert_indices, + topk=topk, + n_expert=num_experts, + n_local_expert=num_experts, + expert_map=None, + align_block_size=align_block_size, + ) + ) + else: + ( + permuted_hidden_states, + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) = _moe_permute( + qhidden_states, None, topk_ids, num_experts, None, align_block_size + ) + + # JIT compilation & warmup + run() + torch.cuda.synchronize() + + # Capture 10 invocations with CUDA graph + graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(graph): + for _ in range(10): + run() + torch.cuda.synchronize() + + # Warmup + for _ in range(5): + graph.replay() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + prepare(i) + torch.cuda.synchronize() + + start_event.record() + graph.replay() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + graph.reset() + return avg + + +def benchmark_unpermute( + num_tokens: int, + num_experts: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + num_iters: int = 100, + use_customized_permute: bool = False, +) -> float: + # init_dtype = torch.float16 if use_fp8_w8a8 else dtype + hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype) + output_hidden_states = torch.empty_like(hidden_states) + if use_fp8_w8a8: + align_block_size = 128 # deepgemm needs 128 m aligned block + qhidden_states, scale = _fp8_quantize(hidden_states, None, None) + else: + align_block_size = None + qhidden_states = hidden_states + + input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32) + + topk_weights, topk_ids, token_expert_indices = fused_topk( + qhidden_states, input_gating, topk, False + ) + + def prepare(): + if use_customized_permute: + (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = ( + moe_permute( + qhidden_states, + topk_weights=topk_weights, + topk_ids=topk_ids, + token_expert_indices=token_expert_indices, + topk=topk, + n_expert=num_experts, + n_local_expert=num_experts, + expert_map=None, + align_block_size=align_block_size, + ) + ) + # convert to fp16/bf16 as gemm output + return ( + permuted_hidden_states.to(dtype), + first_token_off, + inv_perm_idx, + m_indices, + ) + else: + ( + permuted_qhidden_states, + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) = _moe_permute( + qhidden_states, None, topk_ids, num_experts, None, align_block_size + ) + # convert to fp16/bf16 as gemm output + return ( + permuted_qhidden_states.to(dtype), + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) + + def run(input: tuple): + if use_customized_permute: + (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = input + moe_unpermute( + permuted_hidden_states, + topk_weights, + topk_ids, + inv_perm_idx, + first_token_off, + topk, + num_experts, + num_experts, + ) + else: + ( + permuted_hidden_states, + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) = input + _moe_unpermute_and_reduce( + output_hidden_states, permuted_hidden_states, inv_perm, topk_weights + ) + + # JIT compilation & warmup + input = prepare() + run(input) + torch.cuda.synchronize() + + # Capture 10 invocations with CUDA graph + graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(graph): + for _ in range(10): + run(input) + torch.cuda.synchronize() + + # Warmup + for _ in range(5): + graph.replay() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + torch.cuda.synchronize() + start_event.record() + graph.replay() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + graph.reset() + return avg + + +@ray.remote(num_gpus=1) +class BenchmarkWorker: + def __init__(self, seed: int) -> None: + torch.set_default_device("cuda") + current_platform.seed_everything(seed) + self.seed = seed + # Get the device ID to allocate tensors and kernels + # on the respective GPU. This is required for Ray to work + # correctly with multi-GPU tuning on the ROCm platform. + self.device_id = int(ray.get_gpu_ids()[0]) + + def benchmark( + self, + num_tokens: int, + num_experts: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + use_customized_permute: bool = False, + ) -> tuple[dict[str, int], float]: + current_platform.seed_everything(self.seed) + + permute_time = benchmark_permute( + num_tokens, + num_experts, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=100, + use_customized_permute=use_customized_permute, + ) + unpermute_time = benchmark_unpermute( + num_tokens, + num_experts, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=100, + use_customized_permute=use_customized_permute, + ) + return permute_time, unpermute_time + + +def get_weight_block_size_safety(config, default_value=None): + quantization_config = getattr(config, "quantization_config", {}) + if isinstance(quantization_config, dict): + return quantization_config.get("weight_block_size", default_value) + return default_value + + +def main(args: argparse.Namespace): + print(args) + + config = AutoConfig.from_pretrained( + args.model, trust_remote_code=args.trust_remote_code + ) + if config.architectures[0] == "DbrxForCausalLM": + E = config.ffn_config.moe_num_experts + topk = config.ffn_config.moe_top_k + elif config.architectures[0] == "JambaForCausalLM": + E = config.num_experts + topk = config.num_experts_per_tok + elif ( + config.architectures[0] == "DeepseekV3ForCausalLM" + or config.architectures[0] == "DeepseekV2ForCausalLM" + ): + E = config.n_routed_experts + topk = config.num_experts_per_tok + elif config.architectures[0] in ["Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"]: + E = config.num_experts + topk = config.num_experts_per_tok + + else: + # Support for llama4 + config = config.get_text_config() + # Default: Mixtral. + E = config.num_local_experts + topk = config.num_experts_per_tok + + hidden_size = config.hidden_size + dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype + use_fp8_w8a8 = args.dtype == "fp8_w8a8" + use_int8_w8a16 = args.dtype == "int8_w8a16" + use_customized_permute = args.use_customized_permute + + if args.batch_size is None: + batch_sizes = [ + 1, + 2, + 4, + 8, + 16, + 24, + 32, + 48, + 64, + 96, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ] + else: + batch_sizes = [args.batch_size] + + ray.init() + num_gpus = int(ray.available_resources()["GPU"]) + workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)] + + def _distribute(method: str, inputs: list[Any]) -> list[Any]: + outputs = [] + worker_idx = 0 + for input_args in inputs: + worker = workers[worker_idx] + worker_method = getattr(worker, method) + output = worker_method.remote(*input_args) + outputs.append(output) + worker_idx = (worker_idx + 1) % num_gpus + return ray.get(outputs) + + outputs = _distribute( + "benchmark", + [ + ( + batch_size, + E, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + use_customized_permute, + ) + for batch_size in batch_sizes + ], + ) + + for batch_size, (permute, unpermute) in zip(batch_sizes, outputs): + print(f"Batch size: {batch_size}") + print(f"Permute time: {permute:.2f} us") + print(f"Unpermute time: {unpermute:.2f} us") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser() + parser.add_argument( + "--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1" + ) + parser.add_argument( + "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto" + ) + parser.add_argument("--use-customized-permute", action="store_true") + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--batch-size", type=int, required=False) + parser.add_argument("--trust-remote-code", action="store_true") + args = parser.parse_args() + + main(args) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_paged_attention.py b/offline_benchmark_test/benchmarks/kernels/benchmark_paged_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..7e0376c18ecc79690ca49e3995258c5e749bbac7 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_paged_attention.py @@ -0,0 +1,251 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import random +import time +from typing import Optional + +import torch + +from vllm import _custom_ops as ops +from vllm.logger import init_logger +from vllm.platforms import current_platform +from vllm.utils import ( + STR_DTYPE_TO_TORCH_DTYPE, + FlexibleArgumentParser, + create_kv_caches_with_random, +) + +logger = init_logger(__name__) + +NUM_BLOCKS = 128 * 1024 +PARTITION_SIZE = 512 +PARTITION_SIZE_ROCM = 256 + + +@torch.inference_mode() +def main( + version: str, + num_seqs: int, + seq_len: int, + num_query_heads: int, + num_kv_heads: int, + head_size: int, + use_alibi: bool, + block_size: int, + dtype: torch.dtype, + seed: int, + do_profile: bool, + device: str = "cuda", + kv_cache_dtype: Optional[str] = None, +) -> None: + current_platform.seed_everything(seed) + + scale = float(1.0 / (head_size**0.5)) + query = torch.empty( + num_seqs, num_query_heads, head_size, dtype=dtype, device=device + ) + query.uniform_(-scale, scale) + + assert num_query_heads % num_kv_heads == 0 + alibi_slopes = None + if use_alibi: + alibi_slopes = torch.randn(num_query_heads, dtype=torch.float, device=device) + + seq_lens = [seq_len for _ in range(num_seqs)] + max_seq_len = max(seq_lens) + seq_lens = torch.tensor(seq_lens, dtype=torch.int, device=device) + + # Create the block tables. + max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size + block_tables_lst: list[list[int]] = [] + for _ in range(num_seqs): + block_table = [ + random.randint(0, NUM_BLOCKS - 1) for _ in range(max_num_blocks_per_seq) + ] + block_tables_lst.append(block_table) + + block_tables = torch.tensor(block_tables_lst, dtype=torch.int, device=device) + + # Create the KV cache. + key_caches, value_caches = create_kv_caches_with_random( + NUM_BLOCKS, + block_size, + 1, + num_kv_heads, + head_size, + kv_cache_dtype, + dtype, + device=device, + ) + key_cache, value_cache = key_caches[0], value_caches[0] + + # Prepare for the paged attention kernel. + output = torch.empty_like(query) + if version == "v2": + if current_platform.is_rocm(): + global PARTITION_SIZE + if not args.custom_paged_attn and not current_platform.is_navi(): + PARTITION_SIZE = 1024 + else: + PARTITION_SIZE = PARTITION_SIZE_ROCM + num_partitions = (max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE + tmp_output = torch.empty( + size=(num_seqs, num_query_heads, num_partitions, head_size), + dtype=output.dtype, + device=output.device, + ) + exp_sums = torch.empty( + size=(num_seqs, num_query_heads, num_partitions), + dtype=torch.float32, + device=output.device, + ) + max_logits = torch.empty_like(exp_sums) + + def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: + torch.cuda.synchronize() + if profile: + torch.cuda.cudart().cudaProfilerStart() + start_time = time.perf_counter() + + # Using default kv_scale + k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device) + + for _ in range(num_iters): + if version == "v1": + ops.paged_attention_v1( + output, + query, + key_cache, + value_cache, + num_kv_heads, + scale, + block_tables, + seq_lens, + block_size, + max_seq_len, + alibi_slopes, + kv_cache_dtype, + k_scale, + v_scale, + ) + elif version == "v2": + if not args.custom_paged_attn: + ops.paged_attention_v2( + output, + exp_sums, + max_logits, + tmp_output, + query, + key_cache, + value_cache, + num_kv_heads, + scale, + block_tables, + seq_lens, + block_size, + max_seq_len, + alibi_slopes, + kv_cache_dtype, + k_scale, + v_scale, + ) + else: + ops.paged_attention_rocm( + output, + exp_sums, + max_logits, + tmp_output, + query, + key_cache, + value_cache, + num_kv_heads, + scale, + block_tables, + seq_lens, + None, + block_size, + max_seq_len, + alibi_slopes, + kv_cache_dtype, + k_scale, + v_scale, + ) + else: + raise ValueError(f"Invalid version: {version}") + torch.cuda.synchronize() + + end_time = time.perf_counter() + if profile: + torch.cuda.cudart().cudaProfilerStop() + return (end_time - start_time) / num_iters + + # Warmup. + print("Warming up...") + run_benchmark = run_cuda_benchmark + run_benchmark(num_iters=3, profile=False) + + # Benchmark. + if do_profile: + latency = run_benchmark(num_iters=1, profile=True) + else: + latency = run_benchmark(num_iters=100, profile=False) + print(f"Kernel running time: {latency * 1000000:.3f} us") + + +if __name__ == "__main__": + logger.warning( + "This script benchmarks the paged attention kernel. " + "By default this is no longer used in vLLM inference." + ) + + parser = FlexibleArgumentParser(description="Benchmark the paged attention kernel.") + parser.add_argument("--version", type=str, choices=["v1", "v2"], default="v2") + parser.add_argument("--batch-size", type=int, default=8) + parser.add_argument("--seq-len", type=int, default=4096) + parser.add_argument("--num-query-heads", type=int, default=64) + parser.add_argument("--num-kv-heads", type=int, default=8) + parser.add_argument( + "--head-size", + type=int, + choices=[64, 80, 96, 112, 120, 128, 192, 256], + default=128, + ) + parser.add_argument("--block-size", type=int, choices=[16, 32], default=16) + parser.add_argument("--use-alibi", action="store_true") + parser.add_argument( + "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half" + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--profile", action="store_true") + parser.add_argument( + "--kv-cache-dtype", + type=str, + choices=["auto", "fp8", "fp8_e5m2", "fp8_e4m3"], + default="auto", + help="Data type for kv cache storage. If 'auto', will use model " + "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. " + "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)", + ) + parser.add_argument( + "--custom-paged-attn", action="store_true", help="Use custom paged attention" + ) + args = parser.parse_args() + print(args) + + if args.num_query_heads % args.num_kv_heads != 0: + raise ValueError("num_query_heads must be divisible by num_kv_heads") + main( + version=args.version, + num_seqs=args.batch_size, + seq_len=args.seq_len, + num_query_heads=args.num_query_heads, + num_kv_heads=args.num_kv_heads, + head_size=args.head_size, + block_size=args.block_size, + use_alibi=args.use_alibi, + dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype], + seed=args.seed, + do_profile=args.profile, + kv_cache_dtype=args.kv_cache_dtype, + ) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_quant.py b/offline_benchmark_test/benchmarks/kernels/benchmark_quant.py new file mode 100644 index 0000000000000000000000000000000000000000..6ab26f5f1adf73ccdc68c0d5f20d622d0c09b7ba --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_quant.py @@ -0,0 +1,108 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import time + +import torch + +from vllm import _custom_ops as ops +from vllm.platforms import current_platform +from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser + + +@torch.inference_mode() +def main( + num_tokens: int, + hidden_size: int, + static_scale: bool, + quant_dtype: torch.dtype, + dtype: torch.dtype, + seed: int = 0, + do_profile: bool = False, + num_warmup_iters: int = 5, + num_iters: int = 100, +) -> None: + current_platform.seed_everything(seed) + torch.set_default_device("cuda") + + x = torch.randn(num_tokens, hidden_size, dtype=dtype) + scale = torch.randn(1, 1, dtype=torch.float32) if static_scale else None + + def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: + torch.cuda.synchronize() + if profile: + torch.cuda.cudart().cudaProfilerStart() + start_time = time.perf_counter() + + for _ in range(num_iters): + if quant_dtype == torch.int8: + ops.scaled_int8_quant(x, scale) + else: + ops.scaled_fp8_quant(x, scale) + torch.cuda.synchronize() + + end_time = time.perf_counter() + if profile: + torch.cuda.cudart().cudaProfilerStop() + return (end_time - start_time) / num_iters + + # Warmup. + print("Warming up...") + run_benchmark = run_cuda_benchmark + run_benchmark(num_iters=num_warmup_iters, profile=False) + + # Benchmark. + if do_profile: + latency = run_benchmark(num_iters=1, profile=True) + else: + latency = run_benchmark(num_iters=num_iters, profile=False) + print(f"Kernel running time: {latency * 1000000:.3f} us") + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "int8": + return torch.int8 + if dt == "fp8": + return torch.float8_e4m3fn + raise ValueError(f"Unsupported dtype: {dt}") + + parser = FlexibleArgumentParser( + description="Benchmark the quantization (fp8 or int8) kernel." + ) + parser.add_argument("--num-tokens", type=int, default=4096) + parser.add_argument("--hidden-size", type=int, default=8192) + parser.add_argument("--static-scale", action="store_true") + parser.add_argument( + "--quant-dtype", type=str, choices=["fp8", "int8"], default="int8" + ) + parser.add_argument( + "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half" + ) + + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--profile", action="store_true") + parser.add_argument("--num-warmup-iters", type=int, default=5) + parser.add_argument( + "--num-iters", + type=int, + default=100, + help="Number of benchmark iterations. " + "If --profile is set, this number is ignored", + ) + + args = parser.parse_args() + print(args) + + main( + num_tokens=args.num_tokens, + hidden_size=args.hidden_size, + static_scale=args.static_scale, + quant_dtype=to_torch_dtype(args.quant_dtype), + dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype], + seed=args.seed, + do_profile=args.profile, + num_warmup_iters=args.num_warmup_iters, + num_iters=args.num_iters, + ) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_rmsnorm.py b/offline_benchmark_test/benchmarks/kernels/benchmark_rmsnorm.py new file mode 100644 index 0000000000000000000000000000000000000000..4cf633a81358d80202fa383d7312dac4a818ea62 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_rmsnorm.py @@ -0,0 +1,256 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import itertools +from typing import Optional, Union + +import torch +from flashinfer.norm import fused_add_rmsnorm, rmsnorm +from torch import nn + +from vllm import _custom_ops as vllm_ops +from vllm.triton_utils import triton + + +class HuggingFaceRMSNorm(nn.Module): + def __init__(self, hidden_size: int, eps: float = 1e-6) -> None: + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward( + self, + x: torch.Tensor, + residual: Optional[torch.Tensor] = None, + ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]: + orig_dtype = x.dtype + x = x.to(torch.float32) + if residual is not None: + x = x + residual.to(torch.float32) + residual = x.to(orig_dtype) + + variance = x.pow(2).mean(dim=-1, keepdim=True) + x = x * torch.rsqrt(variance + self.variance_epsilon) + x = x.to(orig_dtype) * self.weight + if residual is None: + return x + else: + return x, residual + + +def rmsnorm_naive( + x: torch.Tensor, + weight: torch.Tensor, + residual: Optional[torch.Tensor] = None, + eps: float = 1e-6, +): + naive_norm = HuggingFaceRMSNorm(x.shape[-1], eps=eps) + naive_norm.weight = nn.Parameter(weight) + naive_norm = naive_norm.to(x.device) + + orig_shape = x.shape + x = x.view(-1, x.shape[-1]) + if residual is not None: + residual = residual.view(-1, residual.shape[-1]) + + output = naive_norm(x, residual) + + if isinstance(output, tuple): + output = (output[0].view(orig_shape), output[1].view(orig_shape)) + else: + output = output.view(orig_shape) + return output + + +def rmsnorm_flashinfer( + x: torch.Tensor, + weight: torch.Tensor, + residual: Optional[torch.Tensor] = None, + eps: float = 1e-6, +): + orig_shape = x.shape + x = x.view(-1, x.shape[-1]) + if residual is not None: + residual = residual.view(-1, residual.shape[-1]) + + if residual is not None: + fused_add_rmsnorm(x, residual, weight, eps) + output = (x, residual) + else: + output = rmsnorm(x, weight, eps) + + if isinstance(output, tuple): + output = (output[0].view(orig_shape), output[1].view(orig_shape)) + else: + output = output.view(orig_shape) + return output + + +def rmsnorm_vllm( + x: torch.Tensor, + weight: torch.Tensor, + residual: Optional[torch.Tensor] = None, + eps: float = 1e-6, +): + orig_shape = x.shape + x = x.view(-1, x.shape[-1]) + if residual is not None: + residual = residual.view(-1, residual.shape[-1]) + + if residual is not None: + vllm_ops.fused_add_rms_norm(x, residual, weight, eps) + output = (x, residual) + else: + out = torch.empty_like(x) + vllm_ops.rms_norm(out, x, weight, eps) + output = out + + if isinstance(output, tuple): + output = (output[0].view(orig_shape), output[1].view(orig_shape)) + else: + output = output.view(orig_shape) + return output + + +def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True): + dtype = torch.bfloat16 + x = torch.randn(batch_size, seq_len, hidden_size, dtype=dtype, device="cuda") + weight = torch.ones(hidden_size, dtype=dtype, device="cuda") + residual = torch.randn_like(x) if use_residual else None + + output_naive = rmsnorm_naive( + x.clone(), weight, residual.clone() if residual is not None else None + ) + output_flashinfer = rmsnorm_flashinfer( + x.clone(), weight, residual.clone() if residual is not None else None + ) + output_vllm = rmsnorm_vllm( + x.clone(), weight, residual.clone() if residual is not None else None + ) + + if use_residual: + output_naive = output_naive[0] + output_flashinfer = output_flashinfer[0] + output_vllm = output_vllm[0] + + print(f"Naive output={output_naive}") + print(f"FlashInfer output={output_flashinfer}") + print(f"vLLM output={output_vllm}") + + if torch.allclose( + output_naive, output_flashinfer, atol=1e-2, rtol=1e-2 + ) and torch.allclose(output_naive, output_vllm, atol=1e-2, rtol=1e-2): + print("✅ All implementations match") + else: + print("❌ Implementations differ") + + +batch_size_range = [2**i for i in range(0, 7, 2)] +seq_length_range = [2**i for i in range(6, 11, 1)] +head_num_range = [32, 48] +configs = list(itertools.product(head_num_range, batch_size_range, seq_length_range)) + + +def get_benchmark(use_residual): + @triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["head_num", "batch_size", "seq_len"], + x_vals=[list(_) for _ in configs], + line_arg="provider", + line_vals=["huggingface", "flashinfer", "vllm"], + line_names=["HuggingFace", "FlashInfer", "vLLM"], + styles=[("blue", "-"), ("green", "-"), ("red", "-")], + ylabel="us", + plot_name=f"rmsnorm-perf-{'with' if use_residual else 'without'}-residual", + args={}, + ) + ) + def benchmark(head_num, batch_size, seq_len, provider): + dtype = torch.bfloat16 + hidden_size = head_num * 128 # assuming head_dim = 128 + + x = torch.randn(batch_size, seq_len, hidden_size, dtype=dtype, device="cuda") + weight = torch.ones(hidden_size, dtype=dtype, device="cuda") + residual = torch.randn_like(x) if use_residual else None + + quantiles = [0.5, 0.2, 0.8] + + if provider == "huggingface": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: rmsnorm_naive( + x.clone(), + weight, + residual.clone() if residual is not None else None, + ), + quantiles=quantiles, + ) + elif provider == "flashinfer": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: rmsnorm_flashinfer( + x.clone(), + weight, + residual.clone() if residual is not None else None, + ), + quantiles=quantiles, + ) + else: + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: rmsnorm_vllm( + x.clone(), + weight, + residual.clone() if residual is not None else None, + ), + quantiles=quantiles, + ) + + return 1000 * ms, 1000 * max_ms, 1000 * min_ms + + return benchmark + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument( + "--batch-size", + type=int, + default=4, + help="Batch size", + ) + parser.add_argument( + "--seq-len", + type=int, + default=128, + help="Sequence length", + ) + parser.add_argument( + "--hidden-size", + type=int, + default=4096, + help="Hidden size (2nd dimension) of the sequence", + ) + parser.add_argument( + "--use-residual", action="store_true", help="Whether to use residual connection" + ) + parser.add_argument( + "--save-path", + type=str, + default="./configs/rmsnorm/", + help="Path to save rmsnorm benchmark results", + ) + + args = parser.parse_args() + + # Run correctness test + calculate_diff( + batch_size=args.batch_size, + seq_len=args.seq_len, + hidden_size=args.hidden_size, + use_residual=args.use_residual, + ) + + # Get the benchmark function with proper use_residual setting + benchmark = get_benchmark(args.use_residual) + # Run performance benchmark + benchmark.run(print_data=True, save_path=args.save_path) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_rope.py b/offline_benchmark_test/benchmarks/kernels/benchmark_rope.py new file mode 100644 index 0000000000000000000000000000000000000000..b81baf17a8c674edd72fe9a6397a0b027a8d33b9 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_rope.py @@ -0,0 +1,133 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +from itertools import accumulate +from typing import Optional + +import nvtx +import torch + +from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope +from vllm.platforms import current_platform +from vllm.utils import FlexibleArgumentParser + + +def benchmark_rope_kernels_multi_lora( + is_neox_style: bool, + batch_size: int, + seq_len: int, + num_heads: int, + head_size: int, + rotary_dim: Optional[int], + dtype: torch.dtype, + seed: int, + device: str, + max_position: int = 8192, + base: float = 10000, +) -> None: + current_platform.seed_everything(seed) + torch.set_default_device(device) + if rotary_dim is None: + rotary_dim = head_size + # silulating serving 4 LoRAs + scaling_factors = [1, 2, 4, 8] + # batched RoPE can take multiple scaling factors + batched_rope = get_rope( + head_size, + rotary_dim, + max_position, + base, + is_neox_style, + {"rope_type": "linear", "factor": tuple(scaling_factors)}, + ) + # non-batched RoPE takes only one scaling factor, we create multiple + # instances to simulate the same behavior + non_batched_ropes: list[RotaryEmbedding] = [] + for scaling_factor in scaling_factors: + non_batched_ropes.append( + get_rope( + head_size, + rotary_dim, + max_position, + base, + is_neox_style, + {"rope_type": "linear", "factor": (scaling_factor,)}, + ) + ) + + positions = torch.randint(0, max_position, (batch_size, seq_len)) + query = torch.randn(batch_size, seq_len, num_heads * head_size, dtype=dtype) + key = torch.randn_like(query) + + # create query offsets for batched RoPE, we concat multiple kv cache + # together and each query needs to find the right kv cache of its type + offset_map = torch.tensor( + list( + accumulate( + [0] + + [ + max_position * scaling_factor * 2 + for scaling_factor in scaling_factors[:-1] + ] + ) + ) + ) + query_types = torch.randint( + 0, len(scaling_factors), (batch_size, seq_len), device=device + ) + # map query types to offsets + query_offsets = offset_map[query_types] + # the kernel takes flattened offsets + flatten_offsets = query_offsets.flatten() + + # batched queries of the same type together for non-batched RoPE + queries = [query[query_types == i] for i in range(len(scaling_factors))] + keys = [key[query_types == i] for i in range(len(scaling_factors))] + packed_qkr = zip(queries, keys, non_batched_ropes) + # synchronize before start timing + torch.cuda.synchronize() + with nvtx.annotate("non-batched", color="yellow"): + for q, k, r in packed_qkr: + r.forward(positions, q, k) + torch.cuda.synchronize() + with nvtx.annotate("batched", color="green"): + batched_rope.forward(positions, query, key, flatten_offsets) + torch.cuda.synchronize() + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark the rotary embedding kernels." + ) + parser.add_argument("--is-neox-style", type=bool, default=True) + parser.add_argument("--batch-size", type=int, default=16) + parser.add_argument("--seq-len", type=int, default=512) + parser.add_argument("--num-heads", type=int, default=8) + parser.add_argument( + "--head-size", + type=int, + choices=[64, 80, 96, 112, 120, 128, 192, 256], + default=128, + ) + parser.add_argument("--rotary-dim", type=int, choices=[16, 32], default=32) + parser.add_argument( + "--dtype", type=str, choices=["bfloat16", "float"], default="float" + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument( + "--device", type=str, choices=["cuda:0", "cuda:1"], default="cuda:0" + ) + args = parser.parse_args() + print(args) + + benchmark_rope_kernels_multi_lora( + is_neox_style=args.is_neox_style, + batch_size=args.batch_size, + seq_len=args.seq_len, + num_heads=args.num_heads, + head_size=args.head_size, + rotary_dim=args.rotary_dim, + dtype=getattr(torch, args.dtype), + seed=args.seed, + device=args.device, + ) diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_shapes.py b/offline_benchmark_test/benchmarks/kernels/benchmark_shapes.py new file mode 100644 index 0000000000000000000000000000000000000000..18c459c31d3f84d1ca7b7e31d00d50b521177e96 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_shapes.py @@ -0,0 +1,94 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +WEIGHT_SHAPES = { + "ideal": [[4 * 256 * 32, 256 * 32]], + "mistralai/Mistral-7B-v0.1/TP1": [ + [4096, 6144], + [4096, 4096], + [4096, 28672], + [14336, 4096], + ], + "mistralai/Mistral-7B-v0.1/TP2": [ + [4096, 3072], + [2048, 4096], + [4096, 14336], + [7168, 4096], + ], + "mistralai/Mistral-7B-v0.1/TP4": [ + [4096, 1536], + [1024, 4096], + [4096, 7168], + [3584, 4096], + ], + "meta-llama/Llama-2-7b-hf/TP1": [ + [4096, 12288], + [4096, 4096], + [4096, 22016], + [11008, 4096], + ], + "meta-llama/Llama-2-7b-hf/TP2": [ + [4096, 6144], + [2048, 4096], + [4096, 11008], + [5504, 4096], + ], + "meta-llama/Llama-2-7b-hf/TP4": [ + [4096, 3072], + [1024, 4096], + [4096, 5504], + [2752, 4096], + ], + "meta-llama/Llama-2-13b-hf/TP1": [ + [5120, 15360], + [5120, 5120], + [5120, 27648], + [13824, 5120], + ], + "meta-llama/Llama-2-13b-hf/TP2": [ + [5120, 7680], + [2560, 5120], + [5120, 13824], + [6912, 5120], + ], + "meta-llama/Llama-2-13b-hf/TP4": [ + [5120, 3840], + [1280, 5120], + [5120, 6912], + [3456, 5120], + ], + "meta-llama/Llama-2-70b-hf/TP1": [ + [8192, 10240], + [8192, 8192], + [8192, 57344], + [28672, 8192], + ], + "meta-llama/Llama-2-70b-hf/TP2": [ + [8192, 5120], + [4096, 8192], + [8192, 28672], + [14336, 8192], + ], + "meta-llama/Llama-2-70b-hf/TP4": [ + [8192, 2560], + [2048, 8192], + [8192, 14336], + [7168, 8192], + ], +} + +WEIGHT_SHAPES_MOE = { + "nm-testing/Mixtral-8x7B-Instruct-v0.1": [ + [8, 2, 4096, 28672], + [8, 2, 14336, 4096], + ], + "nm-testing/deepseekv2-lite": [ + [64, 6, 2048, 1408], + ], + "ibm-granite/granite-3.0-1b-a400m": [ + [32, 8, 1024, 1024], + ], + "ibm-granite/granite-3.0-3b-a800m": [ + [40, 8, 1024, 1536], + ], +} diff --git a/offline_benchmark_test/benchmarks/kernels/benchmark_w8a8_block_fp8.py b/offline_benchmark_test/benchmarks/kernels/benchmark_w8a8_block_fp8.py new file mode 100644 index 0000000000000000000000000000000000000000..4fcdbadd65ecd366f51ce5d9053cc700854dcb71 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/benchmark_w8a8_block_fp8.py @@ -0,0 +1,414 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +# Adapted from sglang quantization/tuning_block_wise_kernel.py + +import argparse +import json +import multiprocessing as mp +import os +import time +from datetime import datetime +from typing import Any + +import torch +import tqdm +import triton + +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + _w8a8_block_fp8_matmul, +) +from vllm.platforms import current_platform +from vllm.utils import FlexibleArgumentParser + +mp.set_start_method("spawn", force=True) + +assert current_platform.is_cuda(), ( + "Only support tune w8a8 block fp8 kernel on CUDA device." +) + +DTYPE_MAP = { + "float32": torch.float32, + "float16": torch.float16, + "half": torch.half, + "bfloat16": torch.bfloat16, +} + + +def w8a8_block_matmul( + A: torch.Tensor, + B: torch.Tensor, + As: torch.Tensor, + Bs: torch.Tensor, + block_size: list[int], + config: dict[str, Any], + output_dtype: torch.dtype = torch.float16, +) -> torch.Tensor: + """This function performs matrix multiplication with + block-wise quantization. + + It takes two input tensors `A` and `B` with scales `As` and `Bs`. + The output is returned in the specified `output_dtype`. + + Args: + A: The input tensor, e.g., activation. + B: The input tensor, e.g., weight. + As: The per-token-group quantization scale for `A`. + Bs: The per-block quantization scale for `B`. + block_size: The block size for per-block quantization. + It should be 2-dim, e.g., [128, 128]. + output_dytpe: The dtype of the returned tensor. + + Returns: + torch.Tensor: The result of matmul. + """ + assert len(block_size) == 2 + block_n, block_k = block_size[0], block_size[1] + + assert A.shape[-1] == B.shape[-1] + assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous() + assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1] + M = A.numel() // A.shape[-1] + + assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2 + N, K = B.shape + assert triton.cdiv(N, block_n) == Bs.shape[0] + assert triton.cdiv(K, block_k) == Bs.shape[1] + + C_shape = A.shape[:-1] + (N,) + C = A.new_empty(C_shape, dtype=output_dtype) + + def grid(META): + return ( + triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]), + ) + + if A.dtype == torch.float8_e4m3fn: + kernel = _w8a8_block_fp8_matmul + else: + raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.") + + kernel[grid]( + A, + B, + C, + As, + Bs, + M, + N, + K, + block_n, + block_k, + A.stride(-2), + A.stride(-1), + B.stride(1), + B.stride(0), + C.stride(-2), + C.stride(-1), + As.stride(-2), + As.stride(-1), + Bs.stride(1), + Bs.stride(0), + **config, + ) + + return C + + +def get_configs_compute_bound(): + configs = [] + for num_stages in [2, 3, 4, 5]: + for block_m in [16, 32, 64, 128, 256]: + for block_k in [64, 128]: + for block_n in [32, 64, 128, 256]: + for num_warps in [4, 8]: + for group_size in [1, 16, 32, 64]: + configs.append( + { + "BLOCK_SIZE_M": block_m, + "BLOCK_SIZE_N": block_n, + "BLOCK_SIZE_K": block_k, + "GROUP_SIZE_M": group_size, + "num_warps": num_warps, + "num_stages": num_stages, + } + ) + return configs + + +def get_weight_shapes(tp_size): + # NOTE(HandH1998): The weight shapes only works for DeepSeek-V3. + # Modify them, if you tune for another different model. + # cannot TP + total = [ + (512 + 64, 7168), + ((128 + 64) * 128, 7168), + (128 * (128 + 128), 512), + (7168, 16384), + (7168, 18432), + ] + # N can TP + n_tp = [ + (18432 * 2, 7168), + ((128 + 64) * 128, 7168), + (128 * (128 + 128), 512), + (24576, 1536), + (12288, 7168), + (4096, 7168), + ] + # K can TP + k_tp = [(7168, 18432), (7168, 16384), (7168, 2048)] + + weight_shapes = [] + for t in total: + weight_shapes.append(t) + for n_t in n_tp: + new_t = (n_t[0] // tp_size, n_t[1]) + weight_shapes.append(new_t) + for k_t in k_tp: + new_t = (k_t[0], k_t[1] // tp_size) + weight_shapes.append(new_t) + return weight_shapes + + +def benchmark_config( + A, B, As, Bs, block_size, config, out_dtype=torch.float16, num_iters=10 +): + def run(): + w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype) + + torch.cuda.synchronize() + # JIT complication & warmup + for _ in range(5): + run() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + torch.cuda.synchronize() + start_event.record() + run() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + return avg + + +def tune(M, N, K, block_size, out_dtype, search_space, input_type): + factor_for_scale = 1e-2 + + if input_type == "fp8": + fp8_info = torch.finfo(torch.float8_e4m3fn) + fp8_max, fp8_min = fp8_info.max, fp8_info.min + + A_fp32 = ( + (torch.rand(M, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max + ) + A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) + + B_fp32 = ( + (torch.rand(N, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max + ) + B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) + else: + raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.") + + block_n, block_k = block_size[0], block_size[1] + n_tiles = (N + block_n - 1) // block_n + k_tiles = (K + block_k - 1) // block_k + + As = torch.rand(M, k_tiles, dtype=torch.float32, device="cuda") * factor_for_scale + Bs = ( + torch.rand(n_tiles, k_tiles, dtype=torch.float32, device="cuda") + * factor_for_scale + ) + + best_config = None + best_time = float("inf") + for config in tqdm(search_space): + try: + kernel_time = benchmark_config( + A, + B, + As, + Bs, + block_size, + config, + out_dtype, + num_iters=10, + ) + except triton.runtime.autotuner.OutOfResources: + # Some configurations may be invalid and fail to compile. + continue + + if kernel_time < best_time: + best_time = kernel_time + best_config = config + now = datetime.now() + print(f"{now.ctime()}] Completed tuning for batch_size={M}") + assert best_config is not None + return best_config + + +def save_configs( + N, + K, + block_n, + block_k, + configs, + save_path, + input_type="fp8", +) -> None: + os.makedirs(save_path, exist_ok=True) + device_name = current_platform.get_device_name().replace(" ", "_") + json_file_name = ( + f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8," + f"block_shape=[{block_n},{block_k}].json" + ) + + config_file_path = os.path.join(save_path, json_file_name) + print(f"Writing best config to {config_file_path}...") + + with open(config_file_path, "w") as f: + json.dump(configs, f, indent=4) + f.write("\n") + + +def tune_on_gpu(args_dict): + """Run tuning on a specific GPU.""" + gpu_id = args_dict["gpu_id"] + batch_sizes = args_dict["batch_sizes"] + weight_shapes = args_dict["weight_shapes"] + args = args_dict["args"] + + torch.cuda.set_device(gpu_id) + print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}") + + block_n = args.block_n + block_k = args.block_k + out_dtype = DTYPE_MAP[args.out_dtype] + save_path = args.save_path + input_type = args.input_type + + search_space = get_configs_compute_bound() + search_space = [ + config for config in search_space if block_k % config["BLOCK_SIZE_K"] == 0 + ] + + start = time.time() + for shape in tqdm(weight_shapes, desc=f"GPU {gpu_id} - Shapes"): + N, K = shape[0], shape[1] + print(f"[GPU {gpu_id}] Tune for weight shape of `N: {N}, K: {K}`") + benchmark_results = [ + tune( + batch_size, + N, + K, + [block_n, block_k], + out_dtype, + search_space, + input_type, + ) + for batch_size in tqdm(batch_sizes, desc=f"GPU {gpu_id} - Batch sizes") + ] + best_configs = {M: config for M, config in zip(batch_sizes, benchmark_results)} + save_configs(N, K, block_n, block_k, best_configs, save_path, input_type) + + end = time.time() + print(f"Tuning on GPU {gpu_id} took {end - start:.2f} seconds") + + +def distribute_batch_sizes(batch_sizes, num_gpus): + """Distribute batch sizes across available GPUs.""" + batches_per_gpu = [] + for i in range(num_gpus): + start_idx = i * len(batch_sizes) // num_gpus + end_idx = (i + 1) * len(batch_sizes) // num_gpus + batches_per_gpu.append(batch_sizes[start_idx:end_idx]) + return batches_per_gpu + + +def main(args): + print(args) + num_gpus = torch.cuda.device_count() + if num_gpus == 0: + raise RuntimeError("No GPU available for tuning") + print(f"Found {num_gpus} GPUs for parallel tuning") + + torch.cuda.init() + + if args.batch_size is None: + batch_sizes = [ + 1, + 2, + 4, + 8, + 16, + 24, + 32, + 48, + 64, + 96, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ] + else: + batch_sizes = [args.batch_size] + num_gpus = 1 # If only one batch size, use only one GPU + + weight_shapes = get_weight_shapes(args.tp_size) + + batches_per_gpu = distribute_batch_sizes(batch_sizes, num_gpus) + + process_args = [] + for gpu_id in range(num_gpus): + process_args.append( + { + "gpu_id": gpu_id, + "batch_sizes": batches_per_gpu[gpu_id], + "weight_shapes": weight_shapes, # Each GPU processes all weight shapes + "args": args, + } + ) + + ctx = mp.get_context("spawn") + with ctx.Pool(num_gpus) as pool: + pool.map(tune_on_gpu, process_args) + + print("Multi-GPU tuning completed") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description=""" +Tune triton w8a8 block fp8 for DeepSeek-V3/DeepSeek-R1: + python3 benchmark_w8a8_block_fp8.py --tp-size 8 --input-type fp8 +Then copy to model_executor/layers/quantization/utils/configs + """, + formatter_class=argparse.RawTextHelpFormatter, + ) + + parser.add_argument("--tp-size", "-tp", type=int, default=8) + parser.add_argument("--input-type", type=str, choices=["fp8"], default="fp8") + parser.add_argument( + "--out-dtype", + type=str, + choices=["float32", "float16", "bfloat16", "half"], + default="float16", + ) + parser.add_argument("--block-n", type=int, default=128) + parser.add_argument("--block-k", type=int, default=128) + parser.add_argument("--batch-size", type=int, required=False) + parser.add_argument("--save-path", type=str, default="./") + args = parser.parse_args() + + main(args) diff --git a/offline_benchmark_test/benchmarks/kernels/deepgemm/README.md b/offline_benchmark_test/benchmarks/kernels/deepgemm/README.md new file mode 100644 index 0000000000000000000000000000000000000000..917e814010f894cd812e42849feb5822ad332d04 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/deepgemm/README.md @@ -0,0 +1,129 @@ +# DeepSeek DeepGEMM Kernels Benchmark + +This directory includes benchmarks between DeepSeek's DeepGEMM block fp8 kernels against vLLM's existing triton and CUTLASS-based kernels. + +Currently this just includes dense GEMMs and only works on Hopper GPUs. + +## Setup + +You need to install vLLM in your usual fashion, then install DeepGEMM from source in its own directory: + +``` +git clone --recursive https://github.com/deepseek-ai/DeepGEMM +cd DeepGEMM +python setup.py install +uv pip install -e . +``` + +## Usage + +``` +python benchmark_fp8_block_dense_gemm.py +INFO 02-26 21:55:13 [__init__.py:207] Automatically detected platform cuda. +===== STARTING FP8 GEMM BENCHMARK ===== +PyTorch version: 2.5.1+cu124 +CUDA version: 12.4 +Triton version: 3.1.0 +Using device: NVIDIA H100 80GB HBM3 +WARNING 02-26 21:55:15 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +INFO 02-26 21:55:15 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel. +WARNING 02-26 21:55:16 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=18432,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +WARNING 02-26 21:55:17 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel. +INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel. + +===== PERFORMANCE COMPARISON ===== + +DeepGEMM Implementation: ++------+-------+-------+-----------+--------+--------+ +| m | n | k | Time (μs) | TFLOPS | GB/s | ++------+-------+-------+-----------+--------+--------+ +| 8 | 4096 | 7168 | 102.9 | 4.6 | 286.4 | +| 8 | 7168 | 18432 | 70.8 | 29.8 | 1868.8 | +| 8 | 18432 | 7168 | 69.3 | 30.5 | 1911.8 | +| 64 | 4096 | 7168 | 69.1 | 54.4 | 439.0 | +| 64 | 7168 | 18432 | 69.4 | 243.6 | 1933.6 | +| 64 | 18432 | 7168 | 70.4 | 240.3 | 1917.2 | +| 64 | 24576 | 1536 | 70.1 | 68.9 | 584.6 | +| 64 | 32768 | 512 | 68.4 | 31.4 | 307.1 | +| 64 | 7168 | 16384 | 69.5 | 216.3 | 1718.5 | +| 128 | 4096 | 7168 | 141.1 | 53.3 | 222.1 | +| 128 | 7168 | 18432 | 71.9 | 470.5 | 1896.1 | +| 128 | 18432 | 7168 | 69.3 | 488.2 | 1988.2 | +| 1024 | 4096 | 7168 | 89.7 | 670.1 | 502.5 | +| 1024 | 18432 | 7168 | 279.0 | 969.8 | 635.2 | +| 2048 | 4096 | 7168 | 175.1 | 687.0 | 347.4 | +| 4096 | 4096 | 7168 | 335.4 | 717.0 | 275.1 | ++------+-------+-------+-----------+--------+--------+ + +vLLM Triton Implementation: ++------+-------+-------+-----------+--------+--------+--------------+ +| m | n | k | Time (μs) | TFLOPS | GB/s | vs DeepGEMM | ++------+-------+-------+-----------+--------+--------+--------------+ +| 8 | 4096 | 7168 | 74.0 | 6.3 | 398.2 | 1.39x faster | +| 8 | 7168 | 18432 | 89.6 | 23.6 | 1478.1 | 0.79x slower | +| 8 | 18432 | 7168 | 113.2 | 18.7 | 1170.4 | 0.61x slower | +| 64 | 4096 | 7168 | 79.4 | 47.3 | 382.2 | 0.87x slower | +| 64 | 7168 | 18432 | 98.5 | 171.7 | 1363.0 | 0.70x slower | +| 64 | 18432 | 7168 | 119.5 | 141.5 | 1129.4 | 0.59x slower | +| 64 | 24576 | 1536 | 37.6 | 128.4 | 1089.7 | 1.86x faster | +| 64 | 32768 | 512 | 38.7 | 55.5 | 542.6 | 1.77x faster | +| 64 | 7168 | 16384 | 86.1 | 174.5 | 1386.4 | 0.81x slower | +| 128 | 4096 | 7168 | 90.7 | 82.9 | 345.4 | 1.56x faster | +| 128 | 7168 | 18432 | 144.0 | 234.9 | 946.9 | 0.50x slower | +| 128 | 18432 | 7168 | 229.5 | 147.4 | 600.1 | 0.30x slower | +| 1024 | 4096 | 7168 | 242.3 | 248.2 | 186.1 | 0.37x slower | +| 1024 | 18432 | 7168 | 897.8 | 301.4 | 197.4 | 0.31x slower | +| 2048 | 4096 | 7168 | 463.0 | 259.7 | 131.4 | 0.38x slower | +| 4096 | 4096 | 7168 | 901.8 | 266.7 | 102.3 | 0.37x slower | ++------+-------+-------+-----------+--------+--------+--------------+ + +vLLM CUTLASS Implementation: ++------+-------+-------+-----------+--------+--------+--------------+--------------+ +| m | n | k | Time (μs) | TFLOPS | GB/s | vs DeepGEMM | vs Triton | ++------+-------+-------+-----------+--------+--------+--------------+--------------+ +| 8 | 4096 | 7168 | 34.6 | 13.6 | 852.3 | 2.98x faster | 2.14x faster | +| 8 | 7168 | 18432 | 78.9 | 26.8 | 1677.3 | 0.90x slower | 1.13x faster | +| 8 | 18432 | 7168 | 81.2 | 26.0 | 1631.1 | 0.85x slower | 1.39x faster | +| 64 | 4096 | 7168 | 36.9 | 101.9 | 822.9 | 1.87x faster | 2.15x faster | +| 64 | 7168 | 18432 | 87.4 | 193.4 | 1535.2 | 0.79x slower | 1.13x faster | +| 64 | 18432 | 7168 | 85.0 | 199.0 | 1587.6 | 0.83x slower | 1.41x faster | +| 64 | 24576 | 1536 | 28.0 | 172.8 | 1465.8 | 2.51x faster | 1.35x faster | +| 64 | 32768 | 512 | 28.8 | 74.5 | 728.5 | 2.37x faster | 1.34x faster | +| 64 | 7168 | 16384 | 77.9 | 193.0 | 1532.8 | 0.89x slower | 1.11x faster | +| 128 | 4096 | 7168 | 39.1 | 192.4 | 802.0 | 3.61x faster | 2.32x faster | +| 128 | 7168 | 18432 | 93.7 | 360.8 | 1454.2 | 0.77x slower | 1.54x faster | +| 128 | 18432 | 7168 | 85.7 | 394.8 | 1608.0 | 0.81x slower | 2.68x faster | +| 1024 | 4096 | 7168 | 99.7 | 603.1 | 452.2 | 0.90x slower | 2.43x faster | +| 1024 | 18432 | 7168 | 331.3 | 816.7 | 534.9 | 0.84x slower | 2.71x faster | +| 2048 | 4096 | 7168 | 198.3 | 606.6 | 306.7 | 0.88x slower | 2.34x faster | +| 4096 | 4096 | 7168 | 392.2 | 613.2 | 235.3 | 0.86x slower | 2.30x faster | ++------+-------+-------+-----------+--------+--------+--------------+--------------+ + +===== AVERAGE PERFORMANCE ===== ++----------------+------------+----------+---------------+ +| Implementation | Avg TFLOPS | Avg GB/s | Avg Time (ms) | ++----------------+------------+----------+---------------+ +| DeepGEMM | 310.98 | 1052.10 | 0.11 | +| vLLM Triton | 144.30 | 715.60 | 0.23 | +| vLLM CUTLASS | 286.78 | 1076.67 | 0.11 | ++----------------+------------+----------+---------------+ + +===== AVERAGE SPEEDUPS ===== ++-----------------------------+--------------+ +| Comparison | Speedup | ++-----------------------------+--------------+ +| DeepGEMM vs vLLM Triton | 1.71x faster | +| DeepGEMM vs vLLM CUTLASS | 0.94x slower | +| vLLM CUTLASS vs vLLM Triton | 1.84x faster | ++-----------------------------+--------------+ + +===== ACCURACY COMPARISON ===== ++----------------+-----------------------+ +| Implementation | Avg Diff vs Reference | ++----------------+-----------------------+ +| DeepGEMM | 0.000684 | +| vLLM Triton | 0.000684 | +| vLLM CUTLASS | 0.000684 | ++----------------+-----------------------+ +``` diff --git a/offline_benchmark_test/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py b/offline_benchmark_test/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py new file mode 100644 index 0000000000000000000000000000000000000000..43c54d56ca8c1be027bec0d5f54dd3e53a294771 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py @@ -0,0 +1,456 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +# fmt: off +# ruff: noqa: E501 +import time + +# Import DeepGEMM functions +import deep_gemm +import torch +from deep_gemm import calc_diff, ceil_div, get_col_major_tma_aligned_tensor + +# Import vLLM functions +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + per_token_group_quant_fp8, + w8a8_block_fp8_matmul, +) +from vllm.triton_utils import triton + + +# Copied from +# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L9 +def per_token_cast_to_fp8( + x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """Convert tensor to FP8 format with per-token scaling.""" + assert x.dim() == 2 and x.size(1) % 128 == 0 + m, n = x.shape + x_view = x.view(m, -1, 128) + x_amax = x_view.abs().float().amax(dim=2).view(m, -1).clamp(1e-4) + return (x_view * (448.0 / x_amax.unsqueeze(2))).to( + torch.float8_e4m3fn).view(m, n), (x_amax / 448.0).view(m, -1) + + +# Copied from +# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L17 +def per_block_cast_to_fp8( + x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """Convert tensor to FP8 format with per-block scaling.""" + assert x.dim() == 2 + m, n = x.shape + x_padded = torch.zeros((ceil_div(m, 128) * 128, ceil_div(n, 128) * 128), + dtype=x.dtype, + device=x.device) + x_padded[:m, :n] = x + x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128) + x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4) + x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn) + return x_scaled.view_as(x_padded)[:m, :n].contiguous(), ( + x_amax / 448.0).view(x_view.size(0), x_view.size(2)) + + +def benchmark_shape(m: int, + n: int, + k: int, + warmup: int = 100, + repeat: int = 10000, + verbose: bool = False) -> dict: + """Benchmark all implementations for a specific (m, n, k) shape.""" + if verbose: + print(f"\n=== Benchmarking shape: m={m}, n={n}, k={k} ===") + + # Create test tensors + A = torch.randn((m, k), device='cuda', dtype=torch.bfloat16) + B = torch.randn((n, k), device='cuda', dtype=torch.bfloat16) + + # Reference result in BF16 + torch.cuda.synchronize() + C_ref = A @ B.t() + + # Pre-quantize B for all implementations + # (weights can be pre-quantized offline) + B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B) + B_vllm, B_scale_vllm = per_block_cast_to_fp8(B) + + # Block size configuration + block_size = [128, 128] + + # Pre-quantize A for all implementations + A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A) + A_scale_deepgemm = get_col_major_tma_aligned_tensor(A_scale_deepgemm) + C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16) + A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1]) + A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8( + A, block_size[1], column_major_scales=True) + + # === DeepGEMM Implementation === + def deepgemm_gemm(): + deep_gemm.gemm_fp8_fp8_bf16_nt((A_deepgemm, A_scale_deepgemm), + (B_deepgemm, B_scale_deepgemm), + C_deepgemm) + return C_deepgemm + + # === vLLM Triton Implementation === + def vllm_triton_gemm(): + return w8a8_block_fp8_matmul(A_vllm, + B_vllm, + A_scale_vllm, + B_scale_vllm, + block_size, + output_dtype=torch.bfloat16) + + # === vLLM CUTLASS Implementation === + def vllm_cutlass_gemm(): + return ops.cutlass_scaled_mm(A_vllm_cutlass, + B_vllm.T, + scale_a=A_scale_vllm_cutlass, + scale_b=B_scale_vllm.T, + out_dtype=torch.bfloat16) + + # Run correctness check first + if verbose: + print("Running correctness check...") + C_deepgemm = deepgemm_gemm() + C_vllm_triton = vllm_triton_gemm() + C_vllm_cutlass = vllm_cutlass_gemm() + + deepgemm_diff = calc_diff(C_deepgemm, C_ref) + vllm_triton_diff = calc_diff(C_vllm_triton, C_ref) + vllm_cutlass_diff = calc_diff(C_vllm_cutlass, C_ref) + + if verbose: + print(f"DeepGEMM vs Reference difference: {deepgemm_diff:.6f}") + print(f"vLLM Triton vs Reference difference: {vllm_triton_diff:.6f}") + print(f"vLLM CUTLASS vs Reference difference: {vllm_cutlass_diff:.6f}") + print("vLLM Triton vs DeepGEMM difference: " + f"{calc_diff(C_vllm_triton, C_deepgemm):.6f}") + print("vLLM CUTLASS vs DeepGEMM difference: " + f"{calc_diff(C_vllm_cutlass, C_deepgemm):.6f}") + + # Benchmark implementations + implementations = { + "DeepGEMM": deepgemm_gemm, + "vLLM Triton": vllm_triton_gemm, + "vLLM CUTLASS": vllm_cutlass_gemm + } + + benchmark_results = { + "shape": { + "m": m, + "n": n, + "k": k + }, + "implementations": {} + } + + for name, func in implementations.items(): + # Warmup + for _ in range(warmup): + func() + torch.cuda.synchronize() + + # Timing loop + torch.cuda.synchronize() + start = time.time() + for _ in range(repeat): + func() + torch.cuda.synchronize() + end = time.time() + + # Calculate timing and TFLOPS + avg_time_ms = (end - start) / repeat * 1000 + avg_time_us = avg_time_ms * 1000 + tflops = 2 * m * n * k / (avg_time_ms * 1e-3) / 1e12 + gb_s = (m * k + k * n + m * n * 2) / 1e9 / (avg_time_ms * 1e-3) + + benchmark_results["implementations"][name] = { + "time_ms": avg_time_ms, + "time_us": avg_time_us, + "tflops": tflops, + "gb_s": gb_s, + "diff": { + "DeepGEMM": + 0.0 if name == "DeepGEMM" else calc_diff(func(), C_deepgemm), + "Reference": + deepgemm_diff if name == "DeepGEMM" else + (vllm_triton_diff + if name == "vLLM Triton" else vllm_cutlass_diff) + } + } + + if verbose: + print( + f"{name}: {avg_time_ms:.3f} ms, {tflops:.2f} TFLOPS, {gb_s:.2f} GB/s" + ) + + # Calculate speedups + baseline = benchmark_results["implementations"]["DeepGEMM"]["time_ms"] + for name, data in benchmark_results["implementations"].items(): + if name != "DeepGEMM": + speedup = baseline / data["time_ms"] + benchmark_results["implementations"][name][ + "speedup_vs_deepgemm"] = speedup + if verbose: + print(f"DeepGEMM is {1/speedup:.2f}x " + f"{'faster' if 1/speedup > 1 else 'slower'} than {name}") + + vllm_triton_time = benchmark_results["implementations"]["vLLM Triton"][ + "time_ms"] + vllm_cutlass_time = benchmark_results["implementations"]["vLLM CUTLASS"][ + "time_ms"] + cutlass_vs_triton = vllm_triton_time / vllm_cutlass_time + benchmark_results["implementations"]["vLLM CUTLASS"][ + "speedup_vs_triton"] = cutlass_vs_triton + if verbose: + print( + f"vLLM CUTLASS is {cutlass_vs_triton:.2f}x " + f"{'faster' if cutlass_vs_triton > 1 else 'slower'} than vLLM Triton" + ) + + return benchmark_results + + +def format_table_row(values, widths): + """Format a row with specified column widths.""" + return "| " + " | ".join(f"{val:{w}}" + for val, w in zip(values, widths)) + " |" + + +def print_table(headers, rows, title=None): + """Print a table with headers and rows.""" + if title: + print(f"\n{title}") + + # Calculate column widths based on headers and data + widths = [ + max(len(str(h)), max(len(str(row[i])) for row in rows)) + for i, h in enumerate(headers) + ] + + # Create separator line + separator = "+-" + "-+-".join("-" * w for w in widths) + "-+" + + # Print table + print(separator) + print(format_table_row(headers, widths)) + print(separator) + for row in rows: + print(format_table_row(row, widths)) + print(separator) + + +def format_speedup(value): + """Format speedup value with indicator if it's faster or slower.""" + return f"{value:.2f}x {'faster' if value > 1.0 else 'slower'}" + + +def run_benchmarks(verbose: bool = False): + """Run benchmarks for a set of common shapes.""" + print("===== STARTING FP8 GEMM BENCHMARK =====") + + # Make sure we're using the GPU + if not torch.cuda.is_available(): + print("CUDA not available! Tests require GPU.") + return + + # Print system information + print(f"PyTorch version: {torch.__version__}") + print(f"CUDA version: {torch.version.cuda}") + print(f"Triton version: {triton.__version__}") + print(f"Using device: {torch.cuda.get_device_name()}") + + # Enable TF32 for better performance + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + + # Set seeds for reproducibility + torch.manual_seed(42) + torch.cuda.manual_seed(42) + + # Define benchmark shapes (m, n, k) + shapes = [ + (8, 4096, 7168), + (8, 7168, 18432), + (8, 18432, 7168), + (64, 4096, 7168), + (64, 7168, 18432), + (64, 18432, 7168), + (64, 24576, 1536), + (64, 32768, 512), + (64, 7168, 16384), + (128, 4096, 7168), + (128, 7168, 18432), + (128, 18432, 7168), + (1024, 4096, 7168), + (1024, 18432, 7168), + (2048, 4096, 7168), + (4096, 4096, 7168), + ] + shapes = [ + # (64, 2112, 7168), + (64, 24576, 1536), + (64, 32768, 512), + (64, 7168, 16384), + (64, 4096, 7168), + (64, 7168, 2048), + # (128, 2112, 7168), + (128, 24576, 1536), + (128, 32768, 512), + (128, 7168, 16384), + (128, 4096, 7168), + (128, 7168, 2048), + # (4096, 2112, 7168), + (4096, 24576, 1536), + (4096, 32768, 512), + (4096, 7168, 16384), + (4096, 4096, 7168), + (4096, 7168, 2048), + ] + + all_results = [] + for m, n, k in shapes: + result = benchmark_shape(m, n, k, verbose=verbose) + all_results.append(result) + + # Print results in a nicely formatted table + print("\n===== PERFORMANCE COMPARISON =====") + + # Print DeepGEMM table + deepgemm_headers = ["m", "n", "k", "Time (μs)", "TFLOPS", "GB/s"] + deepgemm_rows = [] + for result in all_results: + shape = result["shape"] + impl_data = result["implementations"]["DeepGEMM"] + deepgemm_rows.append([ + shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}", + f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}" + ]) + + print_table(deepgemm_headers, + deepgemm_rows, + title="DeepGEMM Implementation:") + + # Print vLLM Triton table + triton_headers = [ + "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM" + ] + triton_rows = [] + for result in all_results: + shape = result["shape"] + impl_data = result["implementations"]["vLLM Triton"] + speedup = impl_data.get("speedup_vs_deepgemm", 1.0) + triton_rows.append([ + shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}", + f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}", + format_speedup(speedup) + ]) + + print_table(triton_headers, + triton_rows, + title="vLLM Triton Implementation:") + + # Print vLLM CUTLASS table + cutlass_headers = [ + "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM", + "vs Triton" + ] + cutlass_rows = [] + for result in all_results: + shape = result["shape"] + impl_data = result["implementations"]["vLLM CUTLASS"] + vs_deepgemm = impl_data.get("speedup_vs_deepgemm", 1.0) + vs_triton = impl_data.get("speedup_vs_triton", 1.0) + cutlass_rows.append([ + shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}", + f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}", + format_speedup(vs_deepgemm), + format_speedup(vs_triton) + ]) + + print_table(cutlass_headers, + cutlass_rows, + title="vLLM CUTLASS Implementation:") + + # Calculate and print averages + print("\n===== AVERAGE PERFORMANCE =====") + + implementations = ["DeepGEMM", "vLLM Triton", "vLLM CUTLASS"] + avg_metrics = { + impl: { + "tflops": 0, + "gb_s": 0, + "time_ms": 0 + } + for impl in implementations + } + + for result in all_results: + for impl in implementations: + impl_data = result["implementations"][impl] + avg_metrics[impl]["tflops"] += impl_data["tflops"] + avg_metrics[impl]["gb_s"] += impl_data["gb_s"] + avg_metrics[impl]["time_ms"] += impl_data["time_ms"] + + num_shapes = len(all_results) + avg_headers = ["Implementation", "Avg TFLOPS", "Avg GB/s", "Avg Time (ms)"] + avg_rows = [] + + for impl in implementations: + avg_tflops = avg_metrics[impl]["tflops"] / num_shapes + avg_mem_bw = avg_metrics[impl]["gb_s"] / num_shapes + avg_time = avg_metrics[impl]["time_ms"] / num_shapes + avg_rows.append([ + impl, f"{avg_tflops:.2f}", f"{avg_mem_bw:.2f}", f"{avg_time:.2f}" + ]) + + print_table(avg_headers, avg_rows) + + # Calculate average speedups + avg_speedups = { + "DeepGEMM vs vLLM Triton": 0, + "DeepGEMM vs vLLM CUTLASS": 0, + "vLLM CUTLASS vs vLLM Triton": 0 + } + + for result in all_results: + deepgemm_time = result["implementations"]["DeepGEMM"]["time_ms"] + vllm_triton_time = result["implementations"]["vLLM Triton"]["time_ms"] + vllm_cutlass_time = result["implementations"]["vLLM CUTLASS"][ + "time_ms"] + + avg_speedups[ + "DeepGEMM vs vLLM Triton"] += vllm_triton_time / deepgemm_time + avg_speedups[ + "DeepGEMM vs vLLM CUTLASS"] += vllm_cutlass_time / deepgemm_time + avg_speedups[ + "vLLM CUTLASS vs vLLM Triton"] += vllm_triton_time / vllm_cutlass_time + + print("\n===== AVERAGE SPEEDUPS =====") + speedup_headers = ["Comparison", "Speedup"] + speedup_rows = [] + for comparison, total in avg_speedups.items(): + avg_speedup = total / num_shapes + status = "faster" if avg_speedup > 1 else "slower" + speedup_rows.append([comparison, f"{avg_speedup:.2f}x {status}"]) + + print_table(speedup_headers, speedup_rows) + + # Average accuracy comparison + print("\n===== ACCURACY COMPARISON =====") + avg_diff = {impl: 0 for impl in implementations} + + for result in all_results: + for impl in implementations: + avg_diff[impl] += result["implementations"][impl]["diff"][ + "Reference"] + + diff_headers = ["Implementation", "Avg Diff vs Reference"] + diff_rows = [] + for impl in implementations: + diff_rows.append([impl, f"{avg_diff[impl] / num_shapes:.6f}"]) + + print_table(diff_headers, diff_rows) + + +if __name__ == "__main__": + run_benchmarks(verbose=False) diff --git a/offline_benchmark_test/benchmarks/kernels/graph_machete_bench.py b/offline_benchmark_test/benchmarks/kernels/graph_machete_bench.py new file mode 100644 index 0000000000000000000000000000000000000000..9a4da0ef5a85d0df2178d5e990ad177e3f76bf37 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/graph_machete_bench.py @@ -0,0 +1,64 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import math +import pickle +from collections import defaultdict + +import matplotlib.pyplot as plt +import pandas as pd +import regex as re +import seaborn as sns +from torch.utils.benchmark import Measurement as TMeasurement + +from vllm.utils import FlexibleArgumentParser + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark the latency of processing a single batch of " + "requests till completion." + ) + parser.add_argument("filename", type=str) + + args = parser.parse_args() + + with open(args.filename, "rb") as f: + data = pickle.load(f) + raw_results: list[TMeasurement] = data["results"] + + results = defaultdict(lambda: list()) + for v in raw_results: + result = re.search(r"MKN=\(\d+x(\d+x\d+)\)", v.task_spec.sub_label) + if result is not None: + KN = result.group(1) + else: + raise Exception("MKN not found") + result = re.search(r"MKN=\((\d+)x\d+x\d+\)", v.task_spec.sub_label) + if result is not None: + M = result.group(1) + else: + raise Exception("MKN not found") + + kernel = v.task_spec.description + results[KN].append({"kernel": kernel, "batch_size": M, "median": v.median}) + + rows = int(math.ceil(len(results) / 2)) + fig, axs = plt.subplots(rows, 2, figsize=(12, 5 * rows)) + axs = axs.flatten() + for axs_idx, (shape, data) in enumerate(results.items()): + plt.sca(axs[axs_idx]) + df = pd.DataFrame(data) + sns.lineplot( + data=df, + x="batch_size", + y="median", + hue="kernel", + style="kernel", + markers=True, + dashes=False, + palette="Dark2", + ) + plt.title(f"Shape: {shape}") + plt.ylabel("time (median, s)") + plt.tight_layout() + plt.savefig("graph_machete_bench.pdf") diff --git a/offline_benchmark_test/benchmarks/kernels/requirements.txt b/offline_benchmark_test/benchmarks/kernels/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..1411a4a0b5ab886adfb744e685d150151ab10023 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/requirements.txt @@ -0,0 +1 @@ +pandas \ No newline at end of file diff --git a/offline_benchmark_test/benchmarks/kernels/utils.py b/offline_benchmark_test/benchmarks/kernels/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..4bbb36bb4359259944310adb8e1047a02f126dea --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/utils.py @@ -0,0 +1,214 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import dataclasses +from collections.abc import Iterable +from typing import Any, Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement + + +@dataclasses.dataclass +class CudaGraphBenchParams: + num_ops_in_cuda_graph: int + + +@dataclasses.dataclass +class ArgPool: + """ + When some argument of the benchmarking function is annotated with this type, + the benchmarking class (BenchMM) will collapse the argument to a pick a + single value from the given list of values, during function invocation. + For every invocation during a benchmarking run, it will choose a + different value from the list. + """ + + values: Iterable[Any] + + def __getitem__(self, index): + return self.values[index] + + +class Bench: + class ArgsIterator: + def __init__(self, args_list, kwargs_list): + assert len(args_list) == len(kwargs_list) + self.args_list = args_list + self.kwargs_list = kwargs_list + self.n = len(self.args_list) + self.idx = 0 + + def __next__(self): + while True: + yield (self.args_list[self.idx], self.kwargs_list[self.idx]) + self.idx += 1 + self.idx = self.idx % self.n + + def reset(self): + self.idx = 0 + + @property + def n_args(self): + return self.n + + def __init__( + self, + cuda_graph_params: Optional[CudaGraphBenchParams], + label: str, + sub_label: str, + description: str, + fn: Callable, + *args, + **kwargs, + ): + self.cuda_graph_params = cuda_graph_params + self.use_cuda_graph = self.cuda_graph_params is not None + self.label = label + self.sub_label = sub_label + self.description = description + self.fn = fn + + # Process args + self._args = args + self._kwargs = kwargs + self.args_list, self.kwargs_list = self.collapse_argpool(*args, **kwargs) + self.args_iterator = self.ArgsIterator(self.args_list, self.kwargs_list) + + # Cudagraph runner + self.g = None + if self.use_cuda_graph: + self.g = self.get_cuda_graph_runner() + + # benchmark run params + self.min_run_time = 1 + + def collapse_argpool(self, *args, **kwargs): + argpool_args = [arg for arg in args if isinstance(arg, ArgPool)] + [ + arg for arg in kwargs.values() if isinstance(arg, ArgPool) + ] + if len(argpool_args) == 0: + return [args], [kwargs] + + # Make sure all argpools are of the same size + argpool_size = len(argpool_args[0].values) + assert all([argpool_size == len(arg.values) for arg in argpool_args]) + + # create copies of the args + args_list = [] + kwargs_list = [] + for _ in range(argpool_size): + args_list.append(args) + kwargs_list.append(kwargs.copy()) + + for i in range(argpool_size): + # collapse args; Just pick the ith value + args_list[i] = tuple( + [arg[i] if isinstance(arg, ArgPool) else arg for arg in args_list[i]] + ) + + # collapse kwargs + kwargs_i = kwargs_list[i] + arg_pool_keys = [k for k, v in kwargs_i.items() if isinstance(v, ArgPool)] + for k in arg_pool_keys: + # again just pick the ith value + kwargs_i[k] = kwargs_i[k][i] + kwargs_list[i] = kwargs_i + + return args_list, kwargs_list + + def get_cuda_graph_runner(self): + assert self.use_cuda_graph + assert self.args_iterator is not None + + num_graph_ops = self.cuda_graph_params.num_ops_in_cuda_graph + + # warmup + args_it = self.args_iterator.__next__() + for _ in range(2): + args, kwargs = next(args_it) + self.fn(*args, **kwargs) + + self.args_iterator.reset() + args_it = self.args_iterator.__next__() + stream = torch.cuda.Stream() + with torch.cuda.stream(stream): + g = torch.cuda.CUDAGraph() + with torch.cuda.graph(g): + for _ in range(num_graph_ops): + args, kwargs = next(args_it) + self.fn(*args, **kwargs) + return g + + def run_cudagrah(self) -> TMeasurement: + assert self.use_cuda_graph + globals = {"g": self.g} + + return TBenchmark.Timer( + stmt="g.replay()", + globals=globals, + label=( + f"{self.label}" + f" | cugraph {self.cuda_graph_params.num_ops_in_cuda_graph} ops" + ), + sub_label=self.sub_label, + description=self.description, + ).blocked_autorange(min_run_time=self.min_run_time) + + def run_eager(self) -> TMeasurement: + setup = None + stmt = None + globals = None + + has_arg_pool = self.args_iterator.n_args > 1 + if has_arg_pool: + setup = """ + args_iterator.reset() + args_it = args_iterator.__next__() + """ + stmt = """ + args, kwargs = next(args_it) + fn(*args, **kwargs) + """ + globals = {"fn": self.fn, "args_iterator": self.args_iterator} + else: + # no arg pool. Just use the args and kwargs directly + self.args_iterator.reset() + args_it = self.args_iterator.__next__() + args, kwargs = next(args_it) + + setup = "" + stmt = """ + fn(*args, **kwargs) + """ + globals = {"fn": self.fn, "args": args, "kwargs": kwargs} + + return TBenchmark.Timer( + stmt=stmt, + setup=setup, + globals=globals, + label=self.label, + sub_label=self.sub_label, + description=self.description, + ).blocked_autorange(min_run_time=self.min_run_time) + + def run(self) -> TMeasurement: + timer = None + if self.use_cuda_graph: # noqa SIM108 + timer = self.run_cudagrah() + else: + timer = self.run_eager() + if not timer.meets_confidence() or timer.has_warnings: + print("Doesn't meet confidence - re-running bench ...") + return self.run() + return timer + + def __enter__(self): + return self + + def __exit__(self, exc_type, exc_value, traceback): + if exc_type: + print(f"exc type {exc_type}") + print(f"exc value {exc_value}") + print(f"exc traceback {traceback}") diff --git a/offline_benchmark_test/benchmarks/kernels/weight_shapes.py b/offline_benchmark_test/benchmarks/kernels/weight_shapes.py new file mode 100644 index 0000000000000000000000000000000000000000..a27f02394afbdfa4f3a352a599db349a2601bea5 --- /dev/null +++ b/offline_benchmark_test/benchmarks/kernels/weight_shapes.py @@ -0,0 +1,98 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +# Weight Shapes are in the format +# ([K, N], TP_SPLIT_DIM) +# Example: +# A shape of ([14336, 4096], 0) indicates the following GEMM shape, +# - TP1 : K = 14336, N = 4096 +# - TP2 : K = 7168, N = 4096 +# A shape of ([4096, 6144], 1) indicates the following GEMM shape, +# - TP1 : K = 4096, N = 6144 +# - TP4 : K = 4096, N = 1536 + +# TP1 shapes +WEIGHT_SHAPES = { + "mistralai/Mistral-7B-v0.1": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-7b-hf": [ + ([4096, 12288], 1), + ([4096, 4096], 0), + ([4096, 22016], 1), + ([11008, 4096], 0), + ], + "meta-llama/Llama-3-8b": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-13b-hf": [ + ([5120, 15360], 1), + ([5120, 5120], 0), + ([5120, 27648], 1), + ([13824, 5120], 0), + ], + "meta-llama/Llama-2-70b-hf": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 57344], 1), + ([28672, 8192], 0), + ], + "meta-llama/Llama-3.1-405b-hf": [ + ([16384, 18432], 1), + ([16384, 16384], 0), + ([16384, 106496], 1), + ([53248, 16384], 0), + ], + "meta-llama/Llama-3.1-8B-Instruct": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-3.3-70B-Instruct": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 57344], 1), + ([28672, 8192], 0), + ], + "mistralai/Mistral-Large-Instruct-2407": [ + ([12288, 14336], 1), + ([12288, 12288], 0), + ([12288, 57344], 1), + ([28672, 12288], 0), + ], + "Qwen/Qwen2.5-7B-Instruct": [ + ([3584, 4608], 1), + ([3584, 3584], 0), + ([3584, 37888], 1), + ([18944, 3584], 0), + ], + "Qwen/Qwen2.5-32B-Instruct": [ + ([5120, 7168], 1), + ([5120, 5120], 0), + ([5120, 55296], 1), + ([27648, 5120], 0), + ], + "Qwen/Qwen2.5-72B-Instruct": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 59136], 1), + ([29568, 8192], 0), + ], + "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": [ + ([2048, 3072], 1), + ([2048, 4096], 1), + ([2048, 2048], 0), + ([2048, 576], 0), + ([2048, 21888], 1), + ([10944, 2048], 0), + ([2048, 2816], 1), + ([1408, 2048], 0), + ], +} diff --git a/offline_benchmark_test/benchmarks/overheads/benchmark_hashing.py b/offline_benchmark_test/benchmarks/overheads/benchmark_hashing.py new file mode 100644 index 0000000000000000000000000000000000000000..0957a9c65f06c912ae73588e017602af72ae9772 --- /dev/null +++ b/offline_benchmark_test/benchmarks/overheads/benchmark_hashing.py @@ -0,0 +1,64 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import cProfile +import pstats + +from vllm import LLM, SamplingParams +from vllm.utils import FlexibleArgumentParser + +# A very long prompt, total number of tokens is about 15k. +LONG_PROMPT = ["You are an expert in large language models, aren't you?"] * 1000 +LONG_PROMPT = " ".join(LONG_PROMPT) + + +def main(args): + llm = LLM( + model=args.model, + enforce_eager=True, + enable_prefix_caching=True, + tensor_parallel_size=args.tensor_parallel_size, + ) + + sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len) + profiler = cProfile.Profile() + + print("------warm up------") + for i in range(3): + output = llm.generate(LONG_PROMPT, sampling_params) + print(output[0].outputs[0].text) + + print("------start generating------") + for i in range(3): + profiler.runctx( + "llm.generate(LONG_PROMPT, sampling_params)", globals(), locals() + ) + + # analyze the runtime of hashing function + stats = pstats.Stats(profiler) + stats.sort_stats("cumulative") + total_time = 0 + total_calls = 0 + for func in stats.stats: + if "hash_of_block" in func[2]: + total_time = stats.stats[func][3] + total_calls = stats.stats[func][0] + percentage = (total_time / stats.total_tt) * 100 + print( + f"Hashing took {total_time:.2f} seconds,{percentage:.2f}% of the total runtime." + ) + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark the performance of hashing function in" + "automatic prefix caching." + ) + parser.add_argument("--model", type=str, default="lmsys/longchat-7b-16k") + parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1) + parser.add_argument("--output-len", type=int, default=10) + parser.add_argument( + "--enable-prefix-caching", action="store_true", help="enable prefix caching" + ) + args = parser.parse_args() + main(args) diff --git a/offline_benchmark_test/benchmarks/pyproject.toml b/offline_benchmark_test/benchmarks/pyproject.toml new file mode 100644 index 0000000000000000000000000000000000000000..65b1e09a247e2d74e31b616c38a8544de75ad525 --- /dev/null +++ b/offline_benchmark_test/benchmarks/pyproject.toml @@ -0,0 +1,49 @@ +# This local pyproject file is part of the migration from yapf to ruff format. +# It uses the same core rules as the main pyproject.toml file, but with the +# following differences: +# - ruff line length is overridden to 88 +# - deprecated typing ignores (UP006, UP035) have been removed + +[tool.ruff] +line-length = 88 + +[tool.ruff.lint.per-file-ignores] +"vllm/third_party/**" = ["ALL"] +"vllm/version.py" = ["F401"] +"vllm/_version.py" = ["ALL"] + +[tool.ruff.lint] +select = [ + # pycodestyle + "E", + # Pyflakes + "F", + # pyupgrade + "UP", + # flake8-bugbear + "B", + # flake8-simplify + "SIM", + # isort + "I", + # flake8-logging-format + "G", +] +ignore = [ + # star imports + "F405", "F403", + # lambda expression assignment + "E731", + # Loop control variable not used within loop body + "B007", + # f-string format + "UP032", + # Can remove once 3.10+ is the minimum Python version + "UP007", +] + +[tool.ruff.lint.isort] +known-first-party = ["vllm"] + +[tool.ruff.format] +docstring-code-format = true \ No newline at end of file diff --git a/offline_benchmark_test/benchmarks/run_structured_output_benchmark.sh b/offline_benchmark_test/benchmarks/run_structured_output_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..b043ab83e4608e7b27e2f94e6ec24a05f7c474aa --- /dev/null +++ b/offline_benchmark_test/benchmarks/run_structured_output_benchmark.sh @@ -0,0 +1,129 @@ +#!/bin/bash + +# default values +MODEL=${MODEL:-"Qwen/Qwen2.5-7B-Instruct"} +BACKEND=${BACKEND:-"vllm"} +DATASET=${DATASET:-"xgrammar_bench"} +SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" +OUTPUT_DIR=${OUTPUT_DIR:-"$SCRIPT_DIR/structured_output_benchmark_results"} +PORT=${PORT:-8000} +STRUCTURED_OUTPUT_RATIO=${STRUCTURED_OUTPUT_RATIO:-1} +TOTAL_SECONDS=${TOTAL_SECONDS:-90} +MAX_NEW_TOKENS=${MAX_NEW_TOKENS:-300} +TOKENIZER_MODE=${TOKENIZER_MODE:-"auto"} + +usage() { + echo "Usage: $0 [options]" + echo "Options:" + echo " --model MODEL Model to benchmark (default: $MODEL)" + echo " --backend BACKEND Backend to use (default: $BACKEND)" + echo " --dataset DATASET Dataset to use (default: $DATASET)" + echo " --max-new-tokens N Maximum number of tokens to generate (default: $MAX_NEW_TOKENS)" + echo " --output-dir DIR Output directory for results (default: $OUTPUT_DIR)" + echo " --port PORT Port to use (default: $PORT)" + echo " --structured-output-ratio N Ratio of structured outputs (default: $STRUCTURED_OUTPUT_RATIO)" + echo " --tokenizer-mode MODE Tokenizer mode to use (default: $TOKENIZER_MODE)" + echo " --total-seconds N Total seconds to run the benchmark (default: $TOTAL_SECONDS)" + echo " -h, --help Show this help message and exit" + exit 0 +} + +# parse command line arguments +while [[ $# -gt 0 ]]; do + case $1 in + --model) + MODEL="$2" + shift 2 + ;; + --backend) + BACKEND="$2" + shift 2 + ;; + --dataset) + DATASET="$2" + shift 2 + ;; + --max-new-tokens) + MAX_NEW_TOKENS="$2" + shift 2 + ;; + --output-dir) + OUTPUT_DIR="$2" + shift 2 + ;; + --port) + PORT="$2" + shift 2 + ;; + --structured-output-ratio) + STRUCTURED_OUTPUT_RATIO="$2" + shift 2 + ;; + --tokenizer-mode) + TOKENIZER_MODE="$2" + shift 2 + ;; + --total-seconds) + TOTAL_SECONDS="$2" + shift 2 + ;; + -h|--help) + usage + ;; + *) + echo "Unknown argument: $1\n" + usage + ;; + esac +done + +# Create output directory if it doesn't exist +mkdir -p "$OUTPUT_DIR" + +# Define QPS values to test +QPS_VALUES=(25 20 15 10 5 1) + +# Common parameters +COMMON_PARAMS="--backend $BACKEND \ + --model $MODEL \ + --dataset $DATASET \ + --structured-output-ratio $STRUCTURED_OUTPUT_RATIO \ + --save-results \ + --result-dir $OUTPUT_DIR \ + --output-len $MAX_NEW_TOKENS \ + --port $PORT \ + --tokenizer-mode $TOKENIZER_MODE" + +echo "Starting structured output benchmark with model: $MODEL" +echo "Backend: $BACKEND" +echo "Dataset: $DATASET" +echo "Results will be saved to: $OUTPUT_DIR" +echo "----------------------------------------" + +# Run benchmarks with different QPS values +for qps in "${QPS_VALUES[@]}"; do + echo "Running benchmark with QPS: $qps" + + # Get git hash and branch for the filename + GIT_HASH=$(git rev-parse --short HEAD 2>/dev/null || echo "unknown") + GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "unknown") + + # Construct filename for this run + FILENAME="${BACKEND}_${qps}qps_$(basename $MODEL)_${DATASET}_${GIT_HASH}.json" + + NUM_PROMPTS=$(echo "$TOTAL_SECONDS * $qps" | bc) + NUM_PROMPTS=${NUM_PROMPTS%.*} # Remove fractional part + echo "Running benchmark with $NUM_PROMPTS prompts" + + # Run the benchmark + python "$SCRIPT_DIR/benchmark_serving_structured_output.py" $COMMON_PARAMS \ + --request-rate $qps \ + --result-filename "$FILENAME" \ + --num-prompts $NUM_PROMPTS + + echo "Completed benchmark with QPS: $qps" + echo "----------------------------------------" +done + +echo "All benchmarks completed!" +echo "Results saved to: $OUTPUT_DIR" diff --git a/offline_benchmark_test/benchmarks/sonnet.txt b/offline_benchmark_test/benchmarks/sonnet.txt new file mode 100644 index 0000000000000000000000000000000000000000..34c444e8ce8e2dc701ec80931401c57014ae0bd1 --- /dev/null +++ b/offline_benchmark_test/benchmarks/sonnet.txt @@ -0,0 +1,518 @@ +FROM fairest creatures we desire increase, +That thereby beauty's rose might never die, +But as the riper should by time decease, +His tender heir might bear his memory: +But thou, contracted to thine own bright eyes, +Feed'st thy light'st flame with self-substantial fuel, +Making a famine where abundance lies, +Thyself thy foe, to thy sweet self too cruel. +Thou that art now the world's fresh ornament +And only herald to the gaudy spring, +Within thine own bud buriest thy content +And, tender churl, makest waste in niggarding. +Pity the world, or else this glutton be, +To eat the world's due, by the grave and thee. +When forty winters shall beseige thy brow, +And dig deep trenches in thy beauty's field, +Thy youth's proud livery, so gazed on now, +Will be a tatter'd weed, of small worth held: +Then being ask'd where all thy beauty lies, +Where all the treasure of thy lusty days, +To say, within thine own deep-sunken eyes, +Were an all-eating shame and thriftless praise. +How much more praise deserved thy beauty's use, +If thou couldst answer 'This fair child of mine +Shall sum my count and make my old excuse,' +Proving his beauty by succession thine! +This were to be new made when thou art old, +And see thy blood warm when thou feel'st it cold. +Look in thy glass, and tell the face thou viewest +Now is the time that face should form another; +Whose fresh repair if now thou not renewest, +Thou dost beguile the world, unbless some mother. +For where is she so fair whose unear'd womb +Disdains the tillage of thy husbandry? +Or who is he so fond will be the tomb +Of his self-love, to stop posterity? +Thou art thy mother's glass, and she in thee +Calls back the lovely April of her prime: +So thou through windows of thine age shall see +Despite of wrinkles this thy golden time. +But if thou live, remember'd not to be, +Die single, and thine image dies with thee. +Unthrifty loveliness, why dost thou spend +Upon thyself thy beauty's legacy? +Nature's bequest gives nothing but doth lend, +And being frank she lends to those are free. +Then, beauteous niggard, why dost thou abuse +The bounteous largess given thee to give? +Profitless usurer, why dost thou use +So great a sum of sums, yet canst not live? +For having traffic with thyself alone, +Thou of thyself thy sweet self dost deceive. +Then how, when nature calls thee to be gone, +What acceptable audit canst thou leave? +Thy unused beauty must be tomb'd with thee, +Which, used, lives th' executor to be. +Those hours, that with gentle work did frame +The lovely gaze where every eye doth dwell, +Will play the tyrants to the very same +And that unfair which fairly doth excel: +For never-resting time leads summer on +To hideous winter and confounds him there; +Sap cheque'd with frost and lusty leaves quite gone, +Beauty o'ersnow'd and bareness every where: +Then, were not summer's distillation left, +A liquid prisoner pent in walls of glass, +Beauty's effect with beauty were bereft, +Nor it nor no remembrance what it was: +But flowers distill'd though they with winter meet, +Leese but their show; their substance still lives sweet. +Then let not winter's ragged hand deface +In thee thy summer, ere thou be distill'd: +Make sweet some vial; treasure thou some place +With beauty's treasure, ere it be self-kill'd. +That use is not forbidden usury, +Which happies those that pay the willing loan; +That's for thyself to breed another thee, +Or ten times happier, be it ten for one; +Ten times thyself were happier than thou art, +If ten of thine ten times refigured thee: +Then what could death do, if thou shouldst depart, +Leaving thee living in posterity? +Be not self-will'd, for thou art much too fair +To be death's conquest and make worms thine heir. +Lo! in the orient when the gracious light +Lifts up his burning head, each under eye +Doth homage to his new-appearing sight, +Serving with looks his sacred majesty; +And having climb'd the steep-up heavenly hill, +Resembling strong youth in his middle age, +yet mortal looks adore his beauty still, +Attending on his golden pilgrimage; +But when from highmost pitch, with weary car, +Like feeble age, he reeleth from the day, +The eyes, 'fore duteous, now converted are +From his low tract and look another way: +So thou, thyself out-going in thy noon, +Unlook'd on diest, unless thou get a son. +Music to hear, why hear'st thou music sadly? +Sweets with sweets war not, joy delights in joy. +Why lovest thou that which thou receivest not gladly, +Or else receivest with pleasure thine annoy? +If the true concord of well-tuned sounds, +By unions married, do offend thine ear, +They do but sweetly chide thee, who confounds +In singleness the parts that thou shouldst bear. +Mark how one string, sweet husband to another, +Strikes each in each by mutual ordering, +Resembling sire and child and happy mother +Who all in one, one pleasing note do sing: +Whose speechless song, being many, seeming one, +Sings this to thee: 'thou single wilt prove none.' +Is it for fear to wet a widow's eye +That thou consumest thyself in single life? +Ah! if thou issueless shalt hap to die. +The world will wail thee, like a makeless wife; +The world will be thy widow and still weep +That thou no form of thee hast left behind, +When every private widow well may keep +By children's eyes her husband's shape in mind. +Look, what an unthrift in the world doth spend +Shifts but his place, for still the world enjoys it; +But beauty's waste hath in the world an end, +And kept unused, the user so destroys it. +No love toward others in that bosom sits +That on himself such murderous shame commits. +For shame! deny that thou bear'st love to any, +Who for thyself art so unprovident. +Grant, if thou wilt, thou art beloved of many, +But that thou none lovest is most evident; +For thou art so possess'd with murderous hate +That 'gainst thyself thou stick'st not to conspire. +Seeking that beauteous roof to ruinate +Which to repair should be thy chief desire. +O, change thy thought, that I may change my mind! +Shall hate be fairer lodged than gentle love? +Be, as thy presence is, gracious and kind, +Or to thyself at least kind-hearted prove: +Make thee another self, for love of me, +That beauty still may live in thine or thee. +As fast as thou shalt wane, so fast thou growest +In one of thine, from that which thou departest; +And that fresh blood which youngly thou bestowest +Thou mayst call thine when thou from youth convertest. +Herein lives wisdom, beauty and increase: +Without this, folly, age and cold decay: +If all were minded so, the times should cease +And threescore year would make the world away. +Let those whom Nature hath not made for store, +Harsh featureless and rude, barrenly perish: +Look, whom she best endow'd she gave the more; +Which bounteous gift thou shouldst in bounty cherish: +She carved thee for her seal, and meant thereby +Thou shouldst print more, not let that copy die. +When I do count the clock that tells the time, +And see the brave day sunk in hideous night; +When I behold the violet past prime, +And sable curls all silver'd o'er with white; +When lofty trees I see barren of leaves +Which erst from heat did canopy the herd, +And summer's green all girded up in sheaves +Borne on the bier with white and bristly beard, +Then of thy beauty do I question make, +That thou among the wastes of time must go, +Since sweets and beauties do themselves forsake +And die as fast as they see others grow; +And nothing 'gainst Time's scythe can make defence +Save breed, to brave him when he takes thee hence. +O, that you were yourself! but, love, you are +No longer yours than you yourself here live: +Against this coming end you should prepare, +And your sweet semblance to some other give. +So should that beauty which you hold in lease +Find no determination: then you were +Yourself again after yourself's decease, +When your sweet issue your sweet form should bear. +Who lets so fair a house fall to decay, +Which husbandry in honour might uphold +Against the stormy gusts of winter's day +And barren rage of death's eternal cold? +O, none but unthrifts! Dear my love, you know +You had a father: let your son say so. +Not from the stars do I my judgment pluck; +And yet methinks I have astronomy, +But not to tell of good or evil luck, +Of plagues, of dearths, or seasons' quality; +Nor can I fortune to brief minutes tell, +Pointing to each his thunder, rain and wind, +Or say with princes if it shall go well, +By oft predict that I in heaven find: +But from thine eyes my knowledge I derive, +And, constant stars, in them I read such art +As truth and beauty shall together thrive, +If from thyself to store thou wouldst convert; +Or else of thee this I prognosticate: +Thy end is truth's and beauty's doom and date. +When I consider every thing that grows +Holds in perfection but a little moment, +That this huge stage presenteth nought but shows +Whereon the stars in secret influence comment; +When I perceive that men as plants increase, +Cheered and cheque'd even by the self-same sky, +Vaunt in their youthful sap, at height decrease, +And wear their brave state out of memory; +Then the conceit of this inconstant stay +Sets you most rich in youth before my sight, +Where wasteful Time debateth with Decay, +To change your day of youth to sullied night; +And all in war with Time for love of you, +As he takes from you, I engraft you new. +But wherefore do not you a mightier way +Make war upon this bloody tyrant, Time? +And fortify yourself in your decay +With means more blessed than my barren rhyme? +Now stand you on the top of happy hours, +And many maiden gardens yet unset +With virtuous wish would bear your living flowers, +Much liker than your painted counterfeit: +So should the lines of life that life repair, +Which this, Time's pencil, or my pupil pen, +Neither in inward worth nor outward fair, +Can make you live yourself in eyes of men. +To give away yourself keeps yourself still, +And you must live, drawn by your own sweet skill. +Who will believe my verse in time to come, +If it were fill'd with your most high deserts? +Though yet, heaven knows, it is but as a tomb +Which hides your life and shows not half your parts. +If I could write the beauty of your eyes +And in fresh numbers number all your graces, +The age to come would say 'This poet lies: +Such heavenly touches ne'er touch'd earthly faces.' +So should my papers yellow'd with their age +Be scorn'd like old men of less truth than tongue, +And your true rights be term'd a poet's rage +And stretched metre of an antique song: +But were some child of yours alive that time, +You should live twice; in it and in my rhyme. +Shall I compare thee to a summer's day? +Thou art more lovely and more temperate: +Rough winds do shake the darling buds of May, +And summer's lease hath all too short a date: +Sometime too hot the eye of heaven shines, +And often is his gold complexion dimm'd; +And every fair from fair sometime declines, +By chance or nature's changing course untrimm'd; +But thy eternal summer shall not fade +Nor lose possession of that fair thou owest; +Nor shall Death brag thou wander'st in his shade, +When in eternal lines to time thou growest: +So long as men can breathe or eyes can see, +So long lives this and this gives life to thee. +Devouring Time, blunt thou the lion's paws, +And make the earth devour her own sweet brood; +Pluck the keen teeth from the fierce tiger's jaws, +And burn the long-lived phoenix in her blood; +Make glad and sorry seasons as thou fleets, +And do whate'er thou wilt, swift-footed Time, +To the wide world and all her fading sweets; +But I forbid thee one most heinous crime: +O, carve not with thy hours my love's fair brow, +Nor draw no lines there with thine antique pen; +Him in thy course untainted do allow +For beauty's pattern to succeeding men. +Yet, do thy worst, old Time: despite thy wrong, +My love shall in my verse ever live young. +A woman's face with Nature's own hand painted +Hast thou, the master-mistress of my passion; +A woman's gentle heart, but not acquainted +With shifting change, as is false women's fashion; +An eye more bright than theirs, less false in rolling, +Gilding the object whereupon it gazeth; +A man in hue, all 'hues' in his controlling, +Much steals men's eyes and women's souls amazeth. +And for a woman wert thou first created; +Till Nature, as she wrought thee, fell a-doting, +And by addition me of thee defeated, +By adding one thing to my purpose nothing. +But since she prick'd thee out for women's pleasure, +Mine be thy love and thy love's use their treasure. +So is it not with me as with that Muse +Stirr'd by a painted beauty to his verse, +Who heaven itself for ornament doth use +And every fair with his fair doth rehearse +Making a couplement of proud compare, +With sun and moon, with earth and sea's rich gems, +With April's first-born flowers, and all things rare +That heaven's air in this huge rondure hems. +O' let me, true in love, but truly write, +And then believe me, my love is as fair +As any mother's child, though not so bright +As those gold candles fix'd in heaven's air: +Let them say more than like of hearsay well; +I will not praise that purpose not to sell. +My glass shall not persuade me I am old, +So long as youth and thou are of one date; +But when in thee time's furrows I behold, +Then look I death my days should expiate. +For all that beauty that doth cover thee +Is but the seemly raiment of my heart, +Which in thy breast doth live, as thine in me: +How can I then be elder than thou art? +O, therefore, love, be of thyself so wary +As I, not for myself, but for thee will; +Bearing thy heart, which I will keep so chary +As tender nurse her babe from faring ill. +Presume not on thy heart when mine is slain; +Thou gavest me thine, not to give back again. +As an unperfect actor on the stage +Who with his fear is put besides his part, +Or some fierce thing replete with too much rage, +Whose strength's abundance weakens his own heart. +So I, for fear of trust, forget to say +The perfect ceremony of love's rite, +And in mine own love's strength seem to decay, +O'ercharged with burden of mine own love's might. +O, let my books be then the eloquence +And dumb presagers of my speaking breast, +Who plead for love and look for recompense +More than that tongue that more hath more express'd. +O, learn to read what silent love hath writ: +To hear with eyes belongs to love's fine wit. +Mine eye hath play'd the painter and hath stell'd +Thy beauty's form in table of my heart; +My body is the frame wherein 'tis held, +And perspective it is the painter's art. +For through the painter must you see his skill, +To find where your true image pictured lies; +Which in my bosom's shop is hanging still, +That hath his windows glazed with thine eyes. +Now see what good turns eyes for eyes have done: +Mine eyes have drawn thy shape, and thine for me +Are windows to my breast, where-through the sun +Delights to peep, to gaze therein on thee; +Yet eyes this cunning want to grace their art; +They draw but what they see, know not the heart. +Let those who are in favour with their stars +Of public honour and proud titles boast, +Whilst I, whom fortune of such triumph bars, +Unlook'd for joy in that I honour most. +Great princes' favourites their fair leaves spread +But as the marigold at the sun's eye, +And in themselves their pride lies buried, +For at a frown they in their glory die. +The painful warrior famoused for fight, +After a thousand victories once foil'd, +Is from the book of honour razed quite, +And all the rest forgot for which he toil'd: +Then happy I, that love and am beloved +Where I may not remove nor be removed. +Lord of my love, to whom in vassalage +Thy merit hath my duty strongly knit, +To thee I send this written embassage, +To witness duty, not to show my wit: +Duty so great, which wit so poor as mine +May make seem bare, in wanting words to show it, +But that I hope some good conceit of thine +In thy soul's thought, all naked, will bestow it; +Till whatsoever star that guides my moving +Points on me graciously with fair aspect +And puts apparel on my tatter'd loving, +To show me worthy of thy sweet respect: +Then may I dare to boast how I do love thee; +Till then not show my head where thou mayst prove me. +Weary with toil, I haste me to my bed, +The dear repose for limbs with travel tired; +But then begins a journey in my head, +To work my mind, when body's work's expired: +For then my thoughts, from far where I abide, +Intend a zealous pilgrimage to thee, +And keep my drooping eyelids open wide, +Looking on darkness which the blind do see +Save that my soul's imaginary sight +Presents thy shadow to my sightless view, +Which, like a jewel hung in ghastly night, +Makes black night beauteous and her old face new. +Lo! thus, by day my limbs, by night my mind, +For thee and for myself no quiet find. +How can I then return in happy plight, +That am debarr'd the benefit of rest? +When day's oppression is not eased by night, +But day by night, and night by day, oppress'd? +And each, though enemies to either's reign, +Do in consent shake hands to torture me; +The one by toil, the other to complain +How far I toil, still farther off from thee. +I tell the day, to please them thou art bright +And dost him grace when clouds do blot the heaven: +So flatter I the swart-complexion'd night, +When sparkling stars twire not thou gild'st the even. +But day doth daily draw my sorrows longer +And night doth nightly make grief's strength seem stronger. +When, in disgrace with fortune and men's eyes, +I all alone beweep my outcast state +And trouble deal heaven with my bootless cries +And look upon myself and curse my fate, +Wishing me like to one more rich in hope, +Featured like him, like him with friends possess'd, +Desiring this man's art and that man's scope, +With what I most enjoy contented least; +Yet in these thoughts myself almost despising, +Haply I think on thee, and then my state, +Like to the lark at break of day arising +From sullen earth, sings hymns at heaven's gate; +For thy sweet love remember'd such wealth brings +That then I scorn to change my state with kings. +When to the sessions of sweet silent thought +I summon up remembrance of things past, +I sigh the lack of many a thing I sought, +And with old woes new wail my dear time's waste: +Then can I drown an eye, unused to flow, +For precious friends hid in death's dateless night, +And weep afresh love's long since cancell'd woe, +And moan the expense of many a vanish'd sight: +Then can I grieve at grievances foregone, +And heavily from woe to woe tell o'er +The sad account of fore-bemoaned moan, +Which I new pay as if not paid before. +But if the while I think on thee, dear friend, +All losses are restored and sorrows end. +Thy bosom is endeared with all hearts, +Which I by lacking have supposed dead, +And there reigns love and all love's loving parts, +And all those friends which I thought buried. +How many a holy and obsequious tear +Hath dear religious love stol'n from mine eye +As interest of the dead, which now appear +But things removed that hidden in thee lie! +Thou art the grave where buried love doth live, +Hung with the trophies of my lovers gone, +Who all their parts of me to thee did give; +That due of many now is thine alone: +Their images I loved I view in thee, +And thou, all they, hast all the all of me. +If thou survive my well-contented day, +When that churl Death my bones with dust shall cover, +And shalt by fortune once more re-survey +These poor rude lines of thy deceased lover, +Compare them with the bettering of the time, +And though they be outstripp'd by every pen, +Reserve them for my love, not for their rhyme, +Exceeded by the height of happier men. +O, then vouchsafe me but this loving thought: +'Had my friend's Muse grown with this growing age, +A dearer birth than this his love had brought, +To march in ranks of better equipage: +But since he died and poets better prove, +Theirs for their style I'll read, his for his love.' +Full many a glorious morning have I seen +Flatter the mountain-tops with sovereign eye, +Kissing with golden face the meadows green, +Gilding pale streams with heavenly alchemy; +Anon permit the basest clouds to ride +With ugly rack on his celestial face, +And from the forlorn world his visage hide, +Stealing unseen to west with this disgrace: +Even so my sun one early morn did shine +With all triumphant splendor on my brow; +But out, alack! he was but one hour mine; +The region cloud hath mask'd him from me now. +Yet him for this my love no whit disdaineth; +Suns of the world may stain when heaven's sun staineth. +Why didst thou promise such a beauteous day, +And make me travel forth without my cloak, +To let base clouds o'ertake me in my way, +Hiding thy bravery in their rotten smoke? +'Tis not enough that through the cloud thou break, +To dry the rain on my storm-beaten face, +For no man well of such a salve can speak +That heals the wound and cures not the disgrace: +Nor can thy shame give physic to my grief; +Though thou repent, yet I have still the loss: +The offender's sorrow lends but weak relief +To him that bears the strong offence's cross. +Ah! but those tears are pearl which thy love sheds, +And they are rich and ransom all ill deeds. +No more be grieved at that which thou hast done: +Roses have thorns, and silver fountains mud; +Clouds and eclipses stain both moon and sun, +And loathsome canker lives in sweetest bud. +All men make faults, and even I in this, +Authorizing thy trespass with compare, +Myself corrupting, salving thy amiss, +Excusing thy sins more than thy sins are; +For to thy sensual fault I bring in sense-- +Thy adverse party is thy advocate-- +And 'gainst myself a lawful plea commence: +Such civil war is in my love and hate +That I an accessary needs must be +To that sweet thief which sourly robs from me. +Let me confess that we two must be twain, +Although our undivided loves are one: +So shall those blots that do with me remain +Without thy help by me be borne alone. +In our two loves there is but one respect, +Though in our lives a separable spite, +Which though it alter not love's sole effect, +Yet doth it steal sweet hours from love's delight. +I may not evermore acknowledge thee, +Lest my bewailed guilt should do thee shame, +Nor thou with public kindness honour me, +Unless thou take that honour from thy name: +But do not so; I love thee in such sort +As, thou being mine, mine is thy good report. +As a decrepit father takes delight +To see his active child do deeds of youth, +So I, made lame by fortune's dearest spite, +Take all my comfort of thy worth and truth. +For whether beauty, birth, or wealth, or wit, +Or any of these all, or all, or more, +Entitled in thy parts do crowned sit, +I make my love engrafted to this store: +So then I am not lame, poor, nor despised, +Whilst that this shadow doth such substance give +That I in thy abundance am sufficed +And by a part of all thy glory live. +Look, what is best, that best I wish in thee: +This wish I have; then ten times happy me! \ No newline at end of file diff --git a/offline_benchmark_test/benchmarks/structured_schemas/structured_schema_1.json b/offline_benchmark_test/benchmarks/structured_schemas/structured_schema_1.json new file mode 100644 index 0000000000000000000000000000000000000000..13bd6b6d16c601e7e589952b49119205a407acd1 --- /dev/null +++ b/offline_benchmark_test/benchmarks/structured_schemas/structured_schema_1.json @@ -0,0 +1,19 @@ +{ + "type": "object", + "properties": { + "name": { "type": "string" }, + "email": { "type": "string" }, + "street": { "type": "string" }, + "city": { "type": "string" }, + "state": { "type": "string" }, + "zip": { "type": "string" }, + "phone": { "type": "string" }, + "website": { "type": "string" }, + "company": { "type": "string" }, + "age": { "type": "integer" } + }, + "required": [ + "name", + "email" + ] +} diff --git a/online_apiserver_test/README.md b/online_apiserver_test/README.md new file mode 100644 index 0000000000000000000000000000000000000000..286309e7f5a43953b69223db9d65f287ad0fb4c1 --- /dev/null +++ b/online_apiserver_test/README.md @@ -0,0 +1,51 @@ +# Online API Server Benchmark Scripts + +vLLM 0.9.2 在线评测测试的代码/脚本管理 + +## 更新日志 + +2025/10/16：新增vllm0.9.2在线测试的自动化测试脚本（by 刘煦） + +## 使用方法 + +1. 需要安装jq和netcat +``` + apt-get install jq + apt-get install netcat +``` + +2. 文件说明： + + 1）auto_quick_check_config.json：配置文件 + + 2）online_quick_check.sh：自动化启动脚本 + + 3）run_apiserver.sh：运行服务的脚本 + + 4）benchmarks：评测部分的代码，来源为vllm 0.9.2的GitHub + +``` +# 配置文件说明 auto_quick_check_config.json +{ + "DCU": "BW1000", # DCU型号 + "vllm_version": "0.9.2", # vllm版本号 + "pkg_version": "dtk25.04.1", # DTK版本号 + "dst_path": "./result/", # 结果/日志存放目录 + "items":[ + { + "model_name": "Qwen3-32B", # 模型名称 + "model_path": "/data/models/Qwen3-32B/", # 模型路径 + "dtype": "float16", # 数据类型 + "tensor_parallel": [4], # 使用卡数 + "batch_size": [128, 64, 32, 16], # 并发数（降序排列有助于Prefix Caching功能进行预热，从而提升性能） + "seqlen_tuple": ["512 512", "1024 1024"] # 输入输出序列组合 + } + ] +} +``` + +3. 运行及逻辑： + + 指令：bash online_quick_check.sh + + 逻辑：调用run_apiserver.sh启动服务 -> 监听端口是否启动 -> 启动后运行评测指令评测 -> 汇总数据 diff --git a/online_apiserver_test/auto_quick_check_config.json b/online_apiserver_test/auto_quick_check_config.json new file mode 100644 index 0000000000000000000000000000000000000000..5eba2b094e834c08695d95903ea6dd31b0b7e6b6 --- /dev/null +++ b/online_apiserver_test/auto_quick_check_config.json @@ -0,0 +1,16 @@ +{ + "DCU": "BW1000", + "vllm_version": "0.9.2", + "pkg_version": "dtk25.04.1", + "dst_path": "./result/", + "items":[ + { + "model_name": "Qwen3-32B", + "model_path": "/data/models/Qwen3-32B/", + "dtype": "float16", + "tensor_parallel": [4], + "batch_size": [128, 64, 32, 16], + "seqlen_tuple": ["512 512", "1024 1024"] + } + ] +} \ No newline at end of file diff --git a/online_apiserver_test/benchmarks/README.md b/online_apiserver_test/benchmarks/README.md new file mode 100644 index 0000000000000000000000000000000000000000..fb8690d42db98393c118947d7facce057f3cd48f --- /dev/null +++ b/online_apiserver_test/benchmarks/README.md @@ -0,0 +1,605 @@ +# Benchmarking vLLM + +This README guides you through running benchmark tests with the extensive +datasets supported on vLLM. It’s a living document, updated as new features and datasets +become available. + +**Dataset Overview** + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
DatasetOnlineOfflineData Path
ShareGPT✅✅wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
BurstGPT✅✅wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv
Sonnet✅✅Local file: benchmarks/sonnet.txt
Random✅✅synthetic
HuggingFace-VisionArena✅✅lmarena-ai/VisionArena-Chat
HuggingFace-InstructCoder✅✅likaixin/InstructCoder
HuggingFace-AIMO✅✅AI-MO/aimo-validation-aime , AI-MO/NuminaMath-1.5, AI-MO/NuminaMath-CoT
HuggingFace-Other✅✅lmms-lab/LLaVA-OneVision-Data, Aeala/ShareGPT_Vicuna_unfiltered
Custom✅✅Local file: data.jsonl
+ +✅: supported + +🟡: Partial support + +🚧: to be supported + +**Note**: HuggingFace dataset's `dataset-name` should be set to `hf` + +--- +
+🚀 Example - Online Benchmark + +
+ +First start serving your model + +```bash +vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests +``` + +Then run the benchmarking script + +```bash +# download dataset +# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json +python3 vllm/benchmarks/benchmark_serving.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --endpoint /v1/completions \ + --dataset-name sharegpt \ + --dataset-path /ShareGPT_V3_unfiltered_cleaned_split.json \ + --num-prompts 10 +``` + +If successful, you will see the following output + +``` +============ Serving Benchmark Result ============ +Successful requests: 10 +Benchmark duration (s): 5.78 +Total input tokens: 1369 +Total generated tokens: 2212 +Request throughput (req/s): 1.73 +Output token throughput (tok/s): 382.89 +Total Token throughput (tok/s): 619.85 +---------------Time to First Token---------------- +Mean TTFT (ms): 71.54 +Median TTFT (ms): 73.88 +P99 TTFT (ms): 79.49 +-----Time per Output Token (excl. 1st token)------ +Mean TPOT (ms): 7.91 +Median TPOT (ms): 7.96 +P99 TPOT (ms): 8.03 +---------------Inter-token Latency---------------- +Mean ITL (ms): 7.74 +Median ITL (ms): 7.70 +P99 ITL (ms): 8.39 +================================================== +``` + +**Custom Dataset** + +If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl + +``` +{"prompt": "What is the capital of India?"} +{"prompt": "What is the capital of Iran?"} +{"prompt": "What is the capital of China?"} +``` + +```bash +# start server +VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct --disable-log-requests +``` + +```bash +# run benchmarking script +python3 benchmarks/benchmark_serving.py --port 9001 --save-result --save-detailed \ + --backend vllm \ + --model meta-llama/Llama-3.1-8B-Instruct \ + --endpoint /v1/completions \ + --dataset-name custom \ + --dataset-path \ + --custom-skip-chat-template \ + --num-prompts 80 \ + --max-concurrency 1 \ + --temperature=0.3 \ + --top-p=0.75 \ + --result-dir "./log/" +``` + +You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`. + +**VisionArena Benchmark for Vision Language Models** + +```bash +# need a model with vision capability here +vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests +``` + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend openai-chat \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --endpoint /v1/chat/completions \ + --dataset-name hf \ + --dataset-path lmarena-ai/VisionArena-Chat \ + --hf-split train \ + --num-prompts 1000 +``` + +**InstructCoder Benchmark with Speculative Decoding** + +``` bash +VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \ + --speculative-config $'{"method": "ngram", + "num_speculative_tokens": 5, "prompt_lookup_max": 5, + "prompt_lookup_min": 2}' +``` + +``` bash +python3 benchmarks/benchmark_serving.py \ + --model meta-llama/Meta-Llama-3-8B-Instruct \ + --dataset-name hf \ + --dataset-path likaixin/InstructCoder \ + --num-prompts 2048 +``` + +**Other HuggingFaceDataset Examples** + +```bash +vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests +``` + +**`lmms-lab/LLaVA-OneVision-Data`** + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend openai-chat \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --endpoint /v1/chat/completions \ + --dataset-name hf \ + --dataset-path lmms-lab/LLaVA-OneVision-Data \ + --hf-split train \ + --hf-subset "chart2text(cauldron)" \ + --num-prompts 10 +``` + +**`Aeala/ShareGPT_Vicuna_unfiltered`** + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend openai-chat \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --endpoint /v1/chat/completions \ + --dataset-name hf \ + --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \ + --hf-split train \ + --num-prompts 10 +``` + +**`AI-MO/aimo-validation-aime`** + +``` bash +python3 vllm/benchmarks/benchmark_serving.py \ + --model Qwen/QwQ-32B \ + --dataset-name hf \ + --dataset-path AI-MO/aimo-validation-aime \ + --num-prompts 10 \ + --seed 42 +``` + +**`philschmid/mt-bench`** + +``` bash +python3 vllm/benchmarks/benchmark_serving.py \ + --model Qwen/QwQ-32B \ + --dataset-name hf \ + --dataset-path philschmid/mt-bench \ + --num-prompts 80 +``` + +**Running With Sampling Parameters** + +When using OpenAI-compatible backends such as `vllm`, optional sampling +parameters can be specified. Example client command: + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --endpoint /v1/completions \ + --dataset-name sharegpt \ + --dataset-path /ShareGPT_V3_unfiltered_cleaned_split.json \ + --top-k 10 \ + --top-p 0.9 \ + --temperature 0.5 \ + --num-prompts 10 +``` + +**Running With Ramp-Up Request Rate** + +The benchmark tool also supports ramping up the request rate over the +duration of the benchmark run. This can be useful for stress testing the +server or finding the maximum throughput that it can handle, given some latency budget. + +Two ramp-up strategies are supported: +- `linear`: Increases the request rate linearly from a start value to an end value. +- `exponential`: Increases the request rate exponentially. + +The following arguments can be used to control the ramp-up: +- `--ramp-up-strategy`: The ramp-up strategy to use (`linear` or `exponential`). +- `--ramp-up-start-rps`: The request rate at the beginning of the benchmark. +- `--ramp-up-end-rps`: The request rate at the end of the benchmark. + +
+ +
+📈 Example - Offline Throughput Benchmark + +
+ +```bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset-name sonnet \ + --dataset-path vllm/benchmarks/sonnet.txt \ + --num-prompts 10 +``` + +If successful, you will see the following output + +``` +Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s +Total num prompt tokens: 5014 +Total num output tokens: 1500 +``` + +**VisionArena Benchmark for Vision Language Models** + +``` bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --backend vllm-chat \ + --dataset-name hf \ + --dataset-path lmarena-ai/VisionArena-Chat \ + --num-prompts 1000 \ + --hf-split train +``` + +The `num prompt tokens` now includes image token counts + +``` +Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s +Total num prompt tokens: 14527 +Total num output tokens: 1280 +``` + +**InstructCoder Benchmark with Speculative Decoding** + +``` bash +VLLM_WORKER_MULTIPROC_METHOD=spawn \ +VLLM_USE_V1=1 \ +python3 vllm/benchmarks/benchmark_throughput.py \ + --dataset-name=hf \ + --dataset-path=likaixin/InstructCoder \ + --model=meta-llama/Meta-Llama-3-8B-Instruct \ + --input-len=1000 \ + --output-len=100 \ + --num-prompts=2048 \ + --async-engine \ + --speculative-config $'{"method": "ngram", + "num_speculative_tokens": 5, "prompt_lookup_max": 5, + "prompt_lookup_min": 2}' +``` + +``` +Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s +Total num prompt tokens: 261136 +Total num output tokens: 204800 +``` + +**Other HuggingFaceDataset Examples** + +**`lmms-lab/LLaVA-OneVision-Data`** + +```bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --backend vllm-chat \ + --dataset-name hf \ + --dataset-path lmms-lab/LLaVA-OneVision-Data \ + --hf-split train \ + --hf-subset "chart2text(cauldron)" \ + --num-prompts 10 +``` + +**`Aeala/ShareGPT_Vicuna_unfiltered`** + +```bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --backend vllm-chat \ + --dataset-name hf \ + --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \ + --hf-split train \ + --num-prompts 10 +``` + +**`AI-MO/aimo-validation-aime`** + +```bash +python3 benchmarks/benchmark_throughput.py \ + --model Qwen/QwQ-32B \ + --backend vllm \ + --dataset-name hf \ + --dataset-path AI-MO/aimo-validation-aime \ + --hf-split train \ + --num-prompts 10 +``` + +**Benchmark with LoRA Adapters** + +``` bash +# download dataset +# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json +python3 vllm/benchmarks/benchmark_throughput.py \ + --model meta-llama/Llama-2-7b-hf \ + --backend vllm \ + --dataset_path /ShareGPT_V3_unfiltered_cleaned_split.json \ + --dataset_name sharegpt \ + --num-prompts 10 \ + --max-loras 2 \ + --max-lora-rank 8 \ + --enable-lora \ + --lora-path yard1/llama-2-7b-sql-lora-test + ``` + +
+ +
+🛠️ Example - Structured Output Benchmark + +
+ +Benchmark the performance of structured output generation (JSON, grammar, regex). + +**Server Setup** + +```bash +vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests +``` + +**JSON Schema Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset json \ + --structured-output-ratio 1.0 \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**Grammar-based Generation Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset grammar \ + --structure-type grammar \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**Regex-based Generation Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset regex \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**Choice-based Generation Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset choice \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**XGrammar Benchmark Dataset** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset xgrammar_bench \ + --request-rate 10 \ + --num-prompts 1000 +``` + +
+ +
+📚 Example - Long Document QA Benchmark + +
+ +Benchmark the performance of long document question-answering with prefix caching. + +**Basic Long Document QA Test** + +```bash +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 16 \ + --document-length 2000 \ + --output-len 50 \ + --repeat-count 5 +``` + +**Different Repeat Modes** + +```bash +# Random mode (default) - shuffle prompts randomly +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --document-length 3000 \ + --repeat-count 3 \ + --repeat-mode random + +# Tile mode - repeat entire prompt list in sequence +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --document-length 3000 \ + --repeat-count 3 \ + --repeat-mode tile + +# Interleave mode - repeat each prompt consecutively +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --document-length 3000 \ + --repeat-count 3 \ + --repeat-mode interleave +``` + +
+ +
+🗂️ Example - Prefix Caching Benchmark + +
+ +Benchmark the efficiency of automatic prefix caching. + +**Fixed Prompt with Prefix Caching** + +```bash +python3 benchmarks/benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-prompts 1 \ + --repeat-count 100 \ + --input-length-range 128:256 +``` + +**ShareGPT Dataset with Prefix Caching** + +```bash +# download dataset +# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json + +python3 benchmarks/benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --dataset-path /path/ShareGPT_V3_unfiltered_cleaned_split.json \ + --enable-prefix-caching \ + --num-prompts 20 \ + --repeat-count 5 \ + --input-length-range 128:256 +``` + +
+ +
+⚡ Example - Request Prioritization Benchmark + +
+ +Benchmark the performance of request prioritization in vLLM. + +**Basic Prioritization Test** + +```bash +python3 benchmarks/benchmark_prioritization.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --input-len 128 \ + --output-len 64 \ + --num-prompts 100 \ + --scheduling-policy priority +``` + +**Multiple Sequences per Prompt** + +```bash +python3 benchmarks/benchmark_prioritization.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --input-len 128 \ + --output-len 64 \ + --num-prompts 100 \ + --scheduling-policy priority \ + --n 2 +``` + +
diff --git a/online_apiserver_test/benchmarks/auto_tune.sh b/online_apiserver_test/benchmarks/auto_tune.sh new file mode 100644 index 0000000000000000000000000000000000000000..b257b57ce06f5038338bc8a8f114fcb13c0e203b --- /dev/null +++ b/online_apiserver_test/benchmarks/auto_tune.sh @@ -0,0 +1,276 @@ +#!/bin/bash + +# This script aims to tune the best server parameter combinations to maximize throughput for given requirement. +# The current server parameter combination is max_num_seqs and max_num_batched_tokens +# It also supports additional requirement: e2e latency and prefix cache. + +# Pre-requisite: +# 1. Checkout to your branch, install/ update the correct running env. For TPU, activate conda env and install the corresponding torch, xla version. +# 2. If the model is customized, replace the MODEL's config with the customized config. +# 3. Set variables (ALL REQUIRED) +# BASE: your directory for vllm repo +# MODEL: the model served by vllm +# SYSTEM: the hardware, choice TPU or GPU, for other systems, "get best profile" might not support. +# TP: ways of tensor parallelism +# DOWNLOAD_DIR: directory to download and load model weights. +# INPUT_LEN: request input len +# OUTPUT_LEN: request output len +# MIN_CACHE_HIT_PCT: prefix cache rate +# MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000 +# NUM_SEQS_LIST: a list of `max-num-seqs` you want to loop with. +# NUM_BATCHED_TOKENS_LIST: a list of `max-num-batched-tokens` you want to loop with. +# Note that the default NUM_SEQS_LIST and NUM_BATCHED_TOKENS_LIST are set for medium size input/output len, for extra short context (such as 20:20), you might need to include larger numbers in NUM_SEQS_LIST. +# 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens. +# 5. The final result will be saved in RESULT file. + + +# Example use cases +# 1. Given input_len=1800, output_len=20, what's the best max_num_seqs and max_num_batched_tokens to get highest throughput? +# Use INPUT_LEN=1800, OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=100000000000 +# 2. If we have latency requirement to be lower than 500ms, what's the best server parameter? +# Use INPUT_LEN=1800, OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=500 +# 3. If we want to reach 60% prefix cache, what's the best server parameter? +# Use INPUT_LEN=1800, OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=60, MAX_LATENCY_ALLOWED_MS=500 + +TAG=$(date +"%Y_%m_%d_%H_%M") +BASE="" +MODEL="meta-llama/Llama-3.1-8B-Instruct" +SYSTEM="TPU" +TP=1 +DOWNLOAD_DIR="" +INPUT_LEN=4000 +OUTPUT_LEN=16 +MIN_CACHE_HIT_PCT=0 +MAX_LATENCY_ALLOWED_MS=100000000000 +NUM_SEQS_LIST="128 256" +NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096" + +LOG_FOLDER="$BASE/auto-benchmark/$TAG" +RESULT="$LOG_FOLDER/result.txt" +PROFILE_PATH="$LOG_FOLDER/profile" + +echo "result file: $RESULT" +echo "model: $MODEL" + +rm -rf $LOG_FOLDER +rm -rf $PROFILE_PATH +mkdir -p $LOG_FOLDER +mkdir -p $PROFILE_PATH + +cd "$BASE/vllm" + +pip install -q datasets + +current_hash=$(git rev-parse HEAD) +echo "hash:$current_hash" >> "$RESULT" +echo "current_hash: $current_hash" + +best_throughput=0 +best_max_num_seqs=0 +best_num_batched_tokens=0 +best_goodput=0 + +start_server() { + local gpu_memory_utilization=$1 + local max_num_seqs=$2 + local max_num_batched_tokens=$3 + local vllm_log=$4 + local profile_dir=$5 + + pkill -f vllm + + VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \ + --disable-log-requests \ + --port 8004 \ + --gpu-memory-utilization $gpu_memory_utilization \ + --max-num-seqs $max_num_seqs \ + --max-num-batched-tokens $max_num_batched_tokens \ + --tensor-parallel-size $TP \ + --enable-prefix-caching \ + --load-format dummy \ + --download-dir "$DOWNLOAD_DIR" \ + --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 & + + # wait for 10 minutes... + server_started=0 + for i in {1..60}; do + RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout) + STATUS_CODE=$(echo "$RESPONSE" | tail -n 1) + if [[ "$STATUS_CODE" -eq 200 ]]; then + server_started=1 + break + else + sleep 10 + fi + done + if (( ! server_started )); then + echo "server did not start within 10 minutes. Please check server log at $vllm_log". + return 1 + else + return 0 + fi +} + +update_best_profile() { + local profile_dir=$1 + local profile_index=$2 + sorted_paths=($(find "$profile_dir" -maxdepth 1 -not -path "$profile_dir" | sort)) + selected_profile_file= + if [[ "$SYSTEM" == "TPU" ]]; then + selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb" + fi + if [[ "$SYSTEM" == "GPU" ]]; then + selected_profile_file="${sorted_paths[$profile_index]}" + fi + rm -f $PROFILE_PATH/* + cp $selected_profile_file $PROFILE_PATH +} + +run_benchmark() { + local max_num_seqs=$1 + local max_num_batched_tokens=$2 + local gpu_memory_utilization=$3 + echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens" + local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt" + local profile_dir="$LOG_FOLDER/profile_${max_num_seqs}_${max_num_batched_tokens}" + echo "vllm_log: $vllm_log" + echo + rm -f $vllm_log + mkdir -p $profile_dir + pkill -f vllm + local profile_index=0 + + echo "starting server..." + start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log $profile_dir + result=$? + if [[ "$result" -eq 1 ]]; then + echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens" + else + echo "server started." + fi + echo + + echo "run benchmark test..." + meet_latency_requirement=0 + # get a basic qps by using request-rate inf + bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt" + prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 )) + python benchmarks/benchmark_serving.py \ + --backend vllm \ + --model $MODEL \ + --dataset-name random \ + --random-input-len $INPUT_LEN \ + --random-output-len $OUTPUT_LEN \ + --ignore-eos \ + --disable-tqdm \ + --request-rate inf \ + --percentile-metrics ttft,tpot,itl,e2el \ + --goodput e2el:$MAX_LATENCY_ALLOWED_MS \ + --num-prompts 1000 \ + --random-prefix-len $prefix_len \ + --port 8004 \ + --profile &> "$bm_log" + throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}') + goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + + if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then + meet_latency_requirement=1 + request_rate=inf + fi + + if (( ! meet_latency_requirement )); then + # start from request-rate as int(throughput) + 1 + request_rate=$((${throughput%.*} + 1)) + while ((request_rate > 0)); do + profile_index=$((profile_index+1)) + # clear prefix cache + curl -X POST http://0.0.0.0:8004/reset_prefix_cache + sleep 5 + bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt" + python benchmarks/benchmark_serving.py \ + --backend vllm \ + --model $MODEL \ + --dataset-name random \ + --random-input-len $INPUT_LEN \ + --random-output-len $OUTPUT_LEN \ + --ignore-eos \ + --disable-tqdm \ + --request-rate $request_rate \ + --percentile-metrics ttft,tpot,itl,e2el \ + --goodput e2el:$MAX_LATENCY_ALLOWED_MS \ + --num-prompts 100 \ + --random-prefix-len $prefix_len \ + --port 8004 &> "$bm_log" + throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}') + goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then + meet_latency_requirement=1 + break + fi + request_rate=$((request_rate-1)) + done + fi + # write the results and update the best result. + if ((meet_latency_requirement)); then + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" >> "$RESULT" + if (( $(echo "$throughput > $best_throughput" | bc -l) )); then + best_throughput=$throughput + best_max_num_seqs=$max_num_seqs + best_num_batched_tokens=$max_num_batched_tokens + best_goodput=$goodput + if [[ "$SYSTEM" == "TPU" ]]; then + update_best_profile "$profile_dir/plugins/profile" $profile_index + fi + if [[ "$SYSTEM" == "GPU" ]]; then + update_best_profile "$profile_dir" $profile_index + fi + fi + else + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" >> "$RESULT" + fi + + echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput" + + pkill vllm + sleep 10 + printf '=%.0s' $(seq 1 20) + return 0 +} + +read -r -a num_seqs_list <<< "$NUM_SEQS_LIST" +read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST" + +# first find out the max gpu-memory-utilization without HBM OOM. +gpu_memory_utilization=0.98 +find_gpu_memory_utilization=0 +while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do + start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log" + result=$? + if [[ "$result" -eq 0 ]]; then + find_gpu_memory_utilization=1 + break + else + gpu_memory_utilization=$(echo "$gpu_memory_utilization - 0.01" | bc) + fi +done + +if [[ "$find_gpu_memory_utilization" -eq 1 ]]; then + echo "Using gpu_memory_utilization=$gpu_memory_utilization to serve model." +else + echo "Cannot find a proper gpu_memory_utilization over 0.9 to serve the model, please check logs in $LOG_FOLDER." + exit 1 +fi + +for num_seqs in "${num_seqs_list[@]}"; do + for num_batched_tokens in "${num_batched_tokens_list[@]}"; do + run_benchmark $num_seqs $num_batched_tokens $gpu_memory_utilization + done +done +echo "finish permutations" +echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" +echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT" + diff --git a/online_apiserver_test/benchmarks/backend_request_func.py b/online_apiserver_test/benchmarks/backend_request_func.py new file mode 100644 index 0000000000000000000000000000000000000000..c7229dbb8e90d1167eab8fda1a5fb353e31e81a4 --- /dev/null +++ b/online_apiserver_test/benchmarks/backend_request_func.py @@ -0,0 +1,628 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import io +import json +import os +import sys +import time +import traceback +from dataclasses import dataclass, field +from typing import Optional, Union + +import aiohttp +import huggingface_hub.constants +from tqdm.asyncio import tqdm +from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast + +# NOTE(simon): do not import vLLM here so the benchmark script +# can run without vLLM installed. + +AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60) + + +@dataclass +class RequestFuncInput: + prompt: str + api_url: str + prompt_len: int + output_len: int + model: str + model_name: Optional[str] = None + logprobs: Optional[int] = None + extra_body: Optional[dict] = None + multi_modal_content: Optional[dict] = None + ignore_eos: bool = False + language: Optional[str] = None + + +@dataclass +class RequestFuncOutput: + generated_text: str = "" + success: bool = False + latency: float = 0.0 + output_tokens: int = 0 + ttft: float = 0.0 # Time to first token + itl: list[float] = field(default_factory=list) # list of inter-token latencies + tpot: float = 0.0 # avg next-token latencies + prompt_len: int = 0 + error: str = "" + + +async def async_request_tgi( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith("generate_stream") + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + params = { + "max_new_tokens": request_func_input.output_len, + "do_sample": True, + "temperature": 0.01, # TGI does not accept 0.0 temperature. + "top_p": 0.99, # TGI does not accept 1.0 top_p. + "truncate": request_func_input.prompt_len, + "ignore_eos_token": request_func_input.ignore_eos, + } + payload = { + "inputs": request_func_input.prompt, + "parameters": params, + } + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + if request_func_input.ignore_eos: + output.output_tokens = request_func_input.output_len + else: + output.output_tokens = None + + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post(url=api_url, json=payload) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + chunk_bytes = chunk_bytes.decode("utf-8") + + # NOTE: Sometimes TGI returns a ping response without + # any data, we should skip it. + if chunk_bytes.startswith(":"): + continue + chunk = chunk_bytes.removeprefix("data:") + + data = json.loads(chunk) + timestamp = time.perf_counter() + # First token + if ttft == 0.0: + ttft = time.perf_counter() - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + most_recent_timestamp = timestamp + + output.latency = most_recent_timestamp - st + output.success = True + output.generated_text = data["generated_text"] + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_trt_llm( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith("generate_stream") + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + payload = { + "accumulate_tokens": True, + "text_input": request_func_input.prompt, + "temperature": 0.0, + "top_p": 1.0, + "max_tokens": request_func_input.output_len, + "stream": True, + } + if request_func_input.ignore_eos: + payload["min_length"] = request_func_input.output_len + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post(url=api_url, json=payload) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + + chunk = chunk_bytes.decode("utf-8").removeprefix("data:") + + data = json.loads(chunk) + output.generated_text += data["text_output"] + timestamp = time.perf_counter() + # First token + if ttft == 0.0: + ttft = timestamp - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + most_recent_timestamp = timestamp + + output.latency = most_recent_timestamp - st + output.success = True + + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_deepspeed_mii( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith(("completions", "profile")), ( + "OpenAI Completions API URL must end with 'completions' or 'profile'." + ) + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + payload = { + "model": request_func_input.model, + "prompt": request_func_input.prompt, + "max_tokens": request_func_input.output_len, + "temperature": 0.01, # deepspeed-mii does not accept 0.0 temp. + "top_p": 1.0, + } + headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"} + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + # NOTE: DeepSpeed-MII doesn't support streaming as of Jan 28 2024, + # will use 0 as placeholder. + # See https://github.com/microsoft/DeepSpeed-MII/pull/311 + output.ttft = 0 + + st = time.perf_counter() + try: + async with session.post( + url=api_url, json=payload, headers=headers + ) as response: + if response.status == 200: + parsed_resp = await response.json() + output.latency = time.perf_counter() - st + if "choices" in parsed_resp: + output.generated_text = parsed_resp["choices"][0]["text"] + elif "text" in parsed_resp: + output.generated_text = parsed_resp["text"][0] + else: + output.error = ( + "Unexpected response format: " + "neither 'choices' nor 'text' found" + ) + output.success = False + output.success = True + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_openai_completions( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith(("completions", "profile")), ( + "OpenAI Completions API URL must end with 'completions' or 'profile'." + ) + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + payload = { + "model": request_func_input.model_name + if request_func_input.model_name + else request_func_input.model, + "prompt": request_func_input.prompt, + "temperature": 0.0, + "repetition_penalty": 1.0, + "max_tokens": request_func_input.output_len, + "logprobs": request_func_input.logprobs, + "stream": True, + "stream_options": { + "include_usage": True, + }, + } + if request_func_input.ignore_eos: + payload["ignore_eos"] = request_func_input.ignore_eos + if request_func_input.extra_body: + payload.update(request_func_input.extra_body) + headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"} + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + generated_text = "" + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post( + url=api_url, json=payload, headers=headers + ) as response: + if response.status == 200: + first_chunk_received = False + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + + chunk = chunk_bytes.decode("utf-8").removeprefix("data: ") + if chunk != "[DONE]": + data = json.loads(chunk) + + # NOTE: Some completion API might have a last + # usage summary response without a token so we + # want to check a token was generated + if choices := data.get("choices"): + # Note that text could be empty here + # e.g. for special tokens + text = choices[0].get("text") + timestamp = time.perf_counter() + # First token + if not first_chunk_received: + first_chunk_received = True + ttft = time.perf_counter() - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + most_recent_timestamp = timestamp + generated_text += text or "" + if usage := data.get("usage"): + output.output_tokens = usage.get("completion_tokens") + if first_chunk_received: + output.success = True + else: + output.success = False + output.error = ( + "Never received a valid chunk to calculate TTFT." + "This response will be marked as failed!" + ) + output.generated_text = generated_text + output.latency = most_recent_timestamp - st + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_openai_chat_completions( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith(("chat/completions", "profile")), ( + "OpenAI Chat Completions API URL must end with 'chat/completions'." + ) + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + content = [{"type": "text", "text": request_func_input.prompt}] + if request_func_input.multi_modal_content: + content.append(request_func_input.multi_modal_content) + payload = { + "model": request_func_input.model_name + if request_func_input.model_name + else request_func_input.model, + "messages": [ + {"role": "user", "content": content}, + ], + "temperature": 0.0, + "max_completion_tokens": request_func_input.output_len, + "stream": True, + "stream_options": { + "include_usage": True, + }, + } + if request_func_input.ignore_eos: + payload["ignore_eos"] = request_func_input.ignore_eos + if request_func_input.extra_body: + payload.update(request_func_input.extra_body) + headers = { + "Content-Type": "application/json", + "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}", + } + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + generated_text = "" + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post( + url=api_url, json=payload, headers=headers + ) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + chunk_bytes = chunk_bytes.decode("utf-8") + # NOTE: SSE comments (often used as pings) start with a colon. + # These are not JSON data payload and should be skipped. + if chunk_bytes.startswith(":"): + continue + + chunk = chunk_bytes.removeprefix("data: ") + + if chunk != "[DONE]": + timestamp = time.perf_counter() + data = json.loads(chunk) + + if choices := data.get("choices"): + content = choices[0]["delta"].get("content") + # First token + if ttft == 0.0: + ttft = timestamp - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + generated_text += content or "" + elif usage := data.get("usage"): + output.output_tokens = usage.get("completion_tokens") + + most_recent_timestamp = timestamp + + output.generated_text = generated_text + output.success = True + output.latency = most_recent_timestamp - st + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_openai_audio( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + # Lazy import without PlaceholderModule to avoid vllm dep. + import soundfile + + api_url = request_func_input.api_url + assert api_url.endswith(("transcriptions", "translations")), ( + "OpenAI Chat Completions API URL must end with 'transcriptions' " + ) + "or `translations`." + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + content = [{"type": "text", "text": request_func_input.prompt}] + payload = { + "model": request_func_input.model_name + if request_func_input.model_name + else request_func_input.model, + "temperature": 0.0, + "max_completion_tokens": request_func_input.output_len, + "stream": True, + "language": "en", + # Flattened due to multipart/form-data + "stream_include_usage": True, + "stream_continuous_usage_stats": True, + } + if request_func_input.extra_body: + payload.update(request_func_input.extra_body) + headers = { + "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}", + } + + # Send audio file + def to_bytes(y, sr): + buffer = io.BytesIO() + soundfile.write(buffer, y, sr, format="WAV") + buffer.seek(0) + return buffer + + with to_bytes(*request_func_input.multi_modal_content["audio"]) as f: + form = aiohttp.FormData() + form.add_field("file", f, content_type="audio/wav") + for key, value in payload.items(): + form.add_field(key, str(value)) + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + generated_text = "" + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post( + url=api_url, data=form, headers=headers + ) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + + chunk = chunk_bytes.decode("utf-8").removeprefix("data: ") + if chunk != "[DONE]": + timestamp = time.perf_counter() + data = json.loads(chunk) + + if choices := data.get("choices"): + content = choices[0]["delta"].get("content") + # First token + if ttft == 0.0: + ttft = timestamp - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append( + timestamp - most_recent_timestamp + ) + + generated_text += content or "" + elif usage := data.get("usage"): + output.output_tokens = usage.get( + "completion_tokens" + ) + + most_recent_timestamp = timestamp + + output.generated_text = generated_text + output.success = True + output.latency = most_recent_timestamp - st + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +def get_model(pretrained_model_name_or_path: str) -> str: + if os.getenv("VLLM_USE_MODELSCOPE", "False").lower() == "true": + from modelscope import snapshot_download + + from vllm.model_executor.model_loader.weight_utils import get_lock + + # Use file lock to prevent multiple processes from + # downloading the same model weights at the same time. + with get_lock(pretrained_model_name_or_path): + model_path = snapshot_download( + model_id=pretrained_model_name_or_path, + local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, + ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"], + ) + + return model_path + return pretrained_model_name_or_path + + +def get_tokenizer( + pretrained_model_name_or_path: str, + tokenizer_mode: str = "auto", + trust_remote_code: bool = False, + **kwargs, +) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]: + if pretrained_model_name_or_path is not None and not os.path.exists( + pretrained_model_name_or_path + ): + pretrained_model_name_or_path = get_model(pretrained_model_name_or_path) + if tokenizer_mode == "slow": + if kwargs.get("use_fast", False): + raise ValueError("Cannot use the fast tokenizer in slow tokenizer mode.") + kwargs["use_fast"] = False + if tokenizer_mode == "mistral": + try: + from vllm.transformers_utils.tokenizer import MistralTokenizer + except ImportError as e: + raise ImportError( + "MistralTokenizer requires vllm package.\n" + "Please install it with `pip install vllm` " + "to use mistral tokenizer mode." + ) from e + return MistralTokenizer.from_pretrained(str(pretrained_model_name_or_path)) + else: + return AutoTokenizer.from_pretrained( + pretrained_model_name_or_path, + trust_remote_code=trust_remote_code, + **kwargs, + ) + + +ASYNC_REQUEST_FUNCS = { + "tgi": async_request_tgi, + "vllm": async_request_openai_completions, + "lmdeploy": async_request_openai_completions, + "deepspeed-mii": async_request_deepspeed_mii, + "openai": async_request_openai_completions, + "openai-chat": async_request_openai_chat_completions, + "openai-audio": async_request_openai_audio, + "tensorrt-llm": async_request_trt_llm, + "scalellm": async_request_openai_completions, + "sglang": async_request_openai_completions, + "llama.cpp": async_request_openai_completions, +} + +OPENAI_COMPATIBLE_BACKENDS = [ + k + for k, v in ASYNC_REQUEST_FUNCS.items() + if v in (async_request_openai_completions, async_request_openai_chat_completions) +] diff --git a/online_apiserver_test/benchmarks/benchmark_dataset.py b/online_apiserver_test/benchmarks/benchmark_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..55c0cf851264f37d271f7a8cd1fc5ad17451f386 --- /dev/null +++ b/online_apiserver_test/benchmarks/benchmark_dataset.py @@ -0,0 +1,1168 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +This module defines a framework for sampling benchmark requests from various +datasets. Each dataset subclass of BenchmarkDataset must implement sample +generation. Supported dataset types include: + - ShareGPT + - Random (synthetic) + - Sonnet + - BurstGPT + - HuggingFace + - VisionArena +""" + +import base64 +import io +import json +import logging +import random +from abc import ABC, abstractmethod +from collections.abc import Mapping +from dataclasses import dataclass +from functools import cache +from io import BytesIO +from typing import Any, Callable, Optional, Union + +import numpy as np +import pandas as pd +from datasets import load_dataset +from PIL import Image +from transformers import PreTrainedTokenizerBase + +from vllm.lora.request import LoRARequest +from vllm.lora.utils import get_adapter_absolute_path +from vllm.multimodal import MultiModalDataDict +from vllm.multimodal.image import convert_image_mode +from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer + +logger = logging.getLogger(__name__) + +# ----------------------------------------------------------------------------- +# Data Classes +# ----------------------------------------------------------------------------- + + +@dataclass +class SampleRequest: + """ + Represents a single inference request for benchmarking. + """ + + prompt: Union[str, Any] + prompt_len: int + expected_output_len: int + multi_modal_data: Optional[Union[MultiModalDataDict, dict]] = None + lora_request: Optional[LoRARequest] = None + + +# ----------------------------------------------------------------------------- +# Benchmark Dataset Base Class +# ----------------------------------------------------------------------------- + + +class BenchmarkDataset(ABC): + DEFAULT_SEED = 0 + IS_MULTIMODAL = False + + def __init__( + self, + dataset_path: Optional[str] = None, + random_seed: int = DEFAULT_SEED, + ) -> None: + """ + Initialize the BenchmarkDataset with an optional dataset path and random + seed. Args: + dataset_path (Optional[str]): Path to the dataset. If None, it + indicates that a default or random dataset might be used. + random_seed (int): Seed value for reproducible shuffling or + sampling. Defaults to DEFAULT_SEED. + """ + self.dataset_path = dataset_path + # Set the random seed, ensuring that a None value is replaced with the + # default seed. + self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED + self.data = None + + def apply_multimodal_chat_transformation( + self, prompt: str, mm_content: Optional[MultiModalDataDict] = None + ) -> list[dict]: + """ + Transform a prompt and optional multimodal content into a chat format. + This method is used for chat models that expect a specific conversation + format. + """ + content = [{"text": prompt, "type": "text"}] + if mm_content is not None: + content.append(mm_content) + return [{"role": "user", "content": content}] + + def load_data(self) -> None: + """ + Load data from the dataset path into self.data. + + This method must be overridden by subclasses since the method to load + data will vary depending on the dataset format and source. + + Raises: + NotImplementedError: If a subclass does not implement this method. + """ + # TODO (jenniferzhao): add support for downloading data + raise NotImplementedError("load_data must be implemented in subclasses.") + + def get_random_lora_request( + self, + tokenizer: PreTrainedTokenizerBase, + max_loras: Optional[int] = None, + lora_path: Optional[str] = None, + ) -> tuple[Optional[LoRARequest], AnyTokenizer]: + """ + Optionally select a random LoRA request and return its associated + tokenizer. + + This method is used when LoRA parameters are provided. It randomly + selects a LoRA based on max_loras and retrieves a cached tokenizer for + that LoRA if available. Otherwise, it returns the base tokenizer. + + Args: + tokenizer (PreTrainedTokenizerBase): The base tokenizer to use if no + LoRA is selected. max_loras (Optional[int]): The maximum number of + LoRAs available. If None, LoRA is not used. lora_path + (Optional[str]): Path to the LoRA parameters on disk. If None, LoRA + is not used. + + Returns: + tuple[Optional[LoRARequest], AnyTokenizer]: A tuple where the first + element is a LoRARequest (or None if not applicable) and the second + element is the tokenizer associated with the LoRA request (or the + base tokenizer). + """ + if max_loras is None or lora_path is None: + return None, tokenizer + + # Generate a random LoRA ID in the range [1, max_loras]. + lora_id = random.randint(1, max_loras) + lora_request = LoRARequest( + lora_name=str(lora_id), + lora_int_id=lora_id, + lora_path=lora_path_on_disk(lora_path), + ) + if lora_id not in lora_tokenizer_cache: + lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request) + # Return lora_request and the cached tokenizer if available; otherwise, + # return the base tokenizer + return lora_request, lora_tokenizer_cache[lora_id] or tokenizer + + @abstractmethod + def sample( + self, tokenizer: PreTrainedTokenizerBase, num_requests: int + ) -> list[SampleRequest]: + """ + Abstract method to generate sample requests from the dataset. + + Subclasses must override this method to implement dataset-specific logic + for generating a list of SampleRequest objects. + + Args: + tokenizer (PreTrainedTokenizerBase): The tokenizer to be used + for processing the dataset's text. + num_requests (int): The number of sample requests to generate. + + Returns: + list[SampleRequest]: A list of sample requests generated from the + dataset. + """ + raise NotImplementedError("sample must be implemented in subclasses.") + + def maybe_oversample_requests( + self, requests: list[SampleRequest], num_requests: int + ) -> None: + """ + Oversamples the list of requests if its size is less than the desired + number. + + Args: + requests (List[SampleRequest]): The current list of sampled + requests. num_requests (int): The target number of requests. + """ + if len(requests) < num_requests: + random.seed(self.random_seed) + additional = random.choices(requests, k=num_requests - len(requests)) + requests.extend(additional) + logger.info("Oversampled requests to reach %d total samples.", num_requests) + + +# ----------------------------------------------------------------------------- +# Utility Functions and Global Caches +# ----------------------------------------------------------------------------- + + +def is_valid_sequence( + prompt_len: int, + output_len: int, + min_len: int = 4, + max_prompt_len: int = 1024, + max_total_len: int = 2048, + skip_min_output_len_check: bool = False, +) -> bool: + """ + Validate a sequence based on prompt and output lengths. + + Default pruning criteria are copied from the original `sample_hf_requests` + and `sample_sharegpt_requests` functions in benchmark_serving.py, as well as + from `sample_requests` in benchmark_throughput.py. + """ + # Check for invalid conditions + prompt_too_short = prompt_len < min_len + output_too_short = (not skip_min_output_len_check) and (output_len < min_len) + prompt_too_long = prompt_len > max_prompt_len + combined_too_long = (prompt_len + output_len) > max_total_len + + # Return True if none of the invalid conditions are met + return not ( + prompt_too_short or output_too_short or prompt_too_long or combined_too_long + ) + + +@cache +def lora_path_on_disk(lora_path: str) -> str: + return get_adapter_absolute_path(lora_path) + + +# Global cache for LoRA tokenizers. +lora_tokenizer_cache: dict[int, AnyTokenizer] = {} + + +def process_image(image: Any) -> Mapping[str, Any]: + """ + Process a single image input and return a multimedia content dictionary. + + Supports three input types: + + 1. Dictionary with raw image bytes: - Expects a dict with a 'bytes' key + containing raw image data. - Loads the bytes as a PIL.Image.Image. + + 2. PIL.Image.Image input: - Converts the image to RGB. - Saves the image as + a JPEG in memory. - Encodes the JPEG data as a base64 string. - Returns + a dictionary with the image as a base64 data URL. + + 3. String input: - Treats the string as a URL or local file path. - + Prepends "file://" if the string doesn't start with "http://" or + "file://". - Returns a dictionary with the image URL. + + Raises: + ValueError: If the input is not a supported type. + """ + if isinstance(image, dict) and "bytes" in image: + image = Image.open(BytesIO(image["bytes"])) + if isinstance(image, Image.Image): + image = convert_image_mode(image, "RGB") + with io.BytesIO() as image_data: + image.save(image_data, format="JPEG") + image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8") + return { + "type": "image_url", + "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"}, + } + + if isinstance(image, str): + image_url = ( + image if image.startswith(("http://", "file://")) else f"file://{image}" + ) + return {"type": "image_url", "image_url": {"url": image_url}} + + raise ValueError( + f"Invalid image input {image}. Must be a PIL.Image.Image" + " or str or dictionary with raw image bytes." + ) + + +# ----------------------------------------------------------------------------- +# Random Dataset Implementation (Synthetic Data) +# ----------------------------------------------------------------------------- + + +class RandomDataset(BenchmarkDataset): + # Default values copied from benchmark_serving.py for the random dataset. + DEFAULT_PREFIX_LEN = 0 + DEFAULT_RANGE_RATIO = 0.0 + DEFAULT_INPUT_LEN = 1024 + DEFAULT_OUTPUT_LEN = 128 + + def __init__( + self, + **kwargs, + ) -> None: + super().__init__(**kwargs) + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + prefix_len: int = DEFAULT_PREFIX_LEN, + range_ratio: float = DEFAULT_RANGE_RATIO, + input_len: int = DEFAULT_INPUT_LEN, + output_len: int = DEFAULT_OUTPUT_LEN, + **kwargs, + ) -> list[SampleRequest]: + # Enforce range_ratio < 1 + assert range_ratio < 1.0, ( + "random_range_ratio must be < 1.0 to ensure a valid sampling range" + ) + + vocab_size = tokenizer.vocab_size + num_special_tokens = tokenizer.num_special_tokens_to_add() + real_input_len = input_len - num_special_tokens + + prefix_token_ids = ( + np.random.randint(0, vocab_size, size=prefix_len).tolist() + if prefix_len > 0 + else [] + ) + + # New sampling logic: [X * (1 - b), X * (1 + b)] + input_low = int(real_input_len * (1 - range_ratio)) + input_high = int(real_input_len * (1 + range_ratio)) + output_low = int(output_len * (1 - range_ratio)) + output_high = int(output_len * (1 + range_ratio)) + + # Add logging for debugging + logger.info("Sampling input_len from [%s, %s]", input_low, input_high) + logger.info("Sampling output_len from [%s, %s]", output_low, output_high) + + input_lens = np.random.randint(input_low, input_high + 1, size=num_requests) + output_lens = np.random.randint(output_low, output_high + 1, size=num_requests) + offsets = np.random.randint(0, vocab_size, size=num_requests) + + requests = [] + for i in range(num_requests): + inner_seq = ( + (offsets[i] + i + np.arange(input_lens[i])) % vocab_size + ).tolist() + token_sequence = prefix_token_ids + inner_seq + prompt = tokenizer.decode(token_sequence) + # After decoding the prompt we have to encode and decode it again. + # This is done because in some cases N consecutive tokens + # give a string tokenized into != N number of tokens. + # For example for GPT2Tokenizer: + # [6880, 6881] -> ['Ġcalls', 'here'] -> + # [1650, 939, 486] -> ['Ġcall', 'sh', 'ere'] + # To avoid uncontrolled change of the prompt length, + # the encoded sequence is truncated before being decode again. + total_input_len = prefix_len + int(input_lens[i]) + re_encoded_sequence = tokenizer.encode(prompt, add_special_tokens=False)[ + :total_input_len + ] + prompt = tokenizer.decode(re_encoded_sequence) + total_input_len = len(re_encoded_sequence) + requests.append( + SampleRequest( + prompt=prompt, + prompt_len=total_input_len, + expected_output_len=int(output_lens[i]), + ) + ) + return requests + + +# ----------------------------------------------------------------------------- +# ShareGPT Dataset Implementation +# ----------------------------------------------------------------------------- + + +class ShareGPTDataset(BenchmarkDataset): + """ + Implements the ShareGPT dataset. Loads data from a JSON file and generates + sample requests based on conversation turns. + """ + + def __init__(self, **kwargs) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data(self) -> None: + if self.dataset_path is None: + raise ValueError("dataset_path must be provided for loading data.") + + with open(self.dataset_path, encoding="utf-8") as f: + self.data = json.load(f) + # Filter entries with at least two conversation turns. + self.data = [ + entry + for entry in self.data + if "conversations" in entry and len(entry["conversations"]) >= 2 + ] + random.seed(self.random_seed) + random.shuffle(self.data) + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + lora_path: Optional[str] = None, + max_loras: Optional[int] = None, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + samples: list = [] + for entry in self.data: + if len(samples) >= num_requests: + break + prompt, completion = ( + entry["conversations"][0]["value"], + entry["conversations"][1]["value"], + ) + + lora_request, tokenizer = self.get_random_lora_request( + tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path + ) + prompt_ids = tokenizer(prompt).input_ids + completion_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_ids) + new_output_len = len(completion_ids) if output_len is None else output_len + if not is_valid_sequence( + prompt_len, + new_output_len, + skip_min_output_len_check=output_len is not None, + ): + continue + if enable_multimodal_chat: + prompt = self.apply_multimodal_chat_transformation(prompt, None) + samples.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=new_output_len, + lora_request=lora_request, + ) + ) + self.maybe_oversample_requests(samples, num_requests) + return samples + + +# ----------------------------------------------------------------------------- +# Custom Dataset Implementation +# ----------------------------------------------------------------------------- + + +class CustomDataset(BenchmarkDataset): + """ + Implements the Custom dataset. Loads data from a JSONL file and generates + sample requests based on conversation turns. E.g., + ``` + {"prompt": "What is the capital of India?"} + {"prompt": "What is the capital of Iran?"} + {"prompt": "What is the capital of China?"} + ``` + """ + + def __init__(self, **kwargs) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data(self) -> None: + if self.dataset_path is None: + raise ValueError("dataset_path must be provided for loading data.") + + # self.data will be a list of dictionaries + # e.g., [{"prompt": "What is the capital of India?"}, ...] + # This will be the standardized format which load_data() + # has to convert into depending on the filetype of dataset_path. + # sample() will assume this standardized format of self.data + self.data = [] + + # Load the JSONL file + if self.dataset_path.endswith(".jsonl"): + jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True) + + # check if the JSONL file has a 'prompt' column + if "prompt" not in jsonl_data.columns: + raise ValueError("JSONL file must contain a 'prompt' column.") + + # Convert each row to a dictionary and append to self.data + # This will convert the DataFrame to a list of dictionaries + # where each dictionary corresponds to a row in the DataFrame. + # This is the standardized format we want for self.data + for _, row in jsonl_data.iterrows(): + self.data.append(row.to_dict()) + else: + raise NotImplementedError( + "Only JSONL format is supported for CustomDataset." + ) + + random.seed(self.random_seed) + random.shuffle(self.data) + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + lora_path: Optional[str] = None, + max_loras: Optional[int] = None, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + skip_chat_template: bool = False, + **kwargs, + ) -> list: + sampled_requests = [] + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt = item["prompt"] + + # apply template + if not skip_chat_template: + prompt = tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=True, + tokenize=False, + ) + + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Sonnet Dataset Implementation +# ----------------------------------------------------------------------------- + + +class SonnetDataset(BenchmarkDataset): + """ + Simplified implementation of the Sonnet dataset. Loads poem lines from a + text file and generates sample requests. Default values here copied from + `benchmark_serving.py` for the sonnet dataset. + """ + + DEFAULT_PREFIX_LEN = 200 + DEFAULT_INPUT_LEN = 550 + DEFAULT_OUTPUT_LEN = 150 + + def __init__( + self, + **kwargs, + ) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data(self) -> None: + if not self.dataset_path: + raise ValueError("dataset_path must be provided.") + with open(self.dataset_path, encoding="utf-8") as f: + self.data = f.readlines() + + def sample( + self, + tokenizer, + num_requests: int, + prefix_len: int = DEFAULT_PREFIX_LEN, + input_len: int = DEFAULT_INPUT_LEN, + output_len: int = DEFAULT_OUTPUT_LEN, + return_prompt_formatted: bool = False, + **kwargs, + ) -> list: + # Calculate average token length for a poem line. + tokenized_lines = [tokenizer(line).input_ids for line in self.data] + avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines) + + # Build the base prompt. + base_prompt = "Pick as many lines as you can from these poem lines:\n" + base_msg = [{"role": "user", "content": base_prompt}] + base_fmt = tokenizer.apply_chat_template( + base_msg, add_generation_prompt=True, tokenize=False + ) + base_offset = len(tokenizer(base_fmt).input_ids) + if input_len <= base_offset: + raise ValueError( + f"'input_len' must be higher than the base prompt length " + f"({base_offset})." + ) + + # Determine how many poem lines to use. + num_input_lines = round((input_len - base_offset) / avg_len) + num_prefix_lines = max(round((prefix_len - base_offset) / avg_len), 0) + prefix_lines = self.data[:num_prefix_lines] + + samples = [] + while len(samples) < num_requests: + extra_lines = random.choices( + self.data, k=num_input_lines - num_prefix_lines + ) + prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}" + msg = [{"role": "user", "content": prompt}] + prompt_formatted = tokenizer.apply_chat_template( + msg, add_generation_prompt=True, tokenize=False + ) + prompt_len = len(tokenizer(prompt_formatted).input_ids) + if prompt_len <= input_len: + samples.append( + SampleRequest( + prompt=prompt_formatted if return_prompt_formatted else prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + return samples + + +# ----------------------------------------------------------------------------- +# BurstGPT Dataset Implementation +# ----------------------------------------------------------------------------- + + +class BurstGPTDataset(BenchmarkDataset): + """ + Implements the BurstGPT dataset. Loads data from a CSV file and generates + sample requests based on synthetic prompt generation. Only rows with Model + "GPT-4" and positive response tokens are used. + """ + + def __init__(self, **kwargs) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data( + self, + ): + if self.dataset_path is None: + raise ValueError("dataset_path must be provided for loading data.") + + df = pd.read_csv(self.dataset_path) + # Filter to keep only GPT-4 rows. + gpt4_df = df[df["Model"] == "GPT-4"] + # Remove failed requests (where Response tokens is 0 or less). + gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0] + # Sample the desired number of rows. + self.data = gpt4_df + + def _sample_loaded_data(self, num_requests: int) -> list: + if num_requests <= len(self.data): + data = self.data.sample(n=num_requests, random_state=self.random_seed) + else: + data = self.data.sample( + n=num_requests, + random_state=self.random_seed, + replace=True, + ) + # Convert the dataframe to a list of lists. + return data.values.tolist() + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + max_loras: Optional[int] = None, + lora_path: Optional[str] = None, + **kwargs, + ) -> list[SampleRequest]: + samples = [] + data = self._sample_loaded_data(num_requests=num_requests) + for i in range(num_requests): + input_len = int(data[i][2]) + output_len = int(data[i][3]) + lora_req, tokenizer = self.get_random_lora_request( + tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path + ) + vocab_size = tokenizer.vocab_size + # Generate a synthetic prompt: a list of token IDs computed as (i + + # j) modulo vocab_size. + token_ids = [(i + j) % vocab_size for j in range(input_len)] + prompt = tokenizer.decode(token_ids) + samples.append( + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=output_len, + lora_request=lora_req, + ) + ) + return samples + + +# ----------------------------------------------------------------------------- +# HuggingFace Dataset Base Implementation +# ----------------------------------------------------------------------------- +class HuggingFaceDataset(BenchmarkDataset): + """Base class for datasets hosted on HuggingFace.""" + + SUPPORTED_DATASET_PATHS: Union[set[str], dict[str, Callable]] = set() + + def __init__( + self, + dataset_path: str, + dataset_split: str, + dataset_subset: Optional[str] = None, + **kwargs, + ) -> None: + super().__init__(dataset_path=dataset_path, **kwargs) + + self.dataset_split = dataset_split + self.dataset_subset = dataset_subset + self.load_data() + + def load_data(self) -> None: + """Load data from HuggingFace datasets.""" + self.data = load_dataset( + self.dataset_path, + name=self.dataset_subset, + split=self.dataset_split, + streaming=True, + ) + self.data = self.data.shuffle(seed=self.random_seed) + + +# ----------------------------------------------------------------------------- +# Conversation Dataset Implementation +# ----------------------------------------------------------------------------- + + +class ConversationDataset(HuggingFaceDataset): + """Dataset for conversation data with multimodal support.""" + + SUPPORTED_DATASET_PATHS = { + "lmms-lab/LLaVA-OneVision-Data", + "Aeala/ShareGPT_Vicuna_unfiltered", + } + IS_MULTIMODAL = True + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + # Filter examples with at least 2 conversations + filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2) + sampled_requests = [] + dynamic_output = output_len is None + + for item in filtered_data: + if len(sampled_requests) >= num_requests: + break + conv = item["conversations"] + prompt, completion = conv[0]["value"], conv[1]["value"] + + prompt_ids = tokenizer(prompt).input_ids + completion_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_ids) + completion_len = len(completion_ids) + output_len = completion_len if dynamic_output else output_len + assert isinstance(output_len, int) and output_len > 0 + if dynamic_output and not is_valid_sequence(prompt_len, completion_len): + continue + mm_content = process_image(item["image"]) if "image" in item else None + if enable_multimodal_chat: + # Note: when chat is enabled the request prompt_len is no longer + # accurate and we will be using request output to count the + # actual prompt len and output len + prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=mm_content, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Vision Arena Dataset Implementation +# ----------------------------------------------------------------------------- + + +class VisionArenaDataset(HuggingFaceDataset): + """ + Vision Arena Dataset. + """ + + DEFAULT_OUTPUT_LEN = 128 + SUPPORTED_DATASET_PATHS = { + "lmarena-ai/VisionArena-Chat": lambda x: x["conversation"][0][0]["content"], + "lmarena-ai/vision-arena-bench-v0.1": lambda x: x["turns"][0][0]["content"], + } + IS_MULTIMODAL = True + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + sampled_requests = [] + for item in self.data: + if len(sampled_requests) >= num_requests: + break + parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.dataset_path) + if parser_fn is None: + raise ValueError(f"Unsupported dataset path: {self.dataset_path}") + prompt = parser_fn(item) + mm_content = process_image(item["images"][0]) + prompt_len = len(tokenizer(prompt).input_ids) + if enable_multimodal_chat: + # Note: when chat is enabled the request prompt_len is no longer + # accurate and we will be using request output to count the + # actual prompt len + prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=mm_content, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Instruct Coder Dataset Implementation +# ----------------------------------------------------------------------------- + + +class InstructCoderDataset(HuggingFaceDataset): + """ + InstructCoder Dataset. + https://huggingface.co/datasets/likaixin/InstructCoder + + InstructCoder is the dataset designed for general code editing. It consists + of 114,239 instruction-input-output triplets, and covers multiple distinct + code editing scenario. + """ + + DEFAULT_OUTPUT_LEN = 200 # this is the average default output length + SUPPORTED_DATASET_PATHS = { + "likaixin/InstructCoder", + } + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + sampled_requests = [] + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt = f"{item['input']}\n\n{item['instruction']} Just output \ + the code, do not include any explanation." + + # apply template + prompt = tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=True, + tokenize=False, + ) + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# MT-Bench Dataset Implementation +# ----------------------------------------------------------------------------- + + +class MTBenchDataset(HuggingFaceDataset): + """ + MT-Bench Dataset. + https://huggingface.co/datasets/philschmid/mt-bench + + We create a single turn dataset for MT-Bench. + This is similar to Spec decoding benchmark setup in vLLM + https://github.com/vllm-project/vllm/blob/9d98ab5ec/examples/offline_inference/eagle.py#L14-L18 + """ # noqa: E501 + + DEFAULT_OUTPUT_LEN = 256 # avg len used in SD bench in vLLM + SUPPORTED_DATASET_PATHS = { + "philschmid/mt-bench", + } + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + sampled_requests = [] + + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt = item["turns"][0] + + # apply template + prompt = tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=True, + tokenize=False, + ) + + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# AIMO Dataset Implementation +# ----------------------------------------------------------------------------- + + +class AIMODataset(HuggingFaceDataset): + """ + Dataset class for processing a AIMO dataset with reasoning questions. + """ + + SUPPORTED_DATASET_PATHS = { + "AI-MO/aimo-validation-aime", + "AI-MO/NuminaMath-1.5", + "AI-MO/NuminaMath-CoT", + } + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + **kwargs, + ) -> list: + sampled_requests = [] + dynamic_output = output_len is None + + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt, completion = item["problem"], item["solution"] + + prompt_ids = tokenizer(prompt).input_ids + completion_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_ids) + completion_len = len(completion_ids) + output_len = completion_len if dynamic_output else output_len + assert isinstance(output_len, int) and output_len > 0 + if dynamic_output and not is_valid_sequence( + prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000 + ): + continue + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=None, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Next Edit Prediction Dataset Implementation +# ----------------------------------------------------------------------------- + + +zeta_prompt = """### Instruction: +You are a code completion assistant and your task is to analyze user edits and then rewrite an excerpt that the user provides, suggesting the appropriate edits within the excerpt, taking into account the cursor location. + +### User Edits: + +{} + +### User Excerpt: + +{} + +### Response: + +""" # noqa: E501 + + +def _format_zeta_prompt( + sample: dict, original_start_marker: str = "<|editable_region_start|>" +) -> dict: + """Format the zeta prompt for the Next Edit Prediction (NEP) dataset. + + This function formats examples from the NEP dataset + into prompts and expected outputs. It could be + further extended to support more NEP datasets. + + Args: + sample: The dataset sample containing events, + inputs, and outputs. + original_start_marker: The marker indicating the + start of the editable region. Defaults to + "<|editable_region_start|>". + + Returns: + A dictionary with the formatted prompts and expected outputs. + """ + events = sample["events"] + input = sample["input"] + output = sample["output"] + prompt = zeta_prompt.format(events, input) + + # following the original implementation, extract the focused region + # from the raw output + output_start_index = output.find(original_start_marker) + output_focused_region = output[output_start_index:] + expected_output = output_focused_region + + return {"prompt": prompt, "expected_output": expected_output} + + +class NextEditPredictionDataset(HuggingFaceDataset): + """ + Dataset class for processing a Next Edit Prediction dataset. + """ + + SUPPORTED_DATASET_PATHS = { + "zed-industries/zeta", + } + MAPPING_PROMPT_FUNCS = { + "zed-industries/zeta": _format_zeta_prompt, + } + + def sample(self, tokenizer: PreTrainedTokenizerBase, num_requests: int, **kwargs): + formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.dataset_path) + if formatting_prompt_func is None: + raise ValueError(f"Unsupported dataset path: {self.dataset_path}") + samples = [] + for sample in self.data: + sample = formatting_prompt_func(sample) + samples.append( + SampleRequest( + prompt=sample["prompt"], + prompt_len=len(tokenizer(sample["prompt"]).input_ids), + expected_output_len=len( + tokenizer(sample["expected_output"]).input_ids + ), + ) + ) + if len(samples) >= num_requests: + break + self.maybe_oversample_requests(samples, num_requests) + return samples + + +# ----------------------------------------------------------------------------- +# ASR Dataset Implementation +# ----------------------------------------------------------------------------- + + +class ASRDataset(HuggingFaceDataset): + """ + Dataset class for processing a ASR dataset for transcription. + Tested on the following set: + + +----------------+----------------------------------------+--------------------------+-----------------------------+ + | Dataset | Domain | Speaking Style | hf-subset | + +----------------+----------------------------------------+--------------------------+-----------------------------+ + | TED-LIUM | TED talks | Oratory | release1, release2, release3| + | | | | release3-speaker-adaptation | + | VoxPopuli | European Parliament | Oratory | en, de, it, fr, ... | + | LibriSpeech | Audiobook | Narrated | "LIUM/tedlium" | + | GigaSpeech | Audiobook, podcast, YouTube | Narrated, spontaneous | xs, s, m, l, xl, dev, test | + | SPGISpeech | Financial meetings | Oratory, spontaneous | S, M, L, dev, test | + | AMI | Meetings | Spontaneous | ihm, sdm | + +----------------+----------------------------------------+--------------------------+-----------------------------+ + + """ # noqa: E501 + + SUPPORTED_DATASET_PATHS = { + "openslr/librispeech_asr", + "facebook/voxpopuli", + "LIUM/tedlium", + "edinburghcstr/ami", + "speechcolab/gigaspeech", + "kensho/spgispeech", + } + + DEFAULT_OUTPUT_LEN = 128 + IS_MULTIMODAL = True + + # TODO Whisper-specific. Abstract interface when more models are supported. + TRANSCRIPTION_PREAMBLE = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>" + skip_long_audios: bool = True + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + **kwargs, + ) -> list: + import librosa + + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + prompt = ASRDataset.TRANSCRIPTION_PREAMBLE + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests = [] + skipped = 0 + for item in self.data: + if len(sampled_requests) >= num_requests: + break + audio = item["audio"] + y, sr = audio["array"], audio["sampling_rate"] + duration_s = librosa.get_duration(y=y, sr=sr) + # Whisper max supported duration + if self.skip_long_audios and duration_s > 30: + skipped += 1 + continue + + mm_content = {"audio": (y, sr)} + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=mm_content, + ) + ) + if skipped: + logger.warning( + "%d samples discarded from dataset due to" + " their length being greater than" + " what Whisper supports.", + skipped, + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests diff --git a/online_apiserver_test/benchmarks/benchmark_latency.py b/online_apiserver_test/benchmarks/benchmark_latency.py new file mode 100644 index 0000000000000000000000000000000000000000..4d2ea126b24a514bde2e7e3808d2120ede353cc7 --- /dev/null +++ b/online_apiserver_test/benchmarks/benchmark_latency.py @@ -0,0 +1,186 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Benchmark the latency of processing a single batch of requests.""" + +import argparse +import dataclasses +import json +import os +import time +from typing import Any, Optional + +import numpy as np +from tqdm import tqdm + +import vllm.envs as envs +from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json +from vllm import LLM, SamplingParams +from vllm.engine.arg_utils import EngineArgs +from vllm.inputs import PromptType +from vllm.sampling_params import BeamSearchParams +from vllm.utils import FlexibleArgumentParser + + +def save_to_pytorch_benchmark_format( + args: argparse.Namespace, results: dict[str, Any] +) -> None: + pt_records = convert_to_pytorch_benchmark_format( + args=args, + metrics={"latency": results["latencies"]}, + extra_info={k: results[k] for k in ["avg_latency", "percentiles"]}, + ) + if pt_records: + pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json" + write_to_json(pt_file, pt_records) + + +def main(args: argparse.Namespace): + print(args) + + engine_args = EngineArgs.from_cli_args(args) + + # NOTE(woosuk): If the request cannot be processed in a single batch, + # the engine will automatically process the request in multiple batches. + llm = LLM(**dataclasses.asdict(engine_args)) + assert llm.llm_engine.model_config.max_model_len >= ( + args.input_len + args.output_len + ), ( + "Please ensure that max_model_len is greater than" + " the sum of input_len and output_len." + ) + + sampling_params = SamplingParams( + n=args.n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=args.output_len, + detokenize=not args.disable_detokenize, + ) + print(sampling_params) + dummy_prompt_token_ids = np.random.randint( + 10000, size=(args.batch_size, args.input_len) + ) + dummy_prompts: list[PromptType] = [ + {"prompt_token_ids": batch} for batch in dummy_prompt_token_ids.tolist() + ] + + def llm_generate(): + if not args.use_beam_search: + llm.generate(dummy_prompts, sampling_params=sampling_params, use_tqdm=False) + else: + llm.beam_search( + dummy_prompts, + BeamSearchParams( + beam_width=args.n, + max_tokens=args.output_len, + ignore_eos=True, + ), + ) + + def run_to_completion(profile_dir: Optional[str] = None): + if profile_dir: + llm.start_profile() + llm_generate() + llm.stop_profile() + else: + start_time = time.perf_counter() + llm_generate() + end_time = time.perf_counter() + latency = end_time - start_time + return latency + + print("Warming up...") + for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"): + run_to_completion(profile_dir=None) + + if args.profile: + profile_dir = envs.VLLM_TORCH_PROFILER_DIR + print(f"Profiling (results will be saved to '{profile_dir}')...") + run_to_completion(profile_dir=profile_dir) + return + + # Benchmark. + latencies = [] + for _ in tqdm(range(args.num_iters), desc="Profiling iterations"): + latencies.append(run_to_completion(profile_dir=None)) + latencies = np.array(latencies) + percentages = [10, 25, 50, 75, 90, 99] + percentiles = np.percentile(latencies, percentages) + print(f"Avg latency: {np.mean(latencies)} seconds") + for percentage, percentile in zip(percentages, percentiles): + print(f"{percentage}% percentile latency: {percentile} seconds") + + # Output JSON results if specified + if args.output_json: + results = { + "avg_latency": np.mean(latencies), + "latencies": latencies.tolist(), + "percentiles": dict(zip(percentages, percentiles.tolist())), + } + with open(args.output_json, "w") as f: + json.dump(results, f, indent=4) + save_to_pytorch_benchmark_format(args, results) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the latency of processing a single batch of " + "requests till completion." + ) + parser.add_argument("--input-len", type=int, default=32) + parser.add_argument("--output-len", type=int, default=128) + parser.add_argument("--batch-size", type=int, default=8) + parser.add_argument( + "--n", + type=int, + default=1, + help="Number of generated sequences per prompt.", + ) + parser.add_argument("--use-beam-search", action="store_true") + parser.add_argument( + "--num-iters-warmup", + type=int, + default=10, + help="Number of iterations to run for warmup.", + ) + parser.add_argument( + "--num-iters", type=int, default=30, help="Number of iterations to run." + ) + parser.add_argument( + "--profile", + action="store_true", + help="profile the generation process of a single batch", + ) + parser.add_argument( + "--output-json", + type=str, + default=None, + help="Path to save the latency results in JSON format.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize responses (i.e. do not include " + "detokenization time in the latency measurement)" + ), + ) + + parser = EngineArgs.add_cli_args(parser) + # V1 enables prefix caching by default which skews the latency + # numbers. We need to disable prefix caching by default. + parser.set_defaults(enable_prefix_caching=False) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + if args.profile and not envs.VLLM_TORCH_PROFILER_DIR: + raise OSError( + "The environment variable 'VLLM_TORCH_PROFILER_DIR' is not set. " + "Please set it to a valid path to use torch profiler." + ) + main(args) diff --git a/online_apiserver_test/benchmarks/benchmark_long_document_qa_throughput.py b/online_apiserver_test/benchmarks/benchmark_long_document_qa_throughput.py new file mode 100644 index 0000000000000000000000000000000000000000..6e0f3b51c9d284d22dd1acdc1b794b9a5c0f31b1 --- /dev/null +++ b/online_apiserver_test/benchmarks/benchmark_long_document_qa_throughput.py @@ -0,0 +1,202 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +Offline benchmark to test the long document QA throughput. + +Example usage: + # This workload samples 8 different prompts with a default input + # length of 20000 tokens, then replicates each prompt 2 times + # in random order. + python benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --repeat-count 2 + +Commandline arguments: + --num-documents: The number of documents to sample prompts from. + + --document-length: The length of each document in tokens. + (Optional, default: 20000) + + --output-len: The number of tokens to generate for each prompt. + (Optional, default: 10) + + --repeat-count: The number of times to repeat each prompt. + (Optional, default: 2) + + --repeat-mode: The mode to repeat prompts. The supported modes are: + - 'random': shuffle the prompts randomly. (Default) + - 'tile': the entire prompt list is repeated in sequence. (Potentially + lowest cache hit) + - 'interleave': each prompt is repeated consecutively before + moving to the next element. (Highest cache hit) + + --shuffle-seed: Random seed when the repeat mode is "random". + (Optional, default: 0) + +In the meantime, it also supports all the vLLM engine args to initialize the +LLM engine. You can refer to the `vllm.engine.arg_utils.EngineArgs` for more +details. +""" + +import dataclasses +import random +import time + +from vllm import LLM, SamplingParams +from vllm.engine.arg_utils import EngineArgs +from vllm.utils import FlexibleArgumentParser + + +def test_long_document_qa(llm=None, sampling_params=None, prompts=None): + """ + Test long document QA with the given prompts and sampling parameters. + Print the time spent in processing all the prompts. + + Args: + llm: The language model used for generating responses. + sampling_params: Sampling parameter used to generate the response. + prompts: A list of prompt strings to be processed by the LLM. + """ + start_time = time.time() + llm.generate(prompts, sampling_params=sampling_params) + end_time = time.time() + print(f"Time to execute all requests: {end_time - start_time:.4f} secs") + + +def repeat_prompts(prompts, repeat_count, mode: str): + """ + Repeat each prompt in the list for a specified number of times. + The order of prompts in the output list depends on the mode. + + Args: + prompts: A list of prompts to be repeated. + repeat_count: The number of times each prompt is repeated. + mode: The mode of repetition. Supported modes are: + - 'random': Shuffle the prompts randomly after repetition. + - 'tile': Repeat the entire prompt list in sequence. + Example: [1, 2, 3] -> [1, 2, 3, 1, 2, 3]. + - 'interleave': Repeat each prompt consecutively before moving to + the next. Example: [1, 2, 3] -> [1, 1, 2, 2, 3, 3]. + + Returns: + A list of repeated prompts in the specified order. + + Raises: + ValueError: If an invalid mode is provided. + """ + print("Repeat mode: ", mode) + if mode == "random": + repeated_prompts = prompts * repeat_count + random.shuffle(repeated_prompts) + return repeated_prompts + elif mode == "tile": + return prompts * repeat_count + elif mode == "interleave": + repeated_prompts = [] + for prompt in prompts: + repeated_prompts.extend([prompt] * repeat_count) + return repeated_prompts + else: + raise ValueError( + f"Invalid mode: {mode}, only support 'random', 'tile', 'interleave'" + ) + + +def main(args): + random.seed(args.shuffle_seed) + + # Prepare the prompts: + # we append the document id at the beginning to avoid any of the document + # being the prefix of other documents + prompts = [ + str(i) + " ".join(["hi"] * args.document_length) + for i in range(args.num_documents) + ] + + prompts = repeat_prompts(prompts, args.repeat_count, mode=args.repeat_mode) + + warmup_prompts = [ + "This is warm up request " + str(i) + " ".join(["hi"] * args.document_length) + for i in range(args.num_documents) + ] + + # Create the LLM engine + engine_args = EngineArgs.from_cli_args(args) + llm = LLM(**dataclasses.asdict(engine_args)) + sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len) + + print("------warm up------") + test_long_document_qa( + llm=llm, + prompts=warmup_prompts, + sampling_params=sampling_params, + ) + + print("------start generating------") + test_long_document_qa( + llm=llm, + prompts=prompts, + sampling_params=sampling_params, + ) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the performance with or " + "without automatic prefix caching." + ) + + parser.add_argument( + "--document-length", + type=int, + # Roughly the number of tokens for a system paper, + # excluding images + default=20000, + help="Range of input lengths for sampling prompts, " + 'specified as "min:max" (e.g., "128:256").', + ) + + parser.add_argument( + "--num-documents", + type=int, + default=8, + help="Range of input lengths for sampling prompts, " + 'specified as "min:max" (e.g., "128:256").', + ) + + parser.add_argument("--output-len", type=int, default=10) + + parser.add_argument( + "--repeat-count", + type=int, + default=2, + help="Number of times to repeat each prompt", + ) + + parser.add_argument( + "--repeat-mode", + type=str, + default="random", + help="The mode to repeat prompts. The supported " + 'modes are "random", "tile", and "interleave". ' + "See repeat_prompts() in the source code for details.", + ) + + parser.add_argument( + "--shuffle-seed", + type=int, + default=0, + help='Random seed when the repeat mode is "random"', + ) + + parser = EngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test/benchmarks/benchmark_prefix_caching.py b/online_apiserver_test/benchmarks/benchmark_prefix_caching.py new file mode 100644 index 0000000000000000000000000000000000000000..b5e2613de1cd4a0a2557cf09baa458aa15a6049c --- /dev/null +++ b/online_apiserver_test/benchmarks/benchmark_prefix_caching.py @@ -0,0 +1,278 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +Benchmark the efficiency of prefix caching. + +This script allows you to benchmark the performance of +a model with and without prefix caching using either fixed prompts +or prompts sampled from the ShareGPT dataset. + +Fixed example usage: + python benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-prompts 1 \ + --repeat-count 100 \ + --input-length-range 128:256 + +ShareGPT example usage: + # This command samples 20 prompts with input lengths + # between 128 and 256 tokens from the ShareGPT dataset, + # then replicates each prompt 5 times. + python benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json \ + --enable-prefix-caching \ + --num-prompts 20 \ + --repeat-count 5 \ + --input-length-range 128:256 +""" + +import dataclasses +import json +import random +import time +from typing import Optional + +from transformers import PreTrainedTokenizerBase + +from vllm import LLM, SamplingParams +from vllm.engine.arg_utils import EngineArgs +from vllm.utils import FlexibleArgumentParser + +try: + from vllm.transformers_utils.tokenizer import get_tokenizer +except ImportError: + from backend_request_func import get_tokenizer + +PROMPT = "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as fellows. You need to answer my question about the table.\n# Table\n|Opening|Opening|Sl. No.|Film|Cast|Director|Music Director|Notes|\n|----|----|----|----|----|----|----|----|\n|J A N|9|1|Agni Pushpam|Jayabharathi, Kamalahasan|Jeassy|M. K. Arjunan||\n|J A N|16|2|Priyamvada|Mohan Sharma, Lakshmi, KPAC Lalitha|K. S. Sethumadhavan|V. Dakshinamoorthy||\n|J A N|23|3|Yakshagaanam|Madhu, Sheela|Sheela|M. S. Viswanathan||\n|J A N|30|4|Paalkkadal|Sheela, Sharada|T. K. Prasad|A. T. Ummer||\n|F E B|5|5|Amma|Madhu, Srividya|M. Krishnan Nair|M. K. Arjunan||\n|F E B|13|6|Appooppan|Thikkurissi Sukumaran Nair, Kamal Haasan|P. Bhaskaran|M. S. Baburaj||\n|F E B|20|7|Srishti|Chowalloor Krishnankutty, Ravi Alummoodu|K. T. Muhammad|M. S. Baburaj||\n|F E B|20|8|Vanadevatha|Prem Nazir, Madhubala|Yusufali Kechery|G. Devarajan||\n|F E B|27|9|Samasya|Madhu, Kamalahaasan|K. Thankappan|Shyam||\n|F E B|27|10|Yudhabhoomi|K. P. Ummer, Vidhubala|Crossbelt Mani|R. K. Shekhar||\n|M A R|5|11|Seemantha Puthran|Prem Nazir, Jayabharathi|A. B. Raj|M. K. Arjunan||\n|M A R|12|12|Swapnadanam|Rani Chandra, Dr. Mohandas|K. G. George|Bhaskar Chandavarkar||\n|M A R|19|13|Thulavarsham|Prem Nazir, sreedevi, Sudheer|N. Sankaran Nair|V. Dakshinamoorthy||\n|M A R|20|14|Aruthu|Kaviyoor Ponnamma, Kamalahasan|Ravi|G. Devarajan||\n|M A R|26|15|Swimming Pool|Kamal Haasan, M. G. Soman|J. Sasikumar|M. K. Arjunan||\n\n# Question\nWhat' s the content in the (1,1) cells\n" # noqa: E501 + + +def test_prefix(llm=None, sampling_params=None, prompts=None): + start_time = time.time() + + llm.generate(prompts, sampling_params=sampling_params) + + end_time = time.time() + print(f"cost time {end_time - start_time}") + + +@dataclasses.dataclass +class Request: + prompt: str + prompt_len: int + output_len: int + + +def sample_tokens(tokenizer: PreTrainedTokenizerBase, length: int) -> list[int]: + vocab = tokenizer.get_vocab() + all_special_ids = set(tokenizer.all_special_ids) + + # Remove the special tokens. + return random.choices( + [v for k, v in vocab.items() if k not in all_special_ids], + k=length, + ) + + +def sample_requests_from_dataset( + dataset_path: str, + num_requests: int, + tokenizer: PreTrainedTokenizerBase, + input_length_range: tuple[int, int], + fixed_output_len: Optional[int], +) -> list[Request]: + if fixed_output_len is not None and fixed_output_len < 4: + raise ValueError("output_len too small") + + # Load the dataset. + with open(dataset_path) as f: + dataset = json.load(f) + # Filter out the conversations with less than 2 turns. + dataset = [data for data in dataset if len(data["conversations"]) >= 2] + # Only keep the first two turns of each conversation. + dataset = [ + (data["conversations"][0]["value"], data["conversations"][1]["value"]) + for data in dataset + ] + + # Shuffle the dataset. + random.shuffle(dataset) + + min_len, max_len = input_length_range + assert min_len >= 0 and max_len >= min_len, "input_length_range too small" + + # Filter out sequences that are too long or too short + filtered_requests: list[Request] = [] + + for i in range(len(dataset)): + if len(filtered_requests) == num_requests: + break + + # Tokenize the prompts and completions. + prompt_token_ids = tokenizer(dataset[i][0]).input_ids + prompt = tokenizer.decode(prompt_token_ids) + completion = dataset[i][1] + completion_token_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_token_ids) + output_len = ( + len(completion_token_ids) if fixed_output_len is None else fixed_output_len + ) + if min_len <= prompt_len <= max_len: + filtered_requests.append(Request(prompt, prompt_len, output_len)) + + return filtered_requests + + +def sample_requests_from_random( + num_requests: int, + tokenizer: PreTrainedTokenizerBase, + input_length_range: tuple[int, int], + fixed_output_len: Optional[int], + prefix_len: int, +) -> list[Request]: + requests = [] + prefix_token_ids = sample_tokens(tokenizer, prefix_len) + min_len, max_len = input_length_range + + for i in range(num_requests): + unique_part_token_ids = sample_tokens( + tokenizer, random.randint(min_len - prefix_len, max_len - prefix_len) + ) + prompt_token_ids = prefix_token_ids + unique_part_token_ids + prompt = tokenizer.decode(prompt_token_ids) + prompt_len = len(prompt_token_ids) + assert min_len <= prompt_len <= max_len, ( + f"prompt_len {prompt_len} out of range {min_len}:{max_len}" + ) + requests.append(Request(prompt, prompt_len, fixed_output_len)) + return requests + + +def repeat_and_sort_requests( + requests: list[Request], repeat_count: int, sort: bool = False +) -> list[str]: + repeated_requests = requests * repeat_count + if sort: + repeated_requests.sort(key=lambda x: x[1]) + else: + random.shuffle(repeated_requests) + return [req.prompt for req in repeated_requests] + + +def main(args): + tokenizer = get_tokenizer(args.model, trust_remote_code=True) + input_length_range = tuple(map(int, args.input_length_range.split(":"))) + random.seed(args.seed) + if args.dataset_path is not None: + if args.prefix_len > 0: + raise ValueError( + "prefix-len is not supported when dataset-path is provided." + ) + print(f"Start to sample {args.num_prompts} prompts from {args.dataset_path}") + filtered_requests = sample_requests_from_dataset( + dataset_path=args.dataset_path, + num_requests=args.num_prompts, + tokenizer=tokenizer, + input_length_range=input_length_range, + fixed_output_len=args.output_len, + ) + else: + print(f"Start to sample {args.num_prompts} prompts from random") + filtered_requests = sample_requests_from_random( + num_requests=args.num_prompts, + tokenizer=tokenizer, + input_length_range=input_length_range, + fixed_output_len=args.output_len, + prefix_len=args.prefix_len, + ) + + # Print some helpful stats of the requests. + print(f"Sampled {len(filtered_requests)} requests.") + prompt_lens = [req.prompt_len for req in filtered_requests] + print(f"Average input length: {sum(prompt_lens) / len(prompt_lens)}") + print(f"P50 input length: {sorted(prompt_lens)[len(prompt_lens) // 2]}") + print(f"Min Prompt Length: {min(prompt_lens)}") + print(f"Max Prompt Length: {max(prompt_lens)}") + + engine_args = EngineArgs.from_cli_args(args) + + llm = LLM(**dataclasses.asdict(engine_args)) + + sampling_params = SamplingParams( + temperature=0, + max_tokens=args.output_len, + detokenize=not args.disable_detokenize, + ) + + print("Testing filtered requests") + prompts = repeat_and_sort_requests( + filtered_requests, repeat_count=args.repeat_count, sort=args.sort + ) + + print("------start generating------") + test_prefix( + llm=llm, + prompts=prompts, + sampling_params=sampling_params, + ) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the performance with or without " + "automatic prefix caching." + ) + parser.add_argument( + "--dataset-path", type=str, default=None, help="Path to the dataset." + ) + parser.add_argument("--output-len", type=int, default=10) + parser.add_argument( + "--num-prompts", + type=int, + required=True, + help="Number of the prompts sampled from dataset", + ) + parser.add_argument( + "--repeat-count", + type=int, + default=1, + help="Number of times to repeat each prompt", + ) + parser.add_argument( + "--sort", action="store_true", help="Sort prompts by input length" + ) + parser.add_argument( + "--input-length-range", + type=str, + required=True, + help="Range of input lengths for sampling prompts," + 'specified as "min:max" (e.g., "128:256").', + ) + parser.add_argument( + "--prefix-len", + type=int, + default=0, + help="Specifies the length of a common prefix to be " + "added to the input prompt. The input-length-range will " + "subtract this length when filtering prompts. Only used " + "when dataset-path is not provided.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize responses (i.e. do not include " + "detokenization time in the latency measurement)" + ), + ) + + parser = EngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test/benchmarks/benchmark_prioritization.py b/online_apiserver_test/benchmarks/benchmark_prioritization.py new file mode 100644 index 0000000000000000000000000000000000000000..bb453791c1862bdd0883f0a47af6fddeef764f8f --- /dev/null +++ b/online_apiserver_test/benchmarks/benchmark_prioritization.py @@ -0,0 +1,222 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Benchmark offline prioritization.""" + +import argparse +import dataclasses +import json +import random +import time +from typing import Optional + +from transformers import AutoTokenizer, PreTrainedTokenizerBase + +from vllm.engine.arg_utils import EngineArgs +from vllm.utils import FlexibleArgumentParser + + +# Select a equi-probable random priority +def get_random_flag(): + return 0 if random.random() < 0.5 else 1 + + +def sample_requests( + dataset_path: str, + num_requests: int, + tokenizer: PreTrainedTokenizerBase, + fixed_output_len: Optional[int], +) -> list[tuple[str, int, int, int]]: + if fixed_output_len is not None and fixed_output_len < 4: + raise ValueError("output_len too small") + + # Load the dataset. + with open(dataset_path) as f: + dataset = json.load(f) + # Filter out the conversations with less than 2 turns. + dataset = [data for data in dataset if len(data["conversations"]) >= 2] + # Only keep the first two turns of each conversation. + dataset = [ + (data["conversations"][0]["value"], data["conversations"][1]["value"]) + for data in dataset + ] + + # Shuffle the dataset. + random.shuffle(dataset) + + # Filter out sequences that are too long or too short + filtered_dataset: list[tuple[str, int, int]] = [] + for i in range(len(dataset)): + if len(filtered_dataset) == num_requests: + break + + # Tokenize the prompts and completions. + prompt = dataset[i][0] + prompt_token_ids = tokenizer(prompt).input_ids + completion = dataset[i][1] + completion_token_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_token_ids) + output_len = ( + len(completion_token_ids) if fixed_output_len is None else fixed_output_len + ) + if prompt_len < 4 or output_len < 4: + # Prune too short sequences. + continue + if prompt_len > 1024 or prompt_len + output_len > 2048: + # Prune too long sequences. + continue + + priority = get_random_flag() + + filtered_dataset.append((prompt, prompt_len, output_len, priority)) + + return filtered_dataset + + +def run_vllm( + requests: list[tuple[str, int, int]], + n: int, + engine_args: EngineArgs, + disable_detokenize: bool = False, +) -> float: + from vllm import LLM, SamplingParams + + llm = LLM(**dataclasses.asdict(engine_args)) + + assert all( + llm.llm_engine.model_config.max_model_len >= (request[1] + request[2]) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of" + " input_len and output_len for all requests." + ) + + # Add the requests to the engine. + prompts = [] + sampling_params = [] + priority = [] + for prompt, _, output_len, _priority in requests: + prompts.append(prompt) + priority.append(_priority) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=output_len, + detokenize=not disable_detokenize, + ) + ) + + start = time.perf_counter() + llm.generate(prompts, sampling_params, priority=priority, use_tqdm=True) + end = time.perf_counter() + return end - start + + +def main(args: argparse.Namespace): + print(args) + random.seed(args.seed) + + # Sample the requests. + tokenizer = AutoTokenizer.from_pretrained( + args.tokenizer, trust_remote_code=args.trust_remote_code + ) + if args.dataset is None: + # Synthesize a prompt with the given input length. + prompt = "hi" * (args.input_len - 1) + requests = [ + (prompt, args.input_len, args.output_len, get_random_flag()) + for _ in range(args.num_prompts) + ] + else: + requests = sample_requests( + args.dataset, args.num_prompts, tokenizer, args.output_len + ) + + if args.backend == "vllm": + elapsed_time = run_vllm( + requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize + ) + else: + raise ValueError(f"Unknown backend: {args.backend}") + total_num_tokens = sum( + prompt_len + output_len for _, prompt_len, output_len, priority in requests + ) + print( + f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, " + f"{total_num_tokens / elapsed_time:.2f} tokens/s" + ) + + # Output JSON results if specified + if args.output_json: + results = { + "elapsed_time": elapsed_time, + "num_requests": len(requests), + "total_num_tokens": total_num_tokens, + "requests_per_second": len(requests) / elapsed_time, + "tokens_per_second": total_num_tokens / elapsed_time, + } + with open(args.output_json, "w") as f: + json.dump(results, f, indent=4) + + +def create_argument_parser(): + parser = FlexibleArgumentParser(description="Benchmark the throughput.") + parser.add_argument( + "--backend", type=str, choices=["vllm", "hf", "mii"], default="vllm" + ) + parser.add_argument( + "--dataset", type=str, default=None, help="Path to the dataset." + ) + parser.add_argument( + "--input-len", + type=int, + default=None, + help="Input prompt length for each request", + ) + parser.add_argument( + "--output-len", + type=int, + default=None, + help="Output length for each request. Overrides the " + "output length from the dataset.", + ) + parser.add_argument( + "--n", type=int, default=1, help="Number of generated sequences per prompt." + ) + parser.add_argument( + "--num-prompts", type=int, default=200, help="Number of prompts to process." + ) + parser.add_argument( + "--output-json", + type=str, + default=None, + help="Path to save the throughput results in JSON format.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize responses (i.e. do not include " + "detokenization time in the latency measurement)" + ), + ) + + parser = EngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + if args.tokenizer is None: + args.tokenizer = args.model + if args.dataset is None: + assert args.input_len is not None + assert args.output_len is not None + else: + assert args.input_len is None + + main(args) diff --git a/online_apiserver_test/benchmarks/benchmark_serving.py b/online_apiserver_test/benchmarks/benchmark_serving.py new file mode 100644 index 0000000000000000000000000000000000000000..9b235266dff1aac95d6fe0509f4a5f01641579cc --- /dev/null +++ b/online_apiserver_test/benchmarks/benchmark_serving.py @@ -0,0 +1,1378 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +r"""Benchmark online serving throughput. + +On the server side, run one of the following commands: + vLLM OpenAI API server + vllm serve \ + --swap-space 16 \ + --disable-log-requests + +On the client side, run: + python benchmarks/benchmark_serving.py \ + --backend \ + --model \ + --dataset-name sharegpt \ + --dataset-path \ + --request-rate \ # By default is inf + --num-prompts # By default is 1000 + + when using tgi backend, add + --endpoint /generate_stream + to the end of the command above. +""" + +import argparse +import asyncio +import gc +import json +import os +import random +import time +import warnings +from collections.abc import AsyncGenerator, Iterable +from dataclasses import dataclass +from datetime import datetime +from typing import Any, Literal, Optional + +import numpy as np +from tqdm.asyncio import tqdm +from transformers import PreTrainedTokenizerBase + +from backend_request_func import ( + ASYNC_REQUEST_FUNCS, + OPENAI_COMPATIBLE_BACKENDS, + RequestFuncInput, + RequestFuncOutput, +) + +try: + from vllm.transformers_utils.tokenizer import get_tokenizer +except ImportError: + from backend_request_func import get_tokenizer + +try: + from vllm.utils import FlexibleArgumentParser +except ImportError: + from argparse import ArgumentParser as FlexibleArgumentParser + +from benchmark_dataset import ( + AIMODataset, + ASRDataset, + BurstGPTDataset, + ConversationDataset, + CustomDataset, + HuggingFaceDataset, + InstructCoderDataset, + MTBenchDataset, + NextEditPredictionDataset, + RandomDataset, + SampleRequest, + ShareGPTDataset, + SonnetDataset, + VisionArenaDataset, +) +from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json + +MILLISECONDS_TO_SECONDS_CONVERSION = 1000 + + +@dataclass +class BenchmarkMetrics: + completed: int + total_input: int + total_output: int + request_throughput: float + request_goodput: float + output_throughput: float + total_token_throughput: float + mean_ttft_ms: float + median_ttft_ms: float + std_ttft_ms: float + percentiles_ttft_ms: list[tuple[float, float]] + mean_tpot_ms: float + median_tpot_ms: float + std_tpot_ms: float + percentiles_tpot_ms: list[tuple[float, float]] + mean_itl_ms: float + median_itl_ms: float + std_itl_ms: float + percentiles_itl_ms: list[tuple[float, float]] + # E2EL stands for end-to-end latency per request. + # It is the time taken on the client side from sending + # a request to receiving a complete response. + mean_e2el_ms: float + median_e2el_ms: float + std_e2el_ms: float + percentiles_e2el_ms: list[tuple[float, float]] + + +def _get_current_request_rate( + ramp_up_strategy: Optional[Literal["linear", "exponential"]], + ramp_up_start_rps: Optional[int], + ramp_up_end_rps: Optional[int], + request_index: int, + total_requests: int, + request_rate: float, +) -> float: + if ( + ramp_up_strategy + and ramp_up_start_rps is not None + and ramp_up_end_rps is not None + ): + progress = request_index / max(total_requests - 1, 1) + if ramp_up_strategy == "linear": + increase = (ramp_up_end_rps - ramp_up_start_rps) * progress + return ramp_up_start_rps + increase + elif ramp_up_strategy == "exponential": + ratio = ramp_up_end_rps / ramp_up_start_rps + return ramp_up_start_rps * (ratio**progress) + else: + raise ValueError(f"Unknown ramp-up strategy: {ramp_up_strategy}") + return request_rate + + +async def get_request( + input_requests: list[SampleRequest], + request_rate: float, + burstiness: float = 1.0, + ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None, + ramp_up_start_rps: Optional[int] = None, + ramp_up_end_rps: Optional[int] = None, +) -> AsyncGenerator[tuple[SampleRequest, float], None]: + """ + Asynchronously generates requests at a specified rate + with OPTIONAL burstiness and OPTIONAL ramp-up strategy. + + Args: + input_requests: + A list of input requests, each represented as a SampleRequest. + request_rate: + The rate at which requests are generated (requests/s). + burstiness (optional): + The burstiness factor of the request generation. + Only takes effect when request_rate is not inf. + Default value is 1, which follows a Poisson process. + Otherwise, the request intervals follow a gamma distribution. + A lower burstiness value (0 < burstiness < 1) results + in more bursty requests, while a higher burstiness value + (burstiness > 1) results in a more uniform arrival of requests. + ramp_up_strategy (optional): + The ramp-up strategy. Can be "linear" or "exponential". + If None, uses constant request rate (specified by request_rate). + ramp_up_start_rps (optional): + The starting request rate for ramp-up. + ramp_up_end_rps (optional): + The ending request rate for ramp-up. + """ + assert burstiness > 0, ( + f"A positive burstiness factor is expected, but given {burstiness}." + ) + # Convert to list to get length for ramp-up calculations + if isinstance(input_requests, Iterable) and not isinstance(input_requests, list): + input_requests = list(input_requests) + + total_requests = len(input_requests) + request_index = 0 + + for request in input_requests: + current_request_rate = _get_current_request_rate( + ramp_up_strategy, + ramp_up_start_rps, + ramp_up_end_rps, + request_index, + total_requests, + request_rate, + ) + + yield request, current_request_rate + + request_index += 1 + + if current_request_rate == float("inf"): + # If the request rate is infinity, then we don't need to wait. + continue + + theta = 1.0 / (current_request_rate * burstiness) + + # Sample the request interval from the gamma distribution. + # If burstiness is 1, it follows exponential distribution. + interval = np.random.gamma(shape=burstiness, scale=theta) + # The next request will be sent after the interval. + await asyncio.sleep(interval) + + +def calculate_metrics( + input_requests: list[SampleRequest], + outputs: list[RequestFuncOutput], + dur_s: float, + tokenizer: PreTrainedTokenizerBase, + selected_percentile_metrics: list[str], + selected_percentiles: list[float], + goodput_config_dict: dict[str, float], +) -> tuple[BenchmarkMetrics, list[int]]: + actual_output_lens: list[int] = [] + total_input = 0 + completed = 0 + good_completed = 0 + itls: list[float] = [] + tpots: list[float] = [] + all_tpots: list[float] = [] + ttfts: list[float] = [] + e2els: list[float] = [] + for i in range(len(outputs)): + if outputs[i].success: + output_len = outputs[i].output_tokens + + if not output_len: + # We use the tokenizer to count the number of output tokens + # for some serving backends instead of looking at + # len(outputs[i].itl) since multiple output tokens may be + # bundled together + # Note : this may inflate the output token count slightly + output_len = len( + tokenizer( + outputs[i].generated_text, add_special_tokens=False + ).input_ids + ) + actual_output_lens.append(output_len) + total_input += input_requests[i].prompt_len + tpot = 0 + if output_len > 1: + latency_minus_ttft = outputs[i].latency - outputs[i].ttft + tpot = latency_minus_ttft / (output_len - 1) + tpots.append(tpot) + # Note: if output_len <= 1, we regard tpot as 0 for goodput + all_tpots.append(tpot) + itls += outputs[i].itl + ttfts.append(outputs[i].ttft) + e2els.append(outputs[i].latency) + completed += 1 + else: + actual_output_lens.append(0) + + if goodput_config_dict: + valid_metrics = [] + slo_values = [] + + if "ttft" in goodput_config_dict: + valid_metrics.append(ttfts) + slo_values.append( + goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "tpot" in goodput_config_dict: + valid_metrics.append(all_tpots) + slo_values.append( + goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "e2el" in goodput_config_dict: + valid_metrics.append(e2els) + slo_values.append( + goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + + for req_metric in zip(*valid_metrics): + is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)]) + if is_good_req: + good_completed += 1 + + if completed == 0: + warnings.warn( + "All requests failed. This is likely due to a misconfiguration " + "on the benchmark arguments.", + stacklevel=2, + ) + metrics = BenchmarkMetrics( + completed=completed, + total_input=total_input, + total_output=sum(actual_output_lens), + request_throughput=completed / dur_s, + request_goodput=good_completed / dur_s, + output_throughput=sum(actual_output_lens) / dur_s, + total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s, + mean_ttft_ms=np.mean(ttfts or 0) + * 1000, # ttfts is empty if streaming is not supported by backend + std_ttft_ms=np.std(ttfts or 0) * 1000, + median_ttft_ms=np.median(ttfts or 0) * 1000, + percentiles_ttft_ms=[ + (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles + ], + mean_tpot_ms=np.mean(tpots or 0) * 1000, + std_tpot_ms=np.std(tpots or 0) * 1000, + median_tpot_ms=np.median(tpots or 0) * 1000, + percentiles_tpot_ms=[ + (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles + ], + mean_itl_ms=np.mean(itls or 0) * 1000, + std_itl_ms=np.std(itls or 0) * 1000, + median_itl_ms=np.median(itls or 0) * 1000, + percentiles_itl_ms=[ + (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles + ], + mean_e2el_ms=np.mean(e2els or 0) * 1000, + std_e2el_ms=np.std(e2els or 0) * 1000, + median_e2el_ms=np.median(e2els or 0) * 1000, + percentiles_e2el_ms=[ + (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles + ], + ) + + return metrics, actual_output_lens + + +async def benchmark( + backend: str, + api_url: str, + base_url: str, + model_id: str, + model_name: str, + tokenizer: PreTrainedTokenizerBase, + input_requests: list[SampleRequest], + logprobs: Optional[int], + request_rate: float, + burstiness: float, + disable_tqdm: bool, + profile: bool, + selected_percentile_metrics: list[str], + selected_percentiles: list[float], + ignore_eos: bool, + goodput_config_dict: dict[str, float], + max_concurrency: Optional[int], + lora_modules: Optional[Iterable[str]], + extra_body: Optional[dict], + ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None, + ramp_up_start_rps: Optional[int] = None, + ramp_up_end_rps: Optional[int] = None, +): + if backend in ASYNC_REQUEST_FUNCS: + request_func = ASYNC_REQUEST_FUNCS[backend] + else: + raise ValueError(f"Unknown backend: {backend}") + + print("Starting initial single prompt test run...") + test_prompt, test_prompt_len, test_output_len, test_mm_content = ( + input_requests[0].prompt, + input_requests[0].prompt_len, + input_requests[0].expected_output_len, + input_requests[0].multi_modal_data, + ) + + assert test_mm_content is None or isinstance(test_mm_content, dict) + test_input = RequestFuncInput( + model=model_id, + model_name=model_name, + prompt=test_prompt, + api_url=api_url, + prompt_len=test_prompt_len, + output_len=test_output_len, + logprobs=logprobs, + multi_modal_content=test_mm_content, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + + test_output = await request_func(request_func_input=test_input) + if not test_output.success: + raise ValueError( + "Initial test run failed - Please make sure benchmark arguments " + f"are correctly specified. Error: {test_output.error}" + ) + else: + print("Initial test run completed. Starting main benchmark run...") + + if lora_modules: + # For each input request, choose a LoRA module at random. + lora_modules = iter( + [random.choice(lora_modules) for _ in range(len(input_requests))] + ) + + if profile: + print("Starting profiler...") + profile_input = RequestFuncInput( + model=model_id, + model_name=model_name, + prompt=test_prompt, + api_url=base_url + "/start_profile", + prompt_len=test_prompt_len, + output_len=test_output_len, + logprobs=logprobs, + multi_modal_content=test_mm_content, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler started") + + distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution" + + if ramp_up_strategy is not None: + print( + f"Traffic ramp-up strategy: {ramp_up_strategy}. Will increase " + f"RPS from {ramp_up_start_rps} to {ramp_up_end_rps} RPS over " + "the duration of the benchmark." + ) + else: + print(f"Traffic request rate: {request_rate} RPS.") + + print(f"Burstiness factor: {burstiness} ({distribution})") + print(f"Maximum request concurrency: {max_concurrency}") + + pbar = None if disable_tqdm else tqdm(total=len(input_requests)) + + # This can be used once the minimum Python version is 3.10 or higher, + # and it will simplify the code in limited_request_func. + # semaphore = (asyncio.Semaphore(max_concurrency) + # if max_concurrency else contextlib.nullcontext()) + semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None + + async def limited_request_func(request_func_input, pbar): + if semaphore is None: + return await request_func(request_func_input=request_func_input, pbar=pbar) + async with semaphore: + return await request_func(request_func_input=request_func_input, pbar=pbar) + + benchmark_start_time = time.perf_counter() + tasks: list[asyncio.Task] = [] + + rps_change_events = [] + last_int_rps = -1 + if ramp_up_strategy is not None and ramp_up_start_rps is not None: + last_int_rps = ramp_up_start_rps + rps_change_events.append( + { + "rps": last_int_rps, + "timestamp": datetime.now().isoformat(), + } + ) + + async for request, current_request_rate in get_request( + input_requests, + request_rate, + burstiness, + ramp_up_strategy, + ramp_up_start_rps, + ramp_up_end_rps, + ): + if ramp_up_strategy is not None: + current_int_rps = int(current_request_rate) + if current_int_rps > last_int_rps: + timestamp = datetime.now().isoformat() + for rps_val in range(last_int_rps + 1, current_int_rps + 1): + rps_change_events.append({"rps": rps_val, "timestamp": timestamp}) + last_int_rps = current_int_rps + + prompt, prompt_len, output_len, mm_content = ( + request.prompt, + request.prompt_len, + request.expected_output_len, + request.multi_modal_data, + ) + req_model_id, req_model_name = model_id, model_name + if lora_modules: + req_lora_module = next(lora_modules) + req_model_id, req_model_name = req_lora_module, req_lora_module + + request_func_input = RequestFuncInput( + model=req_model_id, + model_name=req_model_name, + prompt=prompt, + api_url=api_url, + prompt_len=prompt_len, + output_len=output_len, + logprobs=logprobs, + multi_modal_content=mm_content, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + task = limited_request_func(request_func_input=request_func_input, pbar=pbar) + tasks.append(asyncio.create_task(task)) + outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks) + + if profile: + print("Stopping profiler...") + profile_input = RequestFuncInput( + model=model_id, + prompt=test_prompt, + api_url=base_url + "/stop_profile", + prompt_len=test_prompt_len, + output_len=test_output_len, + logprobs=logprobs, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler stopped") + + if pbar is not None: + pbar.close() + + benchmark_duration = time.perf_counter() - benchmark_start_time + + metrics, actual_output_lens = calculate_metrics( + input_requests=input_requests, + outputs=outputs, + dur_s=benchmark_duration, + tokenizer=tokenizer, + selected_percentile_metrics=selected_percentile_metrics, + selected_percentiles=selected_percentiles, + goodput_config_dict=goodput_config_dict, + ) + + print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="=")) + print("{:<40} {:<10}".format("Successful requests:", metrics.completed)) + print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration)) + print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input)) + print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output)) + print( + "{:<40} {:<10.2f}".format( + "Request throughput (req/s):", metrics.request_throughput + ) + ) + if goodput_config_dict: + print( + "{:<40} {:<10.2f}".format( + "Request goodput (req/s):", metrics.request_goodput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Output token throughput (tok/s):", metrics.output_throughput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Total Token throughput (tok/s):", metrics.total_token_throughput + ) + ) + + result = { + "duration": benchmark_duration, + "completed": metrics.completed, + "total_input_tokens": metrics.total_input, + "total_output_tokens": metrics.total_output, + "request_throughput": metrics.request_throughput, + "request_goodput": metrics.request_goodput if goodput_config_dict else None, + "output_throughput": metrics.output_throughput, + "total_token_throughput": metrics.total_token_throughput, + "input_lens": [output.prompt_len for output in outputs], + "output_lens": actual_output_lens, + "ttfts": [output.ttft for output in outputs], + "itls": [output.itl for output in outputs], + "generated_texts": [output.generated_text for output in outputs], + "errors": [output.error for output in outputs], + } + + if rps_change_events: + result["rps_change_events"] = rps_change_events + + def process_one_metric( + # E.g., "ttft" + metric_attribute_name: str, + # E.g., "TTFT" + metric_name: str, + # E.g., "Time to First Token" + metric_header: str, + ): + # This function prints and adds statistics of the specified + # metric. + if metric_attribute_name not in selected_percentile_metrics: + return + print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-")) + print( + "{:<40} {:<10.2f}".format( + f"Mean {metric_name} (ms):", + getattr(metrics, f"mean_{metric_attribute_name}_ms"), + ) + ) + print( + "{:<40} {:<10.2f}".format( + f"Median {metric_name} (ms):", + getattr(metrics, f"median_{metric_attribute_name}_ms"), + ) + ) + result[f"mean_{metric_attribute_name}_ms"] = getattr( + metrics, f"mean_{metric_attribute_name}_ms" + ) + result[f"median_{metric_attribute_name}_ms"] = getattr( + metrics, f"median_{metric_attribute_name}_ms" + ) + result[f"std_{metric_attribute_name}_ms"] = getattr( + metrics, f"std_{metric_attribute_name}_ms" + ) + for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"): + p_word = str(int(p)) if int(p) == p else str(p) + print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value)) + result[f"p{p_word}_{metric_attribute_name}_ms"] = value + + process_one_metric("ttft", "TTFT", "Time to First Token") + process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)") + process_one_metric("itl", "ITL", "Inter-token Latency") + process_one_metric("e2el", "E2EL", "End-to-end Latency") + + print("=" * 50) + + return result + + +def check_goodput_args(args): + # Check and parse goodput arguments + goodput_config_dict = {} + VALID_NAMES = ["ttft", "tpot", "e2el"] + if args.goodput: + goodput_config_dict = parse_goodput(args.goodput) + for slo_name, slo_val in goodput_config_dict.items(): + if slo_name not in VALID_NAMES: + raise ValueError( + f"Invalid metric name found, {slo_name}: {slo_val}. " + "The service level objective name should be one of " + f"{str(VALID_NAMES)}. " + ) + if slo_val < 0: + raise ValueError( + f"Invalid value found, {slo_name}: {slo_val}. " + "The service level objective value should be " + "non-negative." + ) + return goodput_config_dict + + +def parse_goodput(slo_pairs): + goodput_config_dict = {} + try: + for slo_pair in slo_pairs: + slo_name, slo_val = slo_pair.split(":") + goodput_config_dict[slo_name] = float(slo_val) + except ValueError as err: + raise argparse.ArgumentTypeError( + "Invalid format found for service level objectives. " + 'Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is a " + "number in milliseconds." + ) from err + return goodput_config_dict + + +def save_to_pytorch_benchmark_format( + args: argparse.Namespace, results: dict[str, Any], file_name: str +) -> None: + metrics = [ + "median_ttft_ms", + "mean_ttft_ms", + "std_ttft_ms", + "p99_ttft_ms", + "mean_tpot_ms", + "median_tpot_ms", + "std_tpot_ms", + "p99_tpot_ms", + "median_itl_ms", + "mean_itl_ms", + "std_itl_ms", + "p99_itl_ms", + ] + # These raw data might be useful, but they are rather big. They can be added + # later if needed + ignored_metrics = ["ttfts", "itls", "generated_texts", "errors"] + pt_records = convert_to_pytorch_benchmark_format( + args=args, + metrics={k: [results[k]] for k in metrics}, + extra_info={ + k: results[k] + for k in results + if k not in metrics and k not in ignored_metrics + }, + ) + if pt_records: + # Don't use json suffix here as we don't want CI to pick it up + pt_file = f"{os.path.splitext(file_name)[0]}.pytorch.json" + write_to_json(pt_file, pt_records) + + +def main(args: argparse.Namespace): + print(args) + random.seed(args.seed) + np.random.seed(args.seed) + + backend = args.backend + model_id = args.model + model_name = args.served_model_name + tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model + tokenizer_mode = args.tokenizer_mode + + # Validate ramp-up arguments + if args.ramp_up_strategy is not None: + if args.request_rate != float("inf"): + raise ValueError( + "When using ramp-up, do not specify --request-rate. " + "The request rate will be controlled by ramp-up parameters. " + "Please remove the --request-rate argument." + ) + if args.ramp_up_start_rps is None or args.ramp_up_end_rps is None: + raise ValueError( + "When using --ramp-up-strategy, both --ramp-up-start-rps and " + "--ramp-up-end-rps must be specified" + ) + if args.ramp_up_start_rps < 0 or args.ramp_up_end_rps < 0: + raise ValueError("Ramp-up start and end RPS must be non-negative") + if args.ramp_up_start_rps > args.ramp_up_end_rps: + raise ValueError("Ramp-up start RPS must be less than end RPS") + if args.ramp_up_strategy == "exponential" and args.ramp_up_start_rps == 0: + raise ValueError("For exponential ramp-up, the start RPS cannot be 0.") + + if args.base_url is not None: + api_url = f"{args.base_url}{args.endpoint}" + base_url = f"{args.base_url}" + else: + api_url = f"http://{args.host}:{args.port}{args.endpoint}" + base_url = f"http://{args.host}:{args.port}" + + tokenizer = get_tokenizer( + tokenizer_id, + tokenizer_mode=tokenizer_mode, + trust_remote_code=args.trust_remote_code, + ) + + if args.dataset_name is None: + raise ValueError( + "Please specify '--dataset-name' and the corresponding " + "'--dataset-path' if required." + ) + + if args.dataset_name == "custom": + dataset = CustomDataset(dataset_path=args.dataset_path) + input_requests = dataset.sample( + num_requests=args.num_prompts, + tokenizer=tokenizer, + output_len=args.custom_output_len, + skip_chat_template=args.custom_skip_chat_template, + ) + + elif args.dataset_name == "sonnet": + dataset = SonnetDataset(dataset_path=args.dataset_path) + # For the "sonnet" dataset, formatting depends on the backend. + if args.backend == "openai-chat": + input_requests = dataset.sample( + num_requests=args.num_prompts, + input_len=args.sonnet_input_len, + output_len=args.sonnet_output_len, + prefix_len=args.sonnet_prefix_len, + tokenizer=tokenizer, + return_prompt_formatted=False, + ) + else: + assert tokenizer.chat_template or tokenizer.default_chat_template, ( + "Tokenizer/model must have chat template for sonnet dataset." + ) + input_requests = dataset.sample( + num_requests=args.num_prompts, + input_len=args.sonnet_input_len, + output_len=args.sonnet_output_len, + prefix_len=args.sonnet_prefix_len, + tokenizer=tokenizer, + return_prompt_formatted=True, + ) + + elif args.dataset_name == "hf": + # all following datasets are implemented from the + # HuggingFaceDataset base class + if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS: + dataset_class = VisionArenaDataset + args.hf_split = "train" + args.hf_subset = None + elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS: + dataset_class = InstructCoderDataset + args.hf_split = "train" + elif args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS: + dataset_class = MTBenchDataset + args.hf_split = "train" + elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS: + dataset_class = ConversationDataset + elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS: + dataset_class = AIMODataset + args.hf_split = "train" + elif args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS: # noqa: E501 + dataset_class = NextEditPredictionDataset + args.hf_split = "train" + elif args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS: + dataset_class = ASRDataset + args.hf_split = "train" + else: + supported_datasets = set( + [ + dataset_name + for cls in HuggingFaceDataset.__subclasses__() + for dataset_name in cls.SUPPORTED_DATASET_PATHS + ] + ) + raise ValueError( + f"Unsupported dataset path: {args.dataset_path}. " + "Huggingface dataset only supports dataset_path" + f" from one of following: {supported_datasets}. " + "Please consider contributing if you would " + "like to add support for additional dataset formats." + ) + + if dataset_class.IS_MULTIMODAL and backend not in [ + "openai-chat", + "openai-audio", + ]: + # multi-modal benchmark is only available on OpenAI Chat backend. + raise ValueError( + "Multi-modal content is only supported on 'openai-chat' and " + "'openai-audio' backend." + ) + input_requests = dataset_class( + dataset_path=args.dataset_path, + dataset_subset=args.hf_subset, + dataset_split=args.hf_split, + random_seed=args.seed, + ).sample( + num_requests=args.num_prompts, + tokenizer=tokenizer, + output_len=args.hf_output_len, + ) + + else: + # For datasets that follow a similar structure, use a mapping. + dataset_mapping = { + "sharegpt": lambda: ShareGPTDataset( + random_seed=args.seed, dataset_path=args.dataset_path + ).sample( + tokenizer=tokenizer, + num_requests=args.num_prompts, + output_len=args.sharegpt_output_len, + ), + "burstgpt": lambda: BurstGPTDataset( + random_seed=args.seed, dataset_path=args.dataset_path + ).sample(tokenizer=tokenizer, num_requests=args.num_prompts), + "random": lambda: RandomDataset(dataset_path=args.dataset_path).sample( + tokenizer=tokenizer, + num_requests=args.num_prompts, + prefix_len=args.random_prefix_len, + input_len=args.random_input_len, + output_len=args.random_output_len, + range_ratio=args.random_range_ratio, + ), + } + + try: + input_requests = dataset_mapping[args.dataset_name]() + except KeyError as err: + raise ValueError(f"Unknown dataset: {args.dataset_name}") from err + goodput_config_dict = check_goodput_args(args) + + # Collect the sampling parameters. + sampling_params = { + k: v + for k, v in { + "top_p": args.top_p, + "top_k": args.top_k, + "min_p": args.min_p, + "temperature": args.temperature, + }.items() + if v is not None + } + + # Sampling parameters are only supported by openai-compatible backend. + if sampling_params and args.backend not in OPENAI_COMPATIBLE_BACKENDS: + raise ValueError( + "Sampling parameters are only supported by openai-compatible backends." + ) + + if "temperature" not in sampling_params: + sampling_params["temperature"] = 0.0 # Default to greedy decoding. + + if args.backend == "llama.cpp": + # Disable prompt caching in llama.cpp backend + sampling_params["cache_prompt"] = False + + # Avoid GC processing "static" data - reduce pause times. + gc.collect() + gc.freeze() + + benchmark_result = asyncio.run( + benchmark( + backend=backend, + api_url=api_url, + base_url=base_url, + model_id=model_id, + model_name=model_name, + tokenizer=tokenizer, + input_requests=input_requests, + logprobs=args.logprobs, + request_rate=args.request_rate, + burstiness=args.burstiness, + disable_tqdm=args.disable_tqdm, + profile=args.profile, + selected_percentile_metrics=args.percentile_metrics.split(","), + selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")], + ignore_eos=args.ignore_eos, + goodput_config_dict=goodput_config_dict, + max_concurrency=args.max_concurrency, + lora_modules=args.lora_modules, + extra_body=sampling_params, + ramp_up_strategy=args.ramp_up_strategy, + ramp_up_start_rps=args.ramp_up_start_rps, + ramp_up_end_rps=args.ramp_up_end_rps, + ) + ) + + # Save config and results to json + if args.save_result or args.append_result: + result_json: dict[str, Any] = {} + + # Setup + current_dt = datetime.now().strftime("%Y%m%d-%H%M%S") + result_json["date"] = current_dt + result_json["backend"] = backend + result_json["model_id"] = model_id + result_json["tokenizer_id"] = tokenizer_id + result_json["num_prompts"] = args.num_prompts + + # Metadata + if args.metadata: + for item in args.metadata: + if "=" in item: + kvstring = item.split("=") + result_json[kvstring[0].strip()] = kvstring[1].strip() + else: + raise ValueError( + "Invalid metadata format. Please use KEY=VALUE format." + ) + # Traffic + result_json["request_rate"] = ( + args.request_rate if args.request_rate < float("inf") else "inf" + ) + result_json["burstiness"] = args.burstiness + result_json["max_concurrency"] = args.max_concurrency + + if args.ramp_up_strategy is not None: + result_json["ramp_up_strategy"] = args.ramp_up_strategy + result_json["ramp_up_start_rps"] = args.ramp_up_start_rps + result_json["ramp_up_end_rps"] = args.ramp_up_end_rps + + # Merge with benchmark result + result_json = {**result_json, **benchmark_result} + + if not args.save_detailed: + # Remove fields with too many data points + for field in [ + "input_lens", + "output_lens", + "ttfts", + "itls", + "generated_texts", + "errors", + ]: + if field in result_json: + del result_json[field] + if field in benchmark_result: + del benchmark_result[field] + + # Save to file + base_model_id = model_id.split("/")[-1] + max_concurrency_str = ( + f"-concurrency{args.max_concurrency}" + if args.max_concurrency is not None + else "" + ) + if args.ramp_up_strategy is not None: + file_name = f"{backend}-ramp-up-{args.ramp_up_strategy}-{args.ramp_up_start_rps}qps-{args.ramp_up_end_rps}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json" # noqa + else: + file_name = f"{backend}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json" # noqa + if args.result_filename: + file_name = args.result_filename + if args.result_dir: + os.makedirs(args.result_dir, exist_ok=True) + file_name = os.path.join(args.result_dir, file_name) + with open( + file_name, mode="a+" if args.append_result else "w", encoding="utf-8" + ) as outfile: + # Append a newline. + if args.append_result and outfile.tell() != 0: + outfile.write("\n") + json.dump(result_json, outfile) + save_to_pytorch_benchmark_format(args, result_json, file_name) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the online serving throughput." + ) + parser.add_argument( + "--backend", + type=str, + default="vllm", + choices=list(ASYNC_REQUEST_FUNCS.keys()), + ) + parser.add_argument( + "--base-url", + type=str, + default=None, + help="Server or API base url if not using http host and port.", + ) + # Use 127.0.0.1 here instead of localhost to force the use of ipv4 + parser.add_argument("--host", type=str, default="127.0.0.1") + parser.add_argument("--port", type=int, default=8000) + parser.add_argument( + "--endpoint", + type=str, + default="/v1/completions", + help="API endpoint.", + ) + parser.add_argument( + "--dataset-name", + type=str, + default="sharegpt", + choices=["sharegpt", "burstgpt", "sonnet", "random", "hf", "custom"], + help="Name of the dataset to benchmark on.", + ) + parser.add_argument( + "--dataset-path", + type=str, + default=None, + help="Path to the sharegpt/sonnet dataset. " + "Or the huggingface dataset ID if using HF dataset.", + ) + parser.add_argument( + "--max-concurrency", + type=int, + default=None, + help="Maximum number of concurrent requests. This can be used " + "to help simulate an environment where a higher level component " + "is enforcing a maximum number of concurrent requests. While the " + "--request-rate argument controls the rate at which requests are " + "initiated, this argument will control how many are actually allowed " + "to execute at a time. This means that when used in combination, the " + "actual request rate may be lower than specified with --request-rate, " + "if the server is not processing requests fast enough to keep up.", + ) + + parser.add_argument( + "--model", + type=str, + required=True, + help="Name of the model.", + ) + parser.add_argument( + "--tokenizer", + type=str, + help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 + ) + parser.add_argument("--use-beam-search", action="store_true") + parser.add_argument( + "--num-prompts", + type=int, + default=1000, + help="Number of prompts to process.", + ) + parser.add_argument( + "--logprobs", + type=int, + default=None, + help=( + "Number of logprobs-per-token to compute & return as part of " + "the request. If unspecified, then either (1) if beam search " + "is disabled, no logprobs are computed & a single dummy " + "logprob is returned for each token; or (2) if beam search " + "is enabled 1 logprob per token is computed" + ), + ) + parser.add_argument( + "--request-rate", + type=float, + default=float("inf"), + help="Number of requests per second. If this is inf, " + "then all the requests are sent at time 0. " + "Otherwise, we use Poisson process or gamma distribution " + "to synthesize the request arrival times.", + ) + parser.add_argument( + "--burstiness", + type=float, + default=1.0, + help="Burstiness factor of the request generation. " + "Only take effect when request_rate is not inf. " + "Default value is 1, which follows Poisson process. " + "Otherwise, the request intervals follow a gamma distribution. " + "A lower burstiness value (0 < burstiness < 1) results in more " + "bursty requests. A higher burstiness value (burstiness > 1) " + "results in a more uniform arrival of requests.", + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument( + "--trust-remote-code", + action="store_true", + help="Trust remote code from huggingface", + ) + parser.add_argument( + "--disable-tqdm", + action="store_true", + help="Specify to disable tqdm progress bar.", + ) + parser.add_argument( + "--profile", + action="store_true", + help="Use Torch Profiler. The endpoint must be launched with " + "VLLM_TORCH_PROFILER_DIR to enable profiler.", + ) + parser.add_argument( + "--save-result", + action="store_true", + help="Specify to save benchmark results to a json file", + ) + parser.add_argument( + "--save-detailed", + action="store_true", + help="When saving the results, whether to include per request " + "information such as response, error, ttfs, tpots, etc.", + ) + parser.add_argument( + "--append-result", + action="store_true", + help="Append the benchmark result to the existing json file.", + ) + parser.add_argument( + "--metadata", + metavar="KEY=VALUE", + nargs="*", + help="Key-value pairs (e.g, --metadata version=0.3.3 tp=1) " + "for metadata of this run to be saved in the result JSON file " + "for record keeping purposes.", + ) + parser.add_argument( + "--result-dir", + type=str, + default=None, + help="Specify directory to save benchmark json results." + "If not specified, results are saved in the current directory.", + ) + parser.add_argument( + "--result-filename", + type=str, + default=None, + help="Specify the filename to save benchmark json results." + "If not specified, results will be saved in " + "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json" + " format.", + ) + parser.add_argument( + "--ignore-eos", + action="store_true", + help="Set ignore_eos flag when sending the benchmark request." + "Warning: ignore_eos is not supported in deepspeed_mii and tgi.", + ) + parser.add_argument( + "--percentile-metrics", + type=str, + default="ttft,tpot,itl", + help="Comma-separated list of selected metrics to report percentils. " + "This argument specifies the metrics to report percentiles. " + 'Allowed metric names are "ttft", "tpot", "itl", "e2el". ' + 'Default value is "ttft,tpot,itl".', + ) + parser.add_argument( + "--metric-percentiles", + type=str, + default="99", + help="Comma-separated list of percentiles for selected metrics. " + 'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". ' + 'Default value is "99". ' + 'Use "--percentile-metrics" to select metrics.', + ) + parser.add_argument( + "--goodput", + nargs="+", + required=False, + help='Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is in " + 'milliseconds. Multiple "KEY:VALUE" pairs can be provided, ' + "separated by spaces. Allowed request level metric names are " + '"ttft", "tpot", "e2el". For more context on the definition of ' + "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 " + "and the blog: https://hao-ai-lab.github.io/blogs/distserve", + ) + + # group for dataset specific arguments + custom_group = parser.add_argument_group("custom dataset options") + custom_group.add_argument( + "--custom-output-len", + type=int, + default=256, + help="Number of output tokens per request, used only for custom dataset.", + ) + custom_group.add_argument( + "--custom-skip-chat-template", + action="store_true", + help="Skip applying chat template to prompt, used only for custom dataset.", + ) + + sonnet_group = parser.add_argument_group("sonnet dataset options") + sonnet_group.add_argument( + "--sonnet-input-len", + type=int, + default=550, + help="Number of input tokens per request, used only for sonnet dataset.", + ) + sonnet_group.add_argument( + "--sonnet-output-len", + type=int, + default=150, + help="Number of output tokens per request, used only for sonnet dataset.", + ) + sonnet_group.add_argument( + "--sonnet-prefix-len", + type=int, + default=200, + help="Number of prefix tokens per request, used only for sonnet dataset.", + ) + + sharegpt_group = parser.add_argument_group("sharegpt dataset options") + sharegpt_group.add_argument( + "--sharegpt-output-len", + type=int, + default=None, + help="Output length for each request. Overrides the output length " + "from the ShareGPT dataset.", + ) + + random_group = parser.add_argument_group("random dataset options") + random_group.add_argument( + "--random-input-len", + type=int, + default=1024, + help="Number of input tokens per request, used only for random sampling.", + ) + random_group.add_argument( + "--random-output-len", + type=int, + default=128, + help="Number of output tokens per request, used only for random sampling.", + ) + random_group.add_argument( + "--random-range-ratio", + type=float, + default=0.0, + help="Range ratio for sampling input/output length, " + "used only for random sampling. Must be in the range [0, 1) to define " + "a symmetric sampling range" + "[length * (1 - range_ratio), length * (1 + range_ratio)].", + ) + random_group.add_argument( + "--random-prefix-len", + type=int, + default=0, + help=( + "Number of fixed prefix tokens before the random context " + "in a request. " + "The total input length is the sum of `random-prefix-len` and " + "a random " + "context length sampled from [input_len * (1 - range_ratio), " + "input_len * (1 + range_ratio)]." + ), + ) + + hf_group = parser.add_argument_group("hf dataset options") + hf_group.add_argument( + "--hf-subset", type=str, default=None, help="Subset of the HF dataset." + ) + hf_group.add_argument( + "--hf-split", type=str, default=None, help="Split of the HF dataset." + ) + hf_group.add_argument( + "--hf-output-len", + type=int, + default=None, + help="Output length for each request. Overrides the output lengths " + "from the sampled HF dataset.", + ) + + sampling_group = parser.add_argument_group("sampling parameters") + sampling_group.add_argument( + "--top-p", + type=float, + default=None, + help="Top-p sampling parameter. Only has effect on openai-compatible backends.", + ) + sampling_group.add_argument( + "--top-k", + type=int, + default=None, + help="Top-k sampling parameter. Only has effect on openai-compatible backends.", + ) + sampling_group.add_argument( + "--min-p", + type=float, + default=None, + help="Min-p sampling parameter. Only has effect on openai-compatible backends.", + ) + sampling_group.add_argument( + "--temperature", + type=float, + default=None, + help="Temperature sampling parameter. Only has effect on " + "openai-compatible backends. If not specified, default to greedy " + "decoding (i.e. temperature==0.0).", + ) + + parser.add_argument( + "--tokenizer-mode", + type=str, + default="auto", + choices=["auto", "slow", "mistral", "custom"], + help='The tokenizer mode.\n\n* "auto" will use the ' + 'fast tokenizer if available.\n* "slow" will ' + "always use the slow tokenizer. \n* " + '"mistral" will always use the `mistral_common` tokenizer. \n*' + '"custom" will use --tokenizer to select the preregistered tokenizer.', + ) + + parser.add_argument( + "--served-model-name", + type=str, + default=None, + help="The model name used in the API. " + "If not specified, the model name will be the " + "same as the ``--model`` argument. ", + ) + + parser.add_argument( + "--lora-modules", + nargs="+", + default=None, + help="A subset of LoRA module names passed in when " + "launching the server. For each request, the " + "script chooses a LoRA module at random.", + ) + + parser.add_argument( + "--ramp-up-strategy", + type=str, + default=None, + choices=["linear", "exponential"], + help="The ramp-up strategy. This would be used to " + "ramp up the request rate from initial RPS to final " + "RPS rate (specified by --ramp-up-start-rps and --ramp-up-end-rps). " + "over the duration of the benchmark.", + ) + parser.add_argument( + "--ramp-up-start-rps", + type=int, + default=None, + help="The starting request rate for ramp-up (RPS). " + "Needs to be specified when --ramp-up-strategy is used.", + ) + parser.add_argument( + "--ramp-up-end-rps", + type=int, + default=None, + help="The ending request rate for ramp-up (RPS). " + "Needs to be specified when --ramp-up-strategy is used.", + ) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test/benchmarks/benchmark_serving_structured_output.py b/online_apiserver_test/benchmarks/benchmark_serving_structured_output.py new file mode 100644 index 0000000000000000000000000000000000000000..e23a5a9e2233df685998fbae0ae8cc8910b572cf --- /dev/null +++ b/online_apiserver_test/benchmarks/benchmark_serving_structured_output.py @@ -0,0 +1,1043 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +r"""Benchmark online serving throughput with structured outputs. + +On the server side, run one of the following commands: + (vLLM OpenAI API server) + vllm serve --disable-log-requests + +On the client side, run: + python benchmarks/benchmark_serving_structured_output.py \ + --backend \ + --model \ + --dataset json \ + --structured-output-ratio 1.0 \ + --request-rate 10 \ + --num-prompts 1000 + + when using tgi backend, add + --endpoint /generate_stream + to the end of the command above. +""" + +import argparse +import asyncio +import copy +import dataclasses +import json +import os +import random +import time +import uuid +import warnings +from collections.abc import AsyncGenerator +from dataclasses import dataclass +from typing import Optional + +import datasets +import numpy as np +import pandas as pd +from tqdm.asyncio import tqdm +from transformers import PreTrainedTokenizerBase + +from backend_request_func import ( + ASYNC_REQUEST_FUNCS, + RequestFuncInput, + RequestFuncOutput, +) + +try: + from vllm.transformers_utils.tokenizer import get_tokenizer +except ImportError: + from backend_request_func import get_tokenizer + +try: + from vllm.utils import FlexibleArgumentParser +except ImportError: + from argparse import ArgumentParser as FlexibleArgumentParser + +from vllm.v1.structured_output.backend_xgrammar import ( + has_xgrammar_unsupported_json_features, +) + +MILLISECONDS_TO_SECONDS_CONVERSION = 1000 + + +@dataclass +class BenchmarkMetrics: + completed: int + total_input: int + total_output: int + request_throughput: float + request_goodput: float + output_throughput: float + total_token_throughput: float + mean_ttft_ms: float + median_ttft_ms: float + std_ttft_ms: float + percentiles_ttft_ms: list[tuple[float, float]] + mean_tpot_ms: float + median_tpot_ms: float + std_tpot_ms: float + percentiles_tpot_ms: list[tuple[float, float]] + mean_itl_ms: float + median_itl_ms: float + std_itl_ms: float + percentiles_itl_ms: list[tuple[float, float]] + # E2EL stands for end-to-end latency per request. + # It is the time taken on the client side from sending + # a request to receiving a complete response. + mean_e2el_ms: float + median_e2el_ms: float + std_e2el_ms: float + percentiles_e2el_ms: list[tuple[float, float]] + + +@dataclasses.dataclass +class SampleRequest: + """A class representing a single inference request for benchmarking. + + Attributes: + prompt: The input text prompt for the model. + multi_modal_data: Optional dictionary containing multi-modal data (e.g. + images). + prompt_len: The length of the prompt in tokens. + expected_output_len: The expected length of the output in tokens. + """ + + prompt: str + prompt_len: int + expected_output_len: int + schema: dict + structure_type: str + completion: str = None + + +def sample_requests( + tokenizer: PreTrainedTokenizerBase, args: argparse.Namespace +) -> list[SampleRequest]: + if args.dataset == "json" or args.dataset == "json-unique": + if args.json_schema_path is None: + dir_path = os.path.dirname(os.path.realpath(__file__)) + args.json_schema_path = os.path.join( + dir_path, "structured_schemas", "structured_schema_1.json" + ) + json_schemas = [] + with open(args.json_schema_path) as f: + schema = json.load(f) + + if args.dataset == "json-unique": + json_schemas = [copy.deepcopy(schema) for _ in range(args.num_prompts)] + for i in range(len(json_schemas)): + if "properties" not in json_schemas[i]: + json_schemas[i]["properties"] = {} + json_schemas[i]["properties"][f"__optional_field_{uuid.uuid4()}"] = { + "type": "string", + "description": "An unique optional field to avoid cached schemas", + } + else: + json_schemas = [schema] * args.num_prompts + + def gen_prompt(index: int): + return f"Generate an example of a brief user profile given the following schema: {json.dumps(get_schema(index))}" # noqa: E501 + + def get_schema(index: int): + return json_schemas[index % len(json_schemas)] + + requests = [ + SampleRequest( + prompt=gen_prompt(i), + prompt_len=len(tokenizer(gen_prompt(i)).input_ids), + expected_output_len=args.output_len, + schema=get_schema(i), + structure_type=args.structure_type, + ) + for i in range(args.num_prompts) + ] + + elif args.dataset == "grammar": + schema = """ + root ::= select_statement + + select_statement ::= "SELECT " column " from " table " where " condition + + column ::= "col_1 " | "col_2 " + + table ::= "table_1 " | "table_2 " + + condition ::= column "= " number + + number ::= "1 " | "2 " + """ + prompt = "Generate an SQL query to show the 'username' \ + and 'email' from the 'users' table." + + input_len = len(tokenizer(prompt).input_ids) + print(f"Input length of the prompt: {input_len} tokens") + requests = [ + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=schema, + structure_type=args.structure_type, + ) + for _ in range(args.num_prompts) + ] + + elif args.dataset == "regex": + regex = r"\w+@\w+\.com\n" + args.regex = regex + prompt = "Generate an email address for Alan Turing, \ + who works in Enigma. End in .com and new line. \ + Example result: alan.turing@enigma.com\n" + + input_len = len(tokenizer(prompt).input_ids) + print(f"Input length of the prompt: {input_len} tokens") + requests = [ + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=regex, + structure_type=args.structure_type, + ) + for _ in range(args.num_prompts) + ] + + elif args.dataset == "choice": + choice = ["Positive", "Negative"] + args.choice = choice + prompt = "Classify this sentiment: vLLM is wonderful!" + input_len = len(tokenizer(prompt).input_ids) + print(f"Input length of the prompt: {input_len} tokens") + requests = [ + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=choice, + structure_type=args.structure_type, + ) + for _ in range(args.num_prompts) + ] + + elif args.dataset == "xgrammar_bench": + requests: list[SampleRequest] = [] + dataset = datasets.load_dataset("NousResearch/json-mode-eval", split="train") + full_dataset_len = len(dataset) + + def _filter_func(item): + import json + + schema = json.loads(item["schema"]) + return not has_xgrammar_unsupported_json_features(schema) + + dataset = dataset.filter(_filter_func) + num_filtered_out = full_dataset_len - len(dataset) + print( + f"dataset has {len(dataset)} entries after filtering " + f"out {num_filtered_out} entries with unsupported features" + ) + len_dataset = len(dataset) + for data_point_idx in range(args.num_prompts): + idx = data_point_idx + while idx >= len_dataset: + idx -= len_dataset + schema = dataset["schema"][idx] + prompt = tokenizer.apply_chat_template( + dataset["prompt"][idx], tokenize=False, add_generation_prompt=True + ) + input_len = len(tokenizer(prompt).input_ids) + completion = dataset["completion"][idx] + + requests.append( + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=schema, + structure_type=args.structure_type, + completion=completion, + ) + ) + + return requests + + +async def get_request( + input_requests: list[SampleRequest], + request_rate: float, + burstiness: float = 1.0, +) -> AsyncGenerator[tuple[int, SampleRequest], None]: + """ + Asynchronously generates requests at a specified rate + with OPTIONAL burstiness. + + Args: + input_requests: + A list of input requests, each represented as a tuple. + request_rate: + The rate at which requests are generated (requests/s). + burstiness (optional): + The burstiness factor of the request generation. + Only takes effect when request_rate is not inf. + Default value is 1, which follows a Poisson process. + Otherwise, the request intervals follow a gamma distribution. + A lower burstiness value (0 < burstiness < 1) results + in more bursty requests, while a higher burstiness value + (burstiness > 1) results in a more uniform arrival of requests. + """ + input_requests = iter(input_requests) + + # Calculate scale parameter theta to maintain the desired request_rate. + assert burstiness > 0, ( + f"A positive burstiness factor is expected, but given {burstiness}." + ) + theta = 1.0 / (request_rate * burstiness) + + for i, request in enumerate(input_requests): + yield i, request + + if request_rate == float("inf"): + # If the request rate is infinity, then we don't need to wait. + continue + + # Sample the request interval from the gamma distribution. + # If burstiness is 1, it follows exponential distribution. + interval = np.random.gamma(shape=burstiness, scale=theta) + # The next request will be sent after the interval. + await asyncio.sleep(interval) + + +def calculate_metrics( + input_requests: list[tuple[str, int, int]], + outputs: list[RequestFuncOutput], + dur_s: float, + tokenizer: PreTrainedTokenizerBase, + selected_percentile_metrics: list[str], + selected_percentiles: list[float], + goodput_config_dict: Optional[dict[str, float]] = None, +) -> tuple[BenchmarkMetrics, list[int]]: + actual_output_lens: list[int] = [] + total_input = 0 + completed = 0 + good_completed = 0 + itls: list[float] = [] + tpots: list[float] = [] + all_tpots: list[float] = [] + ttfts: list[float] = [] + e2els: list[float] = [] + for i in range(len(outputs)): + if outputs[i].success: + # We use the tokenizer to count the number of output tokens for all + # serving backends instead of looking at len(outputs[i].itl) since + # multiple output tokens may be bundled together + # Note : this may inflate the output token count slightly + output_len = len( + tokenizer(outputs[i].generated_text, add_special_tokens=False).input_ids + ) + actual_output_lens.append(output_len) + total_input += input_requests[i].prompt_len + tpot = 0 + if output_len > 1: + latency_minus_ttft = outputs[i].latency - outputs[i].ttft + tpot = latency_minus_ttft / (output_len - 1) + tpots.append(tpot) + outputs[i].tpot = tpot + # Note: if output_len <= 1, we regard tpot as 0 for goodput + all_tpots.append(tpot) + itls += outputs[i].itl + ttfts.append(outputs[i].ttft) + e2els.append(outputs[i].latency) + completed += 1 + else: + actual_output_lens.append(0) + + if goodput_config_dict: + valid_metrics = [] + slo_values = [] + + if "ttft" in goodput_config_dict: + valid_metrics.append(ttfts) + slo_values.append( + goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "tpot" in goodput_config_dict: + valid_metrics.append(all_tpots) + slo_values.append( + goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "e2el" in goodput_config_dict: + valid_metrics.append(e2els) + slo_values.append( + goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + + for req_metric in zip(*valid_metrics): + is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)]) + if is_good_req: + good_completed += 1 + + if completed == 0: + warnings.warn( + "All requests failed. This is likely due to a misconfiguration " + "on the benchmark arguments.", + stacklevel=2, + ) + metrics = BenchmarkMetrics( + completed=completed, + total_input=total_input, + total_output=sum(actual_output_lens), + request_throughput=completed / dur_s, + request_goodput=good_completed / dur_s, + output_throughput=sum(actual_output_lens) / dur_s, + total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s, + mean_ttft_ms=np.mean(ttfts or 0) + * 1000, # ttfts is empty if streaming is not supported by backend + std_ttft_ms=np.std(ttfts or 0) * 1000, + median_ttft_ms=np.median(ttfts or 0) * 1000, + percentiles_ttft_ms=[ + (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles + ], + mean_tpot_ms=np.mean(tpots or 0) * 1000, + std_tpot_ms=np.std(tpots or 0) * 1000, + median_tpot_ms=np.median(tpots or 0) * 1000, + percentiles_tpot_ms=[ + (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles + ], + mean_itl_ms=np.mean(itls or 0) * 1000, + std_itl_ms=np.std(itls or 0) * 1000, + median_itl_ms=np.median(itls or 0) * 1000, + percentiles_itl_ms=[ + (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles + ], + mean_e2el_ms=np.mean(e2els or 0) * 1000, + std_e2el_ms=np.std(e2els or 0) * 1000, + median_e2el_ms=np.median(e2els or 0) * 1000, + percentiles_e2el_ms=[ + (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles + ], + ) + + return metrics, actual_output_lens + + +async def benchmark( + backend: str, + api_url: str, + base_url: str, + model_id: str, + tokenizer: PreTrainedTokenizerBase, + input_requests: list[SampleRequest], + request_rate: float, + burstiness: float, + disable_tqdm: bool, + profile: bool, + selected_percentile_metrics: list[str], + selected_percentiles: list[str], + ignore_eos: bool, + max_concurrency: Optional[int], + structured_output_ratio: float, + goodput_config_dict: Optional[dict[str, float]] = None, +): + if backend in ASYNC_REQUEST_FUNCS: + request_func = ASYNC_REQUEST_FUNCS[backend] + else: + raise ValueError(f"Unknown backend: {backend}") + + def prepare_extra_body(request) -> dict: + extra_body = {} + # Add the schema to the extra_body + extra_body[request.structure_type] = request.schema + return extra_body + + print("Starting initial single prompt test run...") + structured_output_req_idx = random.sample( + range(len(input_requests)), int(len(input_requests) * structured_output_ratio) + ) + + test_request = input_requests[0] + test_req_extra_body = ( + prepare_extra_body(test_request) if 0 in structured_output_req_idx else None + ) + test_input = RequestFuncInput( + model=model_id, + prompt=test_request.prompt, + api_url=api_url, + prompt_len=test_request.prompt_len, + output_len=test_request.expected_output_len, + ignore_eos=ignore_eos, + extra_body=test_req_extra_body, + ) + test_output = await request_func(request_func_input=test_input) + if not test_output.success: + raise ValueError( + "Initial test run failed - Please make sure benchmark arguments " + f"are correctly specified. Error: {test_output.error}" + ) + else: + print("Initial test run completed. Starting main benchmark run...") + + if profile: + print("Starting profiler...") + profile_input = RequestFuncInput( + model=model_id, + prompt=test_request.prompt, + api_url=base_url + "/start_profile", + prompt_len=test_request.prompt_len, + output_len=test_request.expected_output_len, + ignore_eos=ignore_eos, + extra_body=test_req_extra_body, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler started") + + distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution" + + print(f"Traffic request rate: {request_rate}") + print(f"Burstiness factor: {burstiness} ({distribution})") + print(f"Maximum request concurrency: {max_concurrency}") + + pbar = None if disable_tqdm else tqdm(total=len(input_requests)) + + # This can be used once the minimum Python version is 3.10 or higher, + # and it will simplify the code in limited_request_func. + # semaphore = (asyncio.Semaphore(max_concurrency) + # if max_concurrency else contextlib.nullcontext()) + semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None + + async def limited_request_func(request_func_input, pbar): + if semaphore is None: + return await request_func(request_func_input=request_func_input, pbar=pbar) + async with semaphore: + return await request_func(request_func_input=request_func_input, pbar=pbar) + + benchmark_start_time = time.perf_counter() + tasks: list[asyncio.Task] = [] + expected: list[str] = [] + async for i, request in get_request(input_requests, request_rate, burstiness): + extra_body = ( + prepare_extra_body(request) if i in structured_output_req_idx else None + ) + request_func_input = RequestFuncInput( + model=model_id, + prompt=request.prompt, + api_url=api_url, + prompt_len=request.prompt_len, + output_len=request.expected_output_len, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + expected.append(request.completion) + tasks.append( + asyncio.create_task( + limited_request_func(request_func_input=request_func_input, pbar=pbar) + ) + ) + outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks) + + if profile: + print("Stopping profiler...") + profile_input = RequestFuncInput( + model=model_id, + prompt=test_request.prompt, + api_url=base_url + "/stop_profile", + prompt_len=test_request.prompt_len, + output_len=test_request.expected_output_len, + extra_body={test_request.structure_type: test_request.schema}, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler stopped") + + if pbar is not None: + pbar.close() + + benchmark_duration = time.perf_counter() - benchmark_start_time + + metrics, actual_output_lens = calculate_metrics( + input_requests=input_requests, + outputs=outputs, + dur_s=benchmark_duration, + tokenizer=tokenizer, + selected_percentile_metrics=selected_percentile_metrics, + selected_percentiles=selected_percentiles, + goodput_config_dict=goodput_config_dict, + ) + + print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="=")) + print("{:<40} {:<10}".format("Successful requests:", metrics.completed)) + print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration)) + print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input)) + print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output)) + print( + "{:<40} {:<10.2f}".format( + "Request throughput (req/s):", metrics.request_throughput + ) + ) + if goodput_config_dict: + print( + "{:<40} {:<10.2f}".format( + "Request goodput (req/s):", metrics.request_goodput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Output token throughput (tok/s):", metrics.output_throughput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Total Token throughput (tok/s):", metrics.total_token_throughput + ) + ) + + result = { + "duration": benchmark_duration, + "completed": metrics.completed, + "total_input_tokens": metrics.total_input, + "total_output_tokens": metrics.total_output, + "request_throughput": metrics.request_throughput, + "output_throughput": metrics.output_throughput, + "total_token_throughput": metrics.total_token_throughput, + "ttft_description": pd.Series([output.ttft for output in outputs]) + .describe() + .to_dict(), + "tpot_description": pd.Series([output.tpot for output in outputs]) + .describe() + .to_dict(), + "input_lens": [output.prompt_len for output in outputs], + "output_lens": actual_output_lens, + "ttfts": [output.ttft for output in outputs], + "itls": [output.itl for output in outputs], + "errors": [output.error for output in outputs], + } + + ret = [ + {"generated": output.generated_text, "expected": gt} + for output, gt in zip(outputs, expected) + ] + + def process_one_metric( + # E.g., "ttft" + metric_attribute_name: str, + # E.g., "TTFT" + metric_name: str, + # E.g., "Time to First Token" + metric_header: str, + ): + # This function prints and adds statistics of the specified + # metric. + if metric_attribute_name not in selected_percentile_metrics: + return + print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-")) + print( + "{:<40} {:<10.2f}".format( + f"Mean {metric_name} (ms):", + getattr(metrics, f"mean_{metric_attribute_name}_ms"), + ) + ) + print( + "{:<40} {:<10.2f}".format( + f"Median {metric_name} (ms):", + getattr(metrics, f"median_{metric_attribute_name}_ms"), + ) + ) + result[f"mean_{metric_attribute_name}_ms"] = getattr( + metrics, f"mean_{metric_attribute_name}_ms" + ) + result[f"median_{metric_attribute_name}_ms"] = getattr( + metrics, f"median_{metric_attribute_name}_ms" + ) + result[f"std_{metric_attribute_name}_ms"] = getattr( + metrics, f"std_{metric_attribute_name}_ms" + ) + for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"): + p_word = str(int(p)) if int(p) == p else str(p) + print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value)) + result[f"p{p_word}_{metric_attribute_name}_ms"] = value + + process_one_metric("ttft", "TTFT", "Time to First Token") + process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)") + process_one_metric("itl", "ITL", "Inter-token Latency") + process_one_metric("e2el", "E2EL", "End-to-end Latency") + + print("=" * 50) + + return result, ret + + +def evaluate(ret, args): + def _eval_correctness_json(expected, actual): + # extract json string from string using regex + import regex as re + + actual = actual.replace("\n", "").replace(" ", "").strip() + try: + actual = re.search(r"\{.*\}", actual).group() + actual = json.loads(actual) + except Exception: + return False + + return True + + def _eval_correctness_choice(expected, actual): + return actual in args.choice + + def _eval_correctness_regex(expected, actual): + import regex as re + + return re.match(args.regex, actual) is not None + + def _eval_correctness(expected, actual): + if args.structure_type == "guided_json": + return _eval_correctness_json(expected, actual) + elif args.structure_type == "guided_regex": + return _eval_correctness_regex(expected, actual) + elif args.structure_type == "guided_choice": + return _eval_correctness_choice(expected, actual) + else: + return None + + scores = [] + for res in ret: + score = _eval_correctness(res["expected"], res["generated"]) + res["correctness"] = score + scores.append(score) + + not_none_scores = [score for score in scores if score is not None] + + return ( + (sum(not_none_scores) / len(not_none_scores) * 100) + if len(not_none_scores) > 0 + else None + ) + + +def parse_goodput(slo_pairs): + goodput_config_dict = {} + try: + for slo_pair in slo_pairs: + slo_name, slo_val = slo_pair.split(":") + goodput_config_dict[slo_name] = float(slo_val) + except ValueError as err: + raise argparse.ArgumentTypeError( + "Invalid format found for service level objectives. " + 'Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is a " + "number in milliseconds." + ) from err + return goodput_config_dict + + +def check_goodput_args(args): + goodput_config_dict = {} + VALID_NAMES = ["ttft", "tpot", "e2el"] + if args.goodput: + goodput_config_dict = parse_goodput(args.goodput) + for slo_name, slo_val in goodput_config_dict.items(): + if slo_name not in VALID_NAMES: + raise ValueError( + f"Invalid metric name found, {slo_name}: {slo_val}. " + "The service level objective name should be one of " + f"{str(VALID_NAMES)}. " + ) + if slo_val < 0: + raise ValueError( + f"Invalid value found, {slo_name}: {slo_val}. " + "The service level objective value should be " + "non-negative." + ) + return goodput_config_dict + + +def main(args: argparse.Namespace): + print(args) + random.seed(args.seed) + np.random.seed(args.seed) + + backend = args.backend + model_id = args.model + tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model + + if args.base_url is not None: + api_url = f"{args.base_url}{args.endpoint}" + base_url = f"{args.base_url}" + else: + api_url = f"http://{args.host}:{args.port}{args.endpoint}" + base_url = f"http://{args.host}:{args.port}" + + tokenizer = get_tokenizer( + tokenizer_id, + trust_remote_code=args.trust_remote_code, + tokenizer_mode=args.tokenizer_mode, + ) + + if args.dataset == "grammar": + args.structure_type = "guided_grammar" + elif args.dataset == "regex": + args.structure_type = "guided_regex" + elif args.dataset == "choice": + args.structure_type = "guided_choice" + else: + args.structure_type = "guided_json" + + if args.no_structured_output: + args.structured_output_ratio = 0 + if args.save_results: + result_file_name = f"{args.structured_output_ratio}guided" + result_file_name += f"_{backend}" + result_file_name += f"_{args.request_rate}qps" + result_file_name += f"_{args.model.split('/')[-1]}" + result_file_name += f"_{args.dataset}" + result_file_name += f"_{args.num_prompts}" + result_file_name += f"_out{args.output_len}" + result_file_name += ".txt" + else: + result_file_name = None + + input_requests = sample_requests(tokenizer, args) + + goodput_config_dict = check_goodput_args(args) + + benchmark_result, ret = asyncio.run( + benchmark( + backend=backend, + api_url=api_url, + base_url=base_url, + model_id=model_id, + tokenizer=tokenizer, + input_requests=input_requests, + request_rate=args.request_rate, + burstiness=args.burstiness, + disable_tqdm=args.disable_tqdm, + profile=args.profile, + selected_percentile_metrics=args.percentile_metrics.split(","), + selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")], + ignore_eos=args.ignore_eos, + max_concurrency=args.max_concurrency, + structured_output_ratio=args.structured_output_ratio, + goodput_config_dict=goodput_config_dict, + ) + ) + + # Save config and results to json + score = evaluate(ret, args) + print("correct_rate(%)", score, "\n") + if args.save_results: + results = { + "backend": backend, + "model_id": model_id, + "tokenizer_id": tokenizer_id, + "num_prompts": args.num_prompts, + "request_rate": args.request_rate + if args.request_rate < float("inf") + else "inf", + "burstiness": args.burstiness, + "max_concurrency": args.max_concurrency, + "correct_rate(%)": score, + } + results = {"outputs": ret, **results, **benchmark_result} + + # Save to file + if args.result_filename: + result_file_name = args.result_filename + if args.result_dir: + result_file_name = os.path.join(args.result_dir, result_file_name) + with open(result_file_name, "w", encoding="utf-8") as outfile: + json.dump(results, outfile, indent=4) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the online serving throughput." + ) + parser.add_argument( + "--backend", + type=str, + default="vllm", + choices=list(ASYNC_REQUEST_FUNCS.keys()), + ) + parser.add_argument( + "--base-url", + type=str, + default=None, + help="Server or API base url if not using http host and port.", + ) + # Use 127.0.0.1 here instead of localhost to force the use of ipv4 + parser.add_argument("--host", type=str, default="127.0.0.1") + parser.add_argument("--port", type=int, default=8000) + parser.add_argument( + "--endpoint", + type=str, + default="/v1/completions", + help="API endpoint.", + ) + parser.add_argument( + "--dataset", + default="json", + choices=["json", "json-unique", "grammar", "regex", "choice", "xgrammar_bench"], + ) + parser.add_argument( + "--json-schema-path", type=str, default=None, help="Path to json schema." + ) + parser.add_argument( + "--max-concurrency", + type=int, + default=None, + help="Maximum number of concurrent requests. This can be used " + "to help simulate an environment where a higher level component " + "is enforcing a maximum number of concurrent requests. While the " + "--request-rate argument controls the rate at which requests are " + "initiated, this argument will control how many are actually allowed " + "to execute at a time. This means that when used in combination, the " + "actual request rate may be lower than specified with --request-rate, " + "if the server is not processing requests fast enough to keep up.", + ) + parser.add_argument( + "--model", + type=str, + required=True, + help="Name of the model.", + ) + parser.add_argument( + "--tokenizer", + type=str, + help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 + ) + parser.add_argument( + "--tokenizer-mode", + type=str, + default="auto", + help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 + ) + parser.add_argument( + "--num-prompts", + type=int, + default=1000, + help="Number of prompts to process.", + ) + parser.add_argument( + "--output-len", + type=int, + default=128, + help="Number of output tokens.", + ) + parser.add_argument( + "--request-rate", + type=float, + default=float("inf"), + help="Number of requests per second. If this is inf, " + "then all the requests are sent at time 0. " + "Otherwise, we use Poisson process or gamma distribution " + "to synthesize the request arrival times.", + ) + parser.add_argument( + "--burstiness", + type=float, + default=1.0, + help="Burstiness factor of the request generation. " + "Only take effect when request_rate is not inf. " + "Default value is 1, which follows Poisson process. " + "Otherwise, the request intervals follow a gamma distribution. " + "A lower burstiness value (0 < burstiness < 1) results in more " + "bursty requests. A higher burstiness value (burstiness > 1) " + "results in a more uniform arrival of requests.", + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument( + "--trust-remote-code", + action="store_true", + help="Trust remote code from huggingface", + ) + parser.add_argument( + "--disable-tqdm", + action="store_true", + help="Specify to disable tqdm progress bar.", + ) + parser.add_argument( + "--save-results", + action="store_true", + help="Specify to save benchmark results to a json file", + ) + parser.add_argument( + "--profile", + action="store_true", + help="Use Torch Profiler. The endpoint must be launched with " + "VLLM_TORCH_PROFILER_DIR to enable profiler.", + ) + parser.add_argument( + "--result-dir", + type=str, + default=None, + help="Specify directory to save benchmark json results." + "If not specified, results are saved in the current directory.", + ) + parser.add_argument( + "--result-filename", + type=str, + default=None, + help="Specify the filename to save benchmark json results." + "If not specified, results will be saved in " + "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json" + " format.", + ) + parser.add_argument( + "--ignore-eos", + action="store_true", + help="Set ignore_eos flag when sending the benchmark request." + "Warning: ignore_eos is not supported in deepspeed_mii and tgi.", + ) + parser.add_argument( + "--percentile-metrics", + type=str, + default="ttft,tpot,itl", + help="Comma-separated list of selected metrics to report percentils. " + "This argument specifies the metrics to report percentiles. " + 'Allowed metric names are "ttft", "tpot", "itl", "e2el". ' + 'Default value is "ttft,tpot,itl".', + ) + parser.add_argument( + "--metric-percentiles", + type=str, + default="99", + help="Comma-separated list of percentiles for selected metrics. " + 'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". ' + 'Default value is "99". ' + 'Use "--percentile-metrics" to select metrics.', + ) + parser.add_argument( + "--goodput", + nargs="+", + required=False, + help='Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is in " + 'milliseconds. Multiple "KEY:VALUE" pairs can be provided, ' + "separated by spaces. Allowed request level metric names are " + '"ttft", "tpot", "e2el". For more context on the definition of ' + "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 " + "and the blog: https://hao-ai-lab.github.io/blogs/distserve", + ) + + parser.add_argument( + "--no-structured-output", + action="store_true", + default=False, + help="Whether to disable JSON decoding or not.", + ) + parser.add_argument( + "--structured-output-ratio", + type=float, + default=1.0, + help="Ratio of Structured Outputs requests", + ) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test/benchmarks/benchmark_throughput.py b/online_apiserver_test/benchmarks/benchmark_throughput.py new file mode 100644 index 0000000000000000000000000000000000000000..0ded34c70badd28d972b3c993ddb7881b7b94aff --- /dev/null +++ b/online_apiserver_test/benchmarks/benchmark_throughput.py @@ -0,0 +1,730 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Benchmark offline inference throughput.""" + +import argparse +import dataclasses +import json +import os +import random +import time +import warnings +from typing import Any, Optional, Union + +import torch +import uvloop +from tqdm import tqdm +from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase + +from benchmark_dataset import ( + AIMODataset, + BurstGPTDataset, + ConversationDataset, + InstructCoderDataset, + RandomDataset, + SampleRequest, + ShareGPTDataset, + SonnetDataset, + VisionArenaDataset, +) +from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json +from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs +from vllm.entrypoints.openai.api_server import ( + build_async_engine_client_from_engine_args, +) +from vllm.inputs import TextPrompt, TokensPrompt +from vllm.lora.request import LoRARequest +from vllm.outputs import RequestOutput +from vllm.sampling_params import BeamSearchParams +from vllm.utils import FlexibleArgumentParser, merge_async_iterators + + +def run_vllm( + requests: list[SampleRequest], + n: int, + engine_args: EngineArgs, + disable_detokenize: bool = False, +) -> tuple[float, Optional[list[RequestOutput]]]: + from vllm import LLM, SamplingParams + + llm = LLM(**dataclasses.asdict(engine_args)) + assert all( + llm.llm_engine.model_config.max_model_len + >= (request.prompt_len + request.expected_output_len) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of" + " prompt_len and expected_output_len for all requests." + ) + # Add the requests to the engine. + prompts: list[Union[TextPrompt, TokensPrompt]] = [] + sampling_params: list[SamplingParams] = [] + for request in requests: + prompts.append( + TokensPrompt( + prompt_token_ids=request.prompt["prompt_token_ids"], + multi_modal_data=request.multi_modal_data, + ) + if "prompt_token_ids" in request.prompt + else TextPrompt( + prompt=request.prompt, multi_modal_data=request.multi_modal_data + ) + ) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=request.expected_output_len, + detokenize=not disable_detokenize, + ) + ) + lora_requests: Optional[list[LoRARequest]] = None + if engine_args.enable_lora: + lora_requests = [request.lora_request for request in requests] + + use_beam_search = False + + outputs = None + if not use_beam_search: + start = time.perf_counter() + outputs = llm.generate( + prompts, sampling_params, lora_request=lora_requests, use_tqdm=True + ) + end = time.perf_counter() + else: + assert lora_requests is None, "BeamSearch API does not support LoRA" + prompts = [request.prompt for request in requests] + # output_len should be the same for all requests. + output_len = requests[0].expected_output_len + for request in requests: + assert request.expected_output_len == output_len + start = time.perf_counter() + llm.beam_search( + prompts, + BeamSearchParams( + beam_width=n, + max_tokens=output_len, + ignore_eos=True, + ), + ) + end = time.perf_counter() + return end - start, outputs + + +def run_vllm_chat( + requests: list[SampleRequest], + n: int, + engine_args: EngineArgs, + disable_detokenize: bool = False, +) -> tuple[float, list[RequestOutput]]: + """ + Run vLLM chat benchmark. This function is recommended ONLY for benchmarking + multimodal models as it properly handles multimodal inputs and chat + formatting. For non-multimodal models, use run_vllm() instead. + """ + from vllm import LLM, SamplingParams + + llm = LLM(**dataclasses.asdict(engine_args)) + + assert all( + llm.llm_engine.model_config.max_model_len + >= (request.prompt_len + request.expected_output_len) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of " + "prompt_len and expected_output_len for all requests." + ) + + prompts = [] + sampling_params: list[SamplingParams] = [] + for request in requests: + prompts.append(request.prompt) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=request.expected_output_len, + detokenize=not disable_detokenize, + ) + ) + start = time.perf_counter() + outputs = llm.chat(prompts, sampling_params, use_tqdm=True) + end = time.perf_counter() + return end - start, outputs + + +async def run_vllm_async( + requests: list[SampleRequest], + n: int, + engine_args: AsyncEngineArgs, + disable_frontend_multiprocessing: bool = False, + disable_detokenize: bool = False, +) -> float: + from vllm import SamplingParams + + async with build_async_engine_client_from_engine_args( + engine_args, disable_frontend_multiprocessing + ) as llm: + model_config = await llm.get_model_config() + assert all( + model_config.max_model_len + >= (request.prompt_len + request.expected_output_len) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of" + " prompt_len and expected_output_len for all requests." + ) + + # Add the requests to the engine. + prompts: list[Union[TextPrompt, TokensPrompt]] = [] + sampling_params: list[SamplingParams] = [] + lora_requests: list[Optional[LoRARequest]] = [] + for request in requests: + prompts.append( + TokensPrompt( + prompt_token_ids=request.prompt["prompt_token_ids"], + multi_modal_data=request.multi_modal_data, + ) + if "prompt_token_ids" in request.prompt + else TextPrompt( + prompt=request.prompt, multi_modal_data=request.multi_modal_data + ) + ) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=request.expected_output_len, + detokenize=not disable_detokenize, + ) + ) + lora_requests.append(request.lora_request) + + generators = [] + start = time.perf_counter() + for i, (prompt, sp, lr) in enumerate( + zip(prompts, sampling_params, lora_requests) + ): + generator = llm.generate(prompt, sp, lora_request=lr, request_id=f"test{i}") + generators.append(generator) + all_gens = merge_async_iterators(*generators) + async for i, res in all_gens: + pass + end = time.perf_counter() + return end - start + + +def run_hf( + requests: list[SampleRequest], + model: str, + tokenizer: PreTrainedTokenizerBase, + n: int, + max_batch_size: int, + trust_remote_code: bool, + disable_detokenize: bool = False, +) -> float: + llm = AutoModelForCausalLM.from_pretrained( + model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code + ) + if llm.config.model_type == "llama": + # To enable padding in the HF backend. + tokenizer.pad_token = tokenizer.eos_token + llm = llm.cuda() + + pbar = tqdm(total=len(requests)) + start = time.perf_counter() + batch: list[str] = [] + max_prompt_len = 0 + max_output_len = 0 + for i in range(len(requests)): + prompt = requests[i].prompt + prompt_len = requests[i].prompt_len + output_len = requests[i].expected_output_len + # Add the prompt to the batch. + batch.append(prompt) + max_prompt_len = max(max_prompt_len, prompt_len) + max_output_len = max(max_output_len, output_len) + if len(batch) < max_batch_size and i != len(requests) - 1: + # Check if we can add more requests to the batch. + next_prompt_len = requests[i + 1].prompt_len + next_output_len = requests[i + 1].expected_output_len + if ( + max(max_prompt_len, next_prompt_len) + + max(max_output_len, next_output_len) + ) <= 2048: + # We can add more requests to the batch. + continue + + # Generate the sequences. + input_ids = tokenizer(batch, return_tensors="pt", padding=True).input_ids + llm_outputs = llm.generate( + input_ids=input_ids.cuda(), + do_sample=True, + num_return_sequences=n, + temperature=1.0, + top_p=1.0, + use_cache=True, + max_new_tokens=max_output_len, + ) + if not disable_detokenize: + # Include the decoding time. + tokenizer.batch_decode(llm_outputs, skip_special_tokens=True) + pbar.update(len(batch)) + + # Clear the batch. + batch = [] + max_prompt_len = 0 + max_output_len = 0 + end = time.perf_counter() + return end - start + + +def run_mii( + requests: list[SampleRequest], + model: str, + tensor_parallel_size: int, + output_len: int, +) -> float: + from mii import client, serve + + llm = serve(model, tensor_parallel=tensor_parallel_size) + prompts = [request.prompt for request in requests] + + start = time.perf_counter() + llm.generate(prompts, max_new_tokens=output_len) + end = time.perf_counter() + client = client(model) + client.terminate_server() + return end - start + + +def save_to_pytorch_benchmark_format( + args: argparse.Namespace, results: dict[str, Any] +) -> None: + pt_records = convert_to_pytorch_benchmark_format( + args=args, + metrics={ + "requests_per_second": [results["requests_per_second"]], + "tokens_per_second": [results["tokens_per_second"]], + }, + extra_info={ + k: results[k] for k in ["elapsed_time", "num_requests", "total_num_tokens"] + }, + ) + if pt_records: + # Don't use json suffix here as we don't want CI to pick it up + pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json" + write_to_json(pt_file, pt_records) + + +def get_requests(args, tokenizer): + # Common parameters for all dataset types. + common_kwargs = { + "dataset_path": args.dataset_path, + "random_seed": args.seed, + } + sample_kwargs = { + "tokenizer": tokenizer, + "lora_path": args.lora_path, + "max_loras": args.max_loras, + "num_requests": args.num_prompts, + "input_len": args.input_len, + "output_len": args.output_len, + } + + if args.dataset_path is None or args.dataset_name == "random": + sample_kwargs["range_ratio"] = args.random_range_ratio + sample_kwargs["prefix_len"] = args.prefix_len + dataset_cls = RandomDataset + elif args.dataset_name == "sharegpt": + dataset_cls = ShareGPTDataset + if args.backend == "vllm-chat": + sample_kwargs["enable_multimodal_chat"] = True + elif args.dataset_name == "sonnet": + assert tokenizer.chat_template or tokenizer.default_chat_template, ( + "Tokenizer/model must have chat template for sonnet dataset." + ) + dataset_cls = SonnetDataset + sample_kwargs["prefix_len"] = args.prefix_len + sample_kwargs["return_prompt_formatted"] = True + elif args.dataset_name == "burstgpt": + dataset_cls = BurstGPTDataset + elif args.dataset_name == "hf": + if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS: + dataset_cls = VisionArenaDataset + common_kwargs["dataset_subset"] = None + common_kwargs["dataset_split"] = "train" + sample_kwargs["enable_multimodal_chat"] = True + elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS: + dataset_cls = InstructCoderDataset + common_kwargs["dataset_split"] = "train" + elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS: + dataset_cls = ConversationDataset + common_kwargs["dataset_subset"] = args.hf_subset + common_kwargs["dataset_split"] = args.hf_split + sample_kwargs["enable_multimodal_chat"] = True + elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS: + dataset_cls = AIMODataset + common_kwargs["dataset_subset"] = None + common_kwargs["dataset_split"] = "train" + else: + raise ValueError(f"Unknown dataset name: {args.dataset_name}") + # Remove None values + sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None} + return dataset_cls(**common_kwargs).sample(**sample_kwargs) + + +def main(args: argparse.Namespace): + if args.seed is None: + args.seed = 0 + print(args) + random.seed(args.seed) + # Sample the requests. + tokenizer = AutoTokenizer.from_pretrained( + args.tokenizer, trust_remote_code=args.trust_remote_code + ) + requests = get_requests(args, tokenizer) + is_multi_modal = any(request.multi_modal_data is not None for request in requests) + request_outputs: Optional[list[RequestOutput]] = None + if args.backend == "vllm": + if args.async_engine: + elapsed_time = uvloop.run( + run_vllm_async( + requests, + args.n, + AsyncEngineArgs.from_cli_args(args), + args.disable_frontend_multiprocessing, + args.disable_detokenize, + ) + ) + else: + elapsed_time, request_outputs = run_vllm( + requests, + args.n, + EngineArgs.from_cli_args(args), + args.disable_detokenize, + ) + elif args.backend == "hf": + assert args.tensor_parallel_size == 1 + elapsed_time = run_hf( + requests, + args.model, + tokenizer, + args.n, + args.hf_max_batch_size, + args.trust_remote_code, + args.disable_detokenize, + ) + elif args.backend == "mii": + elapsed_time = run_mii( + requests, args.model, args.tensor_parallel_size, args.output_len + ) + elif args.backend == "vllm-chat": + elapsed_time, request_outputs = run_vllm_chat( + requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize + ) + else: + raise ValueError(f"Unknown backend: {args.backend}") + + if request_outputs: + # Note: with the vllm and vllm-chat backends, + # we have request_outputs, which we use to count tokens. + total_prompt_tokens = 0 + total_output_tokens = 0 + for ro in request_outputs: + if not isinstance(ro, RequestOutput): + continue + total_prompt_tokens += ( + len(ro.prompt_token_ids) if ro.prompt_token_ids else 0 + ) + total_output_tokens += sum(len(o.token_ids) for o in ro.outputs if o) + total_num_tokens = total_prompt_tokens + total_output_tokens + else: + total_num_tokens = sum(r.prompt_len + r.expected_output_len for r in requests) + total_output_tokens = sum(r.expected_output_len for r in requests) + total_prompt_tokens = total_num_tokens - total_output_tokens + + if is_multi_modal and args.backend != "vllm-chat": + print( + "\033[91mWARNING\033[0m: Multi-modal request with " + f"{args.backend} backend detected. The " + "following metrics are not accurate because image tokens are not" + " counted. See vllm-project/vllm/issues/9778 for details." + ) + # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length. + # vllm-chat backend counts the image tokens now + + print( + f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, " + f"{total_num_tokens / elapsed_time:.2f} total tokens/s, " + f"{total_output_tokens / elapsed_time:.2f} output tokens/s" + ) + print(f"Total num prompt tokens: {total_prompt_tokens}") + print(f"Total num output tokens: {total_output_tokens}") + + # Output JSON results if specified + if args.output_json: + results = { + "elapsed_time": elapsed_time, + "num_requests": len(requests), + "total_num_tokens": total_num_tokens, + "requests_per_second": len(requests) / elapsed_time, + "tokens_per_second": total_num_tokens / elapsed_time, + } + with open(args.output_json, "w") as f: + json.dump(results, f, indent=4) + save_to_pytorch_benchmark_format(args, results) + + +def validate_args(args): + """ + Validate command-line arguments. + """ + + # === Deprecation and Defaulting === + if args.dataset is not None: + warnings.warn( + "The '--dataset' argument will be deprecated in the next release. " + "Please use '--dataset-name' and '--dataset-path' instead.", + stacklevel=2, + ) + args.dataset_path = args.dataset + + if not getattr(args, "tokenizer", None): + args.tokenizer = args.model + + # === Backend Validation === + valid_backends = {"vllm", "hf", "mii", "vllm-chat"} + if args.backend not in valid_backends: + raise ValueError(f"Unsupported backend: {args.backend}") + + # === Dataset Configuration === + if not args.dataset and not args.dataset_path: + print("When dataset path is not set, it will default to random dataset") + args.dataset_name = "random" + if args.input_len is None: + raise ValueError("input_len must be provided for a random dataset") + + # === Dataset Name Specific Checks === + # --hf-subset and --hf-split: only used + # when dataset_name is 'hf' + if args.dataset_name != "hf" and ( + getattr(args, "hf_subset", None) is not None + or getattr(args, "hf_split", None) is not None + ): + warnings.warn( + "--hf-subset and --hf-split will be ignored \ + since --dataset-name is not 'hf'.", + stacklevel=2, + ) + elif args.dataset_name == "hf": + if args.dataset_path in ( + VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys() + | ConversationDataset.SUPPORTED_DATASET_PATHS + ): + assert args.backend == "vllm-chat", ( + f"{args.dataset_path} needs to use vllm-chat as the backend." + ) # noqa: E501 + elif args.dataset_path in ( + InstructCoderDataset.SUPPORTED_DATASET_PATHS + | AIMODataset.SUPPORTED_DATASET_PATHS + ): + assert args.backend == "vllm", ( + f"{args.dataset_path} needs to use vllm as the backend." + ) # noqa: E501 + else: + raise ValueError(f"{args.dataset_path} is not supported by hf dataset.") + + # --random-range-ratio: only used when dataset_name is 'random' + if args.dataset_name != "random" and args.random_range_ratio is not None: + warnings.warn( + "--random-range-ratio will be ignored since \ + --dataset-name is not 'random'.", + stacklevel=2, + ) + + # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not + # set. + if ( + args.dataset_name not in {"random", "sonnet", None} + and args.prefix_len is not None + ): + warnings.warn( + "--prefix-len will be ignored since --dataset-name\ + is not 'random', 'sonnet', or not set.", + stacklevel=2, + ) + + # === LoRA Settings === + if getattr(args, "enable_lora", False) and args.backend != "vllm": + raise ValueError("LoRA benchmarking is only supported for vLLM backend") + if getattr(args, "enable_lora", False) and args.lora_path is None: + raise ValueError("LoRA path must be provided when enable_lora is True") + + # === Backend-specific Validations === + if args.backend == "hf" and args.hf_max_batch_size is None: + raise ValueError("HF max batch size is required for HF backend") + if args.backend != "hf" and args.hf_max_batch_size is not None: + raise ValueError("HF max batch size is only for HF backend.") + + if ( + args.backend in {"hf", "mii"} + and getattr(args, "quantization", None) is not None + ): + raise ValueError("Quantization is only for vLLM backend.") + + if args.backend == "mii" and args.dtype != "auto": + raise ValueError("dtype must be auto for MII backend.") + if args.backend == "mii" and args.n != 1: + raise ValueError("n must be 1 for MII backend.") + if args.backend == "mii" and args.tokenizer != args.model: + raise ValueError("Tokenizer must be the same as the model for MII backend.") + + # --data-parallel is not supported currently. + # https://github.com/vllm-project/vllm/issues/16222 + if args.data_parallel_size > 1: + raise ValueError( + "Data parallel is not supported in offline benchmark, \ + please use benchmark serving instead" + ) + + +def create_argument_parser(): + parser = FlexibleArgumentParser(description="Benchmark the throughput.") + parser.add_argument( + "--backend", + type=str, + choices=["vllm", "hf", "mii", "vllm-chat"], + default="vllm", + ) + parser.add_argument( + "--dataset-name", + type=str, + choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"], + help="Name of the dataset to benchmark on.", + default="sharegpt", + ) + parser.add_argument( + "--dataset", + type=str, + default=None, + help="Path to the ShareGPT dataset, will be deprecated in\ + the next release. The dataset is expected to " + "be a json in form of list[dict[..., conversations: " + "list[dict[..., value: ]]]]", + ) + parser.add_argument( + "--dataset-path", type=str, default=None, help="Path to the dataset" + ) + parser.add_argument( + "--input-len", + type=int, + default=None, + help="Input prompt length for each request", + ) + parser.add_argument( + "--output-len", + type=int, + default=None, + help="Output length for each request. Overrides the " + "output length from the dataset.", + ) + parser.add_argument( + "--n", type=int, default=1, help="Number of generated sequences per prompt." + ) + parser.add_argument( + "--num-prompts", type=int, default=1000, help="Number of prompts to process." + ) + parser.add_argument( + "--hf-max-batch-size", + type=int, + default=None, + help="Maximum batch size for HF backend.", + ) + parser.add_argument( + "--output-json", + type=str, + default=None, + help="Path to save the throughput results in JSON format.", + ) + parser.add_argument( + "--async-engine", + action="store_true", + default=False, + help="Use vLLM async engine rather than LLM class.", + ) + parser.add_argument( + "--disable-frontend-multiprocessing", + action="store_true", + default=False, + help="Disable decoupled async engine frontend.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize the response (i.e. do not include " + "detokenization time in the measurement)" + ), + ) + # LoRA + parser.add_argument( + "--lora-path", + type=str, + default=None, + help="Path to the LoRA adapters to use. This can be an absolute path, " + "a relative path, or a Hugging Face model identifier.", + ) + parser.add_argument( + "--prefix-len", + type=int, + default=None, + help=f"Number of prefix tokens to be used in RandomDataset " + "and SonnetDataset. For RandomDataset, the total input " + "length is the sum of prefix-len (default: " + f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length " + "sampled from [input_len * (1 - range_ratio), " + "input_len * (1 + range_ratio)]. For SonnetDataset, " + f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) " + "controls how much of the input is fixed lines versus " + "random lines, but the total input length remains approximately " + "input_len tokens.", + ) + # random dataset + parser.add_argument( + "--random-range-ratio", + type=float, + default=None, + help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) " + "for sampling input/output length, " + "used only for RandomDataset. Must be in the range [0, 1) to " + "define a symmetric sampling range " + "[length * (1 - range_ratio), length * (1 + range_ratio)].", + ) + + # hf dtaset + parser.add_argument( + "--hf-subset", type=str, default=None, help="Subset of the HF dataset." + ) + parser.add_argument( + "--hf-split", type=str, default=None, help="Split of the HF dataset." + ) + + parser = AsyncEngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + if args.tokenizer is None: + args.tokenizer = args.model + validate_args(args) + main(args) diff --git a/online_apiserver_test/benchmarks/benchmark_utils.py b/online_apiserver_test/benchmarks/benchmark_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..283f938df50af101fa2ff82f3c46e123b25e5524 --- /dev/null +++ b/online_apiserver_test/benchmarks/benchmark_utils.py @@ -0,0 +1,74 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import json +import math +import os +from typing import Any + + +def convert_to_pytorch_benchmark_format( + args: argparse.Namespace, metrics: dict[str, list], extra_info: dict[str, Any] +) -> list: + """ + Save the benchmark results in the format used by PyTorch OSS benchmark with + on metric per record + https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database + """ + records = [] + if not os.environ.get("SAVE_TO_PYTORCH_BENCHMARK_FORMAT", False): + return records + + for name, benchmark_values in metrics.items(): + record = { + "benchmark": { + "name": "vLLM benchmark", + "extra_info": { + "args": vars(args), + }, + }, + "model": { + "name": args.model, + }, + "metric": { + "name": name, + "benchmark_values": benchmark_values, + "extra_info": extra_info, + }, + } + + tp = record["benchmark"]["extra_info"]["args"].get("tensor_parallel_size") + # Save tensor_parallel_size parameter if it's part of the metadata + if not tp and "tensor_parallel_size" in extra_info: + record["benchmark"]["extra_info"]["args"]["tensor_parallel_size"] = ( + extra_info["tensor_parallel_size"] + ) + + records.append(record) + + return records + + +class InfEncoder(json.JSONEncoder): + def clear_inf(self, o: Any): + if isinstance(o, dict): + return {k: self.clear_inf(v) for k, v in o.items()} + elif isinstance(o, list): + return [self.clear_inf(v) for v in o] + elif isinstance(o, float) and math.isinf(o): + return "inf" + return o + + def iterencode(self, o: Any, *args, **kwargs) -> Any: + return super().iterencode(self.clear_inf(o), *args, **kwargs) + + +def write_to_json(filename: str, records: list) -> None: + with open(filename, "w") as f: + json.dump( + records, + f, + cls=InfEncoder, + default=lambda o: f"<{type(o).__name__} object is not JSON serializable>", + ) diff --git a/online_apiserver_test/benchmarks/cutlass_benchmarks/sparse_benchmarks.py b/online_apiserver_test/benchmarks/cutlass_benchmarks/sparse_benchmarks.py new file mode 100644 index 0000000000000000000000000000000000000000..9ec270bbd2e988b459f99d876215984bf125d0f9 --- /dev/null +++ b/online_apiserver_test/benchmarks/cutlass_benchmarks/sparse_benchmarks.py @@ -0,0 +1,516 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import itertools +import pickle as pkl +import time +from collections.abc import Iterable +from typing import Callable + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import make_rand_sparse_tensors +from weight_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512] +DEFAULT_TP_SIZES = [1] + + +# bench +def bench_fn( + label: str, sub_label: str, description: str, fn: Callable, *args, **kwargs +) -> TMeasurement: + min_run_time = 1 + + globals = { + "args": args, + "kwargs": kwargs, + "fn": fn, + } + return TBenchmark.Timer( + stmt="fn(*args, **kwargs)", + globals=globals, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + +def bench_int8( + dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str +) -> Iterable[TMeasurement]: + assert dtype == torch.int8 + b_compressed, e, a, b = make_rand_sparse_tensors(torch.int8, m, n, k) + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + + out = ops.cutlass_scaled_sparse_mm( + a, b_compressed, e, scale_a, scale_b, torch.bfloat16 + ) + out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16) + + if not torch.allclose(out, out_ref): + print("Incorrect results") + print(out) + print(out_ref) + else: + print("Correct results") + + timers = [] + # pytorch impl - bfloat16 + timers.append( + bench_fn( + label, + sub_label, + "pytorch_bf16_bf16_bf16_matmul-no-scales", + torch.mm, + a.to(dtype=torch.bfloat16), + b.to(dtype=torch.bfloat16), + ) + ) + + # pytorch impl - float16 + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp16_fp16_fp16_matmul-no-scales", + torch.mm, + a.to(dtype=torch.float16), + b.to(dtype=torch.float16), + ) + ) + + # cutlass impl + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_mm", + ops.cutlass_scaled_mm, + a, + b, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_mm_bias", + ops.cutlass_scaled_mm, + a, + b, + scale_a, + scale_b, + torch.bfloat16, + bias, + ) + ) + + # cutlass sparse impl + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_sparse_mm", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass sparse with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_sparse_mm_bias", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + bias, + ) + ) + + return timers + + +def bench_fp8( + dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str +) -> Iterable[TMeasurement]: + assert dtype == torch.float8_e4m3fn + b_compressed, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n, k) + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + + out = ops.cutlass_scaled_sparse_mm( + a, b_compressed, e, scale_a, scale_b, torch.bfloat16 + ) + out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16) + + if not torch.allclose(out, out_ref): + print("Incorrect results") + print(out) + print(out_ref) + else: + print("Correct results") + + timers = [] + + # pytorch impl w. bf16 + timers.append( + bench_fn( + label, + sub_label, + "pytorch_bf16_bf16_bf16_matmul-no-scales", + torch.mm, + a.to(dtype=torch.bfloat16, device="cuda"), + b.to(dtype=torch.bfloat16, device="cuda"), + ) + ) + + # pytorch impl: bf16 output, without fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_bf16_scaled_mm", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.bfloat16, + ) + ) + + # pytorch impl: bf16 output, with fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.bfloat16, + use_fast_accum=True, + ) + ) + + # pytorch impl: fp16 output, without fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_fp16_scaled_mm", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.float16, + ) + ) + + # pytorch impl: fp16 output, with fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.float16, + use_fast_accum=True, + ) + ) + + # cutlass impl: bf16 output + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_bf16_scaled_mm", + ops.cutlass_scaled_mm, + a, + b, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass impl: bf16 output + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_bf16_scaled_sparse_mm", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass impl: fp16 output + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_fp16_scaled_sparse_mm", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.float16, + ) + ) + + # cutlass impl: bf16 output, with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_bf16_scaled_sparse_mm_bias", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + bias, + ) + ) + + # cutlass impl: fp16 output, with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_fp16_scaled_sparse_mm_bias", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.float16, + bias.to(dtype=torch.float16), + ) + ) + + return timers + + +def bench( + dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str +) -> Iterable[TMeasurement]: + if dtype == torch.int8: + return bench_int8(dtype, m, k, n, label, sub_label) + if dtype == torch.float8_e4m3fn: + return bench_fp8(dtype, m, k, n, label, sub_label) + raise ValueError("unsupported type") + + +# runner +def print_timers(timers: Iterable[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def run( + dtype: torch.dtype, MKNs: Iterable[tuple[int, int, int]] +) -> Iterable[TMeasurement]: + results = [] + for m, k, n in MKNs: + timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm", f"MKN=({m}x{k}x{n})") + print_timers(timers) + results.extend(timers) + + return results + + +# output makers +def make_output( + data: Iterable[TMeasurement], + MKNs: Iterable[tuple[int, int, int]], + base_description: str, + timestamp=None, +): + print(f"== All Results {base_description} ====") + print_timers(data) + + # pickle all the results + timestamp = int(time.time()) if timestamp is None else timestamp + with open(f"{base_description}-{timestamp}.pkl", "wb") as f: + pkl.dump(data, f) + + +# argparse runners + + +def run_square_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment)) + MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes)) + data = run(args.dtype, MKNs) + + make_output(data, MKNs, f"square_bench-{args.dtype}") + + +def run_range_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment)) + n = len(dim_sizes) + Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes + Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes + Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes + MKNs = list(zip(Ms, Ks, Ns)) + data = run(args.dtype, MKNs) + + make_output(data, MKNs, f"range_bench-{args.dtype}") + + +def run_model_bench(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]: + KNs = [] + for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]): + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + KNs.append(KN) + return KNs + + model_bench_data = [] + models_tps = list(itertools.product(args.models, args.tp_sizes)) + for model, tp_size in models_tps: + Ms = args.batch_sizes + KNs = model_shapes(model, tp_size) + MKNs = [] + for m in Ms: + for k, n in KNs: + MKNs.append((m, k, n)) + + data = run(args.dtype, MKNs) + model_bench_data.append(data) + + # Print all results + for data, model_tp in zip(model_bench_data, models_tps): + model, tp_size = model_tp + print(f"== Results {args.dtype} {model}-TP{tp_size} ====") + print_timers(data) + + timestamp = int(time.time()) + + all_data = [] + for d in model_bench_data: + all_data.extend(d) + # pickle all data + with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f: + pkl.dump(all_data, f) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "int8": + return torch.int8 + if dt == "fp8": + return torch.float8_e4m3fn + raise ValueError("unsupported dtype") + + parser = FlexibleArgumentParser( + description=""" +Benchmark Cutlass GEMM. + + To run square GEMMs: + python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64 + + To run constant N and K and sweep M: + python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384 + + To run dimensions from a model: + python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1 + + Output: + - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs. + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + + parser.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['int8', 'fp8']", + ) + subparsers = parser.add_subparsers(dest="cmd") + + square_parser = subparsers.add_parser("square_bench") + square_parser.add_argument("--dim-start", type=int, required=True) + square_parser.add_argument("--dim-end", type=int, required=True) + square_parser.add_argument("--dim-increment", type=int, required=True) + square_parser.set_defaults(func=run_square_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--dim-start", type=int, required=True) + range_parser.add_argument("--dim-end", type=int, required=True) + range_parser.add_argument("--dim-increment", type=int, required=True) + range_parser.add_argument("--m-constant", type=int, default=None) + range_parser.add_argument("--n-constant", type=int, default=None) + range_parser.add_argument("--k-constant", type=int, default=None) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/online_apiserver_test/benchmarks/cutlass_benchmarks/utils.py b/online_apiserver_test/benchmarks/cutlass_benchmarks/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b4f3c6bf94eda0e1bf1d253def3b17d273415dcc --- /dev/null +++ b/online_apiserver_test/benchmarks/cutlass_benchmarks/utils.py @@ -0,0 +1,100 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +# Cutlass bench utils +from collections.abc import Iterable + +import torch + +import vllm._custom_ops as ops + + +def to_fp8(tensor: torch.Tensor) -> torch.Tensor: + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to( + dtype=torch.float8_e4m3fn + ) + + +def to_int8(tensor: torch.Tensor) -> torch.Tensor: + return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8) + + +def to_bf16(tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(dtype=torch.bfloat16) + + +def to_fp16(tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(dtype=torch.float16) + + +def make_rand_tensors( + dtype: torch.dtype, m: int, n: int, k: int +) -> tuple[torch.Tensor, torch.Tensor]: + a = torch.randn((m, k), device="cuda") * 5 + b = torch.randn((n, k), device="cuda").t() * 5 + + if dtype == torch.int8: + return to_int8(a), to_int8(b) + if dtype == torch.float8_e4m3fn: + return to_fp8(a), to_fp8(b) + + raise ValueError("unsupported dtype") + + +def prune_to_2_4(tensor): + # Reshape tensor to [N, 4] where N is number of groups of 4 + original_shape = tensor.shape + reshaped = tensor.reshape(-1, 4) + + # Get indices of top 2 absolute values in each group of 4 + _, indices = torch.topk(torch.abs(reshaped), k=2, dim=1) + + # Create binary mask + mask = torch.zeros_like(reshaped) + mask.scatter_(dim=1, index=indices, src=torch.ones_like(indices, dtype=mask.dtype)) + + # Apply mask and reshape back + pruned = reshaped * mask + + # Turn all -0.0 to 0.0 + pruned[pruned == -0.0] = 0.0 + + return pruned.reshape(original_shape) + + +def make_rand_sparse_tensors( + dtype: torch.dtype, m: int, n: int, k: int +) -> tuple[torch.Tensor, torch.Tensor]: + a = torch.randn((m, k), device="cuda") * 5 + b = torch.randn((n, k), device="cuda").t() * 5 + + b = prune_to_2_4(b.t()).t() + + if dtype == torch.int8: + a, b = to_int8(a), to_int8(b) + elif dtype == torch.float8_e4m3fn: + a, b = to_fp8(a), to_fp8(b) + elif dtype == torch.float16: + a, b = to_fp16(a), to_fp16(b) + elif dtype == torch.bfloat16: + a, b = to_bf16(a), to_bf16(b) + else: + raise ValueError("unsupported dtype") + + b_compressed, e = ops.cutlass_sparse_compress(b.t()) + + # Compressed B, Metadata, Original A, B + return b_compressed, e, a, b + + +def make_n_rand_sparse_tensors( + num_tensors: int, dtype: torch.dtype, m: int, n: int, k: int +) -> tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]: + ABs = [] + for _ in range(num_tensors): + b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k) + if b_comp is not None: + ABs.append(make_rand_sparse_tensors(dtype, m, n, k)) + BComps, Es, As, Bs = zip(*ABs) + return list(BComps), list(Es), list(As), list(Bs) diff --git a/online_apiserver_test/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py b/online_apiserver_test/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py new file mode 100644 index 0000000000000000000000000000000000000000..a5a5b52f603977cc0c967e455724bc7c6a2bf585 --- /dev/null +++ b/online_apiserver_test/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py @@ -0,0 +1,372 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import itertools +import pickle as pkl +import time +from collections.abc import Iterable +from typing import Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import make_rand_tensors +from weight_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + w8a8_block_fp8_matmul, +) +from vllm.utils import FlexibleArgumentParser, cdiv + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512] +DEFAULT_TP_SIZES = [1] + + +# bench +def bench_fn( + label: str, sub_label: str, description: str, fn: Callable, *args, **kwargs +) -> TMeasurement: + min_run_time = 1 + + globals = { + "args": args, + "kwargs": kwargs, + "fn": fn, + } + return TBenchmark.Timer( + stmt="fn(*args, **kwargs)", + globals=globals, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + +def bench_int8( + dtype: torch.dtype, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + """Benchmark INT8-based kernels.""" + assert dtype == torch.int8 + a, b = make_rand_tensors(torch.int8, m, n, k) + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + azp = torch.zeros((m,), device="cuda", dtype=torch.int32) + azp_adj = torch.zeros((n,), device="cuda", dtype=torch.int32) + + bench_fns = { + "pytorch_bf16_bf16_bf16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16) + ), + "pytorch_fp16_fp16_fp16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.float16), b.to(dtype=torch.float16) + ), + "cutlass_i8_i8_bf16_scaled_mm": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16 + ), + "cutlass_i8_i8_bf16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16, bias + ), + "cutlass_i8_i8_bf16_scaled_mm_azp": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj + ), + "cutlass_i8_i8_bf16_scaled_mm_azp_bias": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj, None, bias + ), + "cutlass_i8_i8_bf16_scaled_mm_azp_pt": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj, azp + ), + "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj, azp, bias + ), + } + + timers = [] + for name, fn in bench_fns.items(): + # If bench_kernels is None, run all. Otherwise, run only exact matches. + if bench_kernels is None or name in bench_kernels: + print(f"Running {name}") + timers.append(bench_fn(label, sub_label, name, fn)) + + return timers + + +def bench_fp8( + dtype: torch.dtype, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + """Benchmark FP8-based kernels.""" + assert dtype == torch.float8_e4m3fn + a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k) + a_cont = a.contiguous() + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + + block_scale_a = torch.rand((m, cdiv(k, 128)), device="cuda", dtype=torch.float32) + block_scale_b = torch.rand( + cdiv(k, 128), cdiv(n, 128), device="cuda", dtype=torch.float32 + ) + block_scale_a_M_major = block_scale_a.t().contiguous().t() + block_scale_b_K_major = block_scale_b.t().contiguous().t() + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + + print(m, k, n) + + bench_fns = { + "pytorch_bf16_bf16_bf16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16) + ), + "pytorch_fp16_fp16_fp16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.float16), b.to(dtype=torch.float16) + ), + "pytorch_fp8_fp8_fp16_scaled_mm": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.float16 + ), + "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.float16, use_fast_accum=True + ), + "pytorch_fp8_fp8_bf16_scaled_mm": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.bfloat16 + ), + "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.bfloat16, use_fast_accum=True + ), + "cutlass_fp8_fp8_bf16_scaled_mm": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16 + ), + "cutlass_fp8_fp8_fp16_scaled_mm": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.float16 + ), + "cutlass_fp8_fp8_bf16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16, bias + ), + "cutlass_fp8_fp8_fp16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.float16, bias.to(dtype=torch.float16) + ), + "triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_block_fp8_matmul( + a_cont, b.t(), block_scale_a, block_scale_b.t(), (128, 128) + ), + "cutlass_fp8_fp8_fp16_scaled_mm_blockwise": lambda: ops.cutlass_scaled_mm( + a, b, block_scale_a_M_major, block_scale_b_K_major, torch.float16 + ), + } + + timers = [] + for name, fn in bench_fns.items(): + # If bench_kernels is None, run all. Otherwise, run only exact matches. + if bench_kernels is None or name in bench_kernels: + print(f"Running {name}") + timers.append(bench_fn(label, sub_label, name, fn)) + + return timers + + +def bench( + dtype: torch.dtype, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + if dtype == torch.int8: + return bench_int8(dtype, m, k, n, label, sub_label, bench_kernels) + if dtype == torch.float8_e4m3fn: + return bench_fp8(dtype, m, k, n, label, sub_label, bench_kernels) + raise ValueError("unsupported type") + + +# runner +def print_timers(timers: Iterable[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def run( + dtype: torch.dtype, + MKNs: Iterable[tuple[int, int, int]], + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + results = [] + for m, k, n in MKNs: + timers = bench( + dtype, + m, + k, + n, + f"scaled-{dtype}-gemm", + f"MKN=({m}x{k}x{n})", + bench_kernels=bench_kernels, + ) + print_timers(timers) + results.extend(timers) + return results + + +def make_output( + data: Iterable[TMeasurement], + MKNs: Iterable[tuple[int, int, int]], + base_description: str, + timestamp=None, +): + print(f"== All Results {base_description} ====") + print_timers(data) + + # pickle all the results + timestamp = int(time.time()) if timestamp is None else timestamp + with open(f"{base_description}-{timestamp}.pkl", "wb") as f: + pkl.dump(data, f) + + +def run_square_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment)) + MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes)) + data = run(args.dtype, MKNs, bench_kernels=args.kernels) + make_output(data, MKNs, f"square_bench-{args.dtype}") + + +def run_range_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment)) + n = len(dim_sizes) + Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes + Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes + Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes + MKNs = list(zip(Ms, Ks, Ns)) + data = run(args.dtype, MKNs, bench_kernels=args.kernels) + make_output(data, MKNs, f"range_bench-{args.dtype}") + + +def run_model_bench(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]: + KNs = [] + for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]): + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + KNs.append(KN) + return KNs + + model_bench_data = [] + models_tps = list(itertools.product(args.models, args.tp_sizes)) + for model, tp_size in models_tps: + Ms = args.batch_sizes + KNs = model_shapes(model, tp_size) + MKNs = [] + for m in Ms: + for k, n in KNs: + MKNs.append((m, k, n)) + + data = run(args.dtype, MKNs, bench_kernels=args.kernels) + model_bench_data.append(data) + + # Print all results + for data, model_tp in zip(model_bench_data, models_tps): + model, tp_size = model_tp + print(f"== Results {args.dtype} {model}-TP{tp_size} ====") + print_timers(data) + + timestamp = int(time.time()) + + all_data = [] + for d in model_bench_data: + all_data.extend(d) + # pickle all data + with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f: + pkl.dump(all_data, f) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "int8": + return torch.int8 + if dt == "fp8": + return torch.float8_e4m3fn + raise ValueError("unsupported dtype") + + parser = FlexibleArgumentParser( + description=""" +Benchmark Cutlass GEMM. + + To run square GEMMs: + python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64 + + To run constant N and K and sweep M: + python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384 + + To run dimensions from a model: + python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1 + + Output: + - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs. + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + + parser.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['int8', 'fp8']", + ) + parser.add_argument( + "--kernels", + nargs="+", + type=str, + default=None, + help="Exact names of the kernels to benchmark. If not set, runs all kernels.", + ) + + subparsers = parser.add_subparsers(dest="cmd") + + square_parser = subparsers.add_parser("square_bench") + square_parser.add_argument("--dim-start", type=int, required=True) + square_parser.add_argument("--dim-end", type=int, required=True) + square_parser.add_argument("--dim-increment", type=int, required=True) + square_parser.set_defaults(func=run_square_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--dim-start", type=int, required=True) + range_parser.add_argument("--dim-end", type=int, required=True) + range_parser.add_argument("--dim-increment", type=int, required=True) + range_parser.add_argument("--m-constant", type=int, default=None) + range_parser.add_argument("--n-constant", type=int, default=None) + range_parser.add_argument("--k-constant", type=int, default=None) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/online_apiserver_test/benchmarks/cutlass_benchmarks/weight_shapes.py b/online_apiserver_test/benchmarks/cutlass_benchmarks/weight_shapes.py new file mode 100644 index 0000000000000000000000000000000000000000..25b96ef56620ea7dbd97846cdd57ab0e97a6dfd1 --- /dev/null +++ b/online_apiserver_test/benchmarks/cutlass_benchmarks/weight_shapes.py @@ -0,0 +1,46 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +# Weight Shapes are in the format +# ([K, N], TP_SPLIT_DIM) +# Example: +# A shape of ([14336, 4096], 0) indicates the following GEMM shape, +# - TP1 : K = 14336, N = 4096 +# - TP2 : K = 7168, N = 4096 +# A shape of ([4096, 6144], 1) indicates the following GEMM shape, +# - TP1 : K = 4096, N = 6144 +# - TP4 : K = 4096, N = 1536 + +# TP1 shapes +WEIGHT_SHAPES = { + "mistralai/Mistral-7B-v0.1": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-7b-hf": [ + ([4096, 12288], 1), + ([4096, 4096], 0), + ([4096, 22016], 1), + ([11008, 4096], 0), + ], + "meta-llama/Llama-3-8b": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-13b-hf": [ + ([5120, 15360], 1), + ([5120, 5120], 0), + ([5120, 27648], 1), + ([13824, 5120], 0), + ], + "meta-llama/Llama-2-70b-hf": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 57344], 1), + ([28672, 8192], 0), + ], +} diff --git a/online_apiserver_test/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh b/online_apiserver_test/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..94999630bae122fa86d68c77cf9fc3941fbe4b14 --- /dev/null +++ b/online_apiserver_test/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh @@ -0,0 +1,145 @@ +#!/bin/bash + +# benchmark the overhead of disaggregated prefill. +# methodology: +# - send all request to prefill vLLM instance. It will buffer KV cache. +# - then send all request to decode instance. +# - The TTFT of decode instance is the overhead. + +set -ex + +kill_gpu_processes() { + # kill all processes on GPU. + pgrep pt_main_thread | xargs -r kill -9 + pgrep python3 | xargs -r kill -9 + sleep 10 + + # remove vllm config file + rm -rf ~/.config/vllm + + # Print the GPU memory usage + # so that we know if all GPU processes are killed. + gpu_memory_usage=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0) + # The memory usage should be 0 MB. + echo "GPU 0 Memory Usage: $gpu_memory_usage MB" +} + +wait_for_server() { + # wait for vllm server to start + # return 1 if vllm server crashes + local port=$1 + timeout 1200 bash -c " + until curl -s localhost:${port}/v1/completions > /dev/null; do + sleep 1 + done" && return 0 || return 1 +} + + +benchmark() { + + export VLLM_LOGGING_LEVEL=DEBUG + export VLLM_HOST_IP=$(hostname -I | awk '{print $1}') + + # compare chunked prefill with disaggregated prefill + + results_folder="./results" + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + dataset_name="sonnet" + dataset_path="../sonnet_4x.txt" + num_prompts=10 + qps=$1 + prefix_len=50 + input_len=2048 + output_len=$2 + + + CUDA_VISIBLE_DEVICES=0 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8100 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + + CUDA_VISIBLE_DEVICES=1 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8200 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + wait_for_server 8100 + wait_for_server 8200 + + # let the prefill instance finish prefill + python3 ../benchmark_serving.py \ + --backend vllm \ + --model $model \ + --dataset-name $dataset_name \ + --dataset-path $dataset_path \ + --sonnet-input-len $input_len \ + --sonnet-output-len "$output_len" \ + --sonnet-prefix-len $prefix_len \ + --num-prompts $num_prompts \ + --port 8100 \ + --save-result \ + --result-dir $results_folder \ + --result-filename disagg_prefill_tp1.json \ + --request-rate "inf" + + + # send the request to decode. + # The TTFT of this command will be the overhead of disagg prefill impl. + python3 ../benchmark_serving.py \ + --backend vllm \ + --model $model \ + --dataset-name $dataset_name \ + --dataset-path $dataset_path \ + --sonnet-input-len $input_len \ + --sonnet-output-len "$output_len" \ + --sonnet-prefix-len $prefix_len \ + --num-prompts $num_prompts \ + --port 8200 \ + --save-result \ + --result-dir $results_folder \ + --result-filename disagg_prefill_tp1_overhead.json \ + --request-rate "$qps" + kill_gpu_processes + +} + + +main() { + + (which wget && which curl) || (apt-get update && apt-get install -y wget curl) + (which jq) || (apt-get -y install jq) + (which socat) || (apt-get -y install socat) + + pip install quart httpx datasets + + cd "$(dirname "$0")" + + cd .. + # create sonnet-4x.txt + echo "" > sonnet_4x.txt + for _ in {1..4} + do + cat sonnet.txt >> sonnet_4x.txt + done + cd disagg_benchmarks + + rm -rf results + mkdir results + + default_qps=1 + default_output_len=1 + benchmark $default_qps $default_output_len + +} + + +main "$@" diff --git a/online_apiserver_test/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh b/online_apiserver_test/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..eb5d891d0d4a58b75f975ae43655ad539bc6a71e --- /dev/null +++ b/online_apiserver_test/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh @@ -0,0 +1,163 @@ +#!/bin/bash + +# Requirement: 2x GPUs. + + +# Model: meta-llama/Meta-Llama-3.1-8B-Instruct +# Query: 1024 input tokens, 6 output tokens, QPS 2/4/6/8, 100 requests +# Resource: 2x GPU +# Approaches: +# 2. Chunked prefill: 2 vllm instance with tp=4, equivalent to 1 tp=4 instance with QPS 4 +# 3. Disaggregated prefill: 1 prefilling instance and 1 decoding instance +# Prefilling instance: max_output_token=1 +# Decoding instance: force the input tokens be the same across requests to bypass prefilling + +set -ex + +kill_gpu_processes() { + # kill all processes on GPU. + pgrep pt_main_thread | xargs -r kill -9 + pgrep python3 | xargs -r kill -9 + for port in 8000 8100 8200; do lsof -t -i:$port | xargs -r kill -9; done + sleep 1 +} + +wait_for_server() { + # wait for vllm server to start + # return 1 if vllm server crashes + local port=$1 + timeout 1200 bash -c " + until curl -s localhost:${port}/v1/completions > /dev/null; do + sleep 1 + done" && return 0 || return 1 +} + + +launch_chunked_prefill() { + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + # disagg prefill + CUDA_VISIBLE_DEVICES=0 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8100 \ + --max-model-len 10000 \ + --enable-chunked-prefill \ + --gpu-memory-utilization 0.6 & + CUDA_VISIBLE_DEVICES=1 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8200 \ + --max-model-len 10000 \ + --enable-chunked-prefill \ + --gpu-memory-utilization 0.6 & + wait_for_server 8100 + wait_for_server 8200 + python3 round_robin_proxy.py & + sleep 1 +} + + +launch_disagg_prefill() { + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + # disagg prefill + CUDA_VISIBLE_DEVICES=0 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8100 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + CUDA_VISIBLE_DEVICES=1 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8200 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + wait_for_server 8100 + wait_for_server 8200 + python3 disagg_prefill_proxy_server.py & + sleep 1 +} + + +benchmark() { + results_folder="./results" + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + dataset_name="sonnet" + dataset_path="../sonnet_4x.txt" + num_prompts=100 + qps=$1 + prefix_len=50 + input_len=1024 + output_len=$2 + tag=$3 + + python3 ../benchmark_serving.py \ + --backend vllm \ + --model $model \ + --dataset-name $dataset_name \ + --dataset-path $dataset_path \ + --sonnet-input-len $input_len \ + --sonnet-output-len "$output_len" \ + --sonnet-prefix-len $prefix_len \ + --num-prompts $num_prompts \ + --port 8000 \ + --save-result \ + --result-dir $results_folder \ + --result-filename "$tag"-qps-"$qps".json \ + --request-rate "$qps" + + sleep 2 +} + + +main() { + + (which wget && which curl) || (apt-get update && apt-get install -y wget curl) + (which jq) || (apt-get -y install jq) + (which socat) || (apt-get -y install socat) + (which lsof) || (apt-get -y install lsof) + + pip install quart httpx matplotlib aiohttp datasets + + cd "$(dirname "$0")" + + cd .. + # create sonnet-4x.txt so that we can sample 2048 tokens for input + echo "" > sonnet_4x.txt + for _ in {1..4} + do + cat sonnet.txt >> sonnet_4x.txt + done + cd disagg_benchmarks + + rm -rf results + mkdir results + + default_output_len=6 + + export VLLM_HOST_IP=$(hostname -I | awk '{print $1}') + + launch_chunked_prefill + for qps in 2 4 6 8; do + benchmark $qps $default_output_len chunked_prefill + done + kill_gpu_processes + + launch_disagg_prefill + for qps in 2 4 6 8; do + benchmark $qps $default_output_len disagg_prefill + done + kill_gpu_processes + + python3 visualize_benchmark_results.py + +} + + +main "$@" diff --git a/online_apiserver_test/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py b/online_apiserver_test/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py new file mode 100644 index 0000000000000000000000000000000000000000..f62d8102e2d9f29c62f177d62453be5d8496cc1e --- /dev/null +++ b/online_apiserver_test/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py @@ -0,0 +1,63 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import os + +import aiohttp +from quart import Quart, make_response, request + +AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60) + +app = Quart(__name__) + + +async def forward_request(url, data): + async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session: + headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"} + async with session.post(url=url, json=data, headers=headers) as response: + if response.status == 200: + # if response.headers.get('Transfer-Encoding') == 'chunked': + if True: + async for chunk_bytes in response.content.iter_chunked(1024): + yield chunk_bytes + else: + content = await response.read() + yield content + + +@app.route("/v1/completions", methods=["POST"]) +async def handle_request(): + try: + original_request_data = await request.get_json() + + prefill_request = original_request_data.copy() + # change max_tokens = 1 to let it only do prefill + prefill_request["max_tokens"] = 1 + + # finish prefill + async for _ in forward_request( + "http://localhost:8100/v1/completions", prefill_request + ): + continue + + # return decode + generator = forward_request( + "http://localhost:8200/v1/completions", original_request_data + ) + response = await make_response(generator) + response.timeout = None + + return response + + except Exception as e: + import sys + import traceback + + exc_info = sys.exc_info() + print("Error occurred in disagg prefill proxy server") + print(e) + print("".join(traceback.format_exception(*exc_info))) + + +if __name__ == "__main__": + app.run(port=8000) diff --git a/online_apiserver_test/benchmarks/disagg_benchmarks/round_robin_proxy.py b/online_apiserver_test/benchmarks/disagg_benchmarks/round_robin_proxy.py new file mode 100644 index 0000000000000000000000000000000000000000..b1df2f255822dad046f5dfcdc1d6538006463510 --- /dev/null +++ b/online_apiserver_test/benchmarks/disagg_benchmarks/round_robin_proxy.py @@ -0,0 +1,63 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import asyncio +import itertools + +import aiohttp +from aiohttp import web + + +class RoundRobinProxy: + def __init__(self, target_ports): + self.target_ports = target_ports + self.port_cycle = itertools.cycle(self.target_ports) + + async def handle_request(self, request): + target_port = next(self.port_cycle) + target_url = f"http://localhost:{target_port}{request.path_qs}" + + async with aiohttp.ClientSession() as session: + try: + # Forward the request + async with session.request( + method=request.method, + url=target_url, + headers=request.headers, + data=request.content, + ) as response: + # Start sending the response + resp = web.StreamResponse( + status=response.status, headers=response.headers + ) + await resp.prepare(request) + + # Stream the response content + async for chunk in response.content.iter_any(): + await resp.write(chunk) + + await resp.write_eof() + return resp + + except Exception as e: + return web.Response(text=f"Error: {str(e)}", status=500) + + +async def main(): + proxy = RoundRobinProxy([8100, 8200]) + app = web.Application() + app.router.add_route("*", "/{path:.*}", proxy.handle_request) + + runner = web.AppRunner(app) + await runner.setup() + site = web.TCPSite(runner, "localhost", 8000) + await site.start() + + print("Proxy server started on http://localhost:8000") + + # Keep the server running + await asyncio.Event().wait() + + +if __name__ == "__main__": + asyncio.run(main()) diff --git a/online_apiserver_test/benchmarks/disagg_benchmarks/visualize_benchmark_results.py b/online_apiserver_test/benchmarks/disagg_benchmarks/visualize_benchmark_results.py new file mode 100644 index 0000000000000000000000000000000000000000..74fa56d076cf14bc066468be24f6053b45166001 --- /dev/null +++ b/online_apiserver_test/benchmarks/disagg_benchmarks/visualize_benchmark_results.py @@ -0,0 +1,47 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import json + +import matplotlib.pyplot as plt +import pandas as pd + +if __name__ == "__main__": + data = [] + for name in ["disagg_prefill", "chunked_prefill"]: + for qps in [2, 4, 6, 8]: + with open(f"results/{name}-qps-{qps}.json") as f: + x = json.load(f) + x["name"] = name + x["qps"] = qps + data.append(x) + + df = pd.DataFrame.from_dict(data) + dis_df = df[df["name"] == "disagg_prefill"] + chu_df = df[df["name"] == "chunked_prefill"] + + plt.style.use("bmh") + plt.rcParams["font.size"] = 20 + + for key in [ + "mean_ttft_ms", + "median_ttft_ms", + "p99_ttft_ms", + "mean_itl_ms", + "median_itl_ms", + "p99_itl_ms", + ]: + fig, ax = plt.subplots(figsize=(11, 7)) + plt.plot( + dis_df["qps"], dis_df[key], label="disagg_prefill", marker="o", linewidth=4 + ) + plt.plot( + chu_df["qps"], chu_df[key], label="chunked_prefill", marker="o", linewidth=4 + ) + ax.legend() + + ax.set_xlabel("QPS") + ax.set_ylabel(key) + ax.set_ylim(bottom=0) + fig.savefig(f"results/{key}.png") + plt.close(fig) diff --git a/online_apiserver_test/benchmarks/fused_kernels/layernorm_rms_benchmarks.py b/online_apiserver_test/benchmarks/fused_kernels/layernorm_rms_benchmarks.py new file mode 100644 index 0000000000000000000000000000000000000000..901524214469e8677cbb58c3078e8b1d724585bb --- /dev/null +++ b/online_apiserver_test/benchmarks/fused_kernels/layernorm_rms_benchmarks.py @@ -0,0 +1,228 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import pickle as pkl +import time +from collections.abc import Iterable +from dataclasses import dataclass +from itertools import product +from typing import Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from tqdm import tqdm + +import vllm._custom_ops as ops +from vllm.model_executor.layers.layernorm import RMSNorm + + +@dataclass +class bench_params_t: + num_tokens: int + hidden_size: int + add_residual: bool + dtype: torch.dtype + + def description(self): + return ( + f"N {self.num_tokens} " + f"x D {self.hidden_size} " + f"x R {self.add_residual} " + f"x DT {self.dtype}" + ) + + +def get_bench_params() -> list[bench_params_t]: + ## Test Fixtures + NUM_TOKENS = [2**x for x in range(11)] + HIDDEN_SIZES = list(range(1024, 8129, 1024)) + ADD_RESIDUAL = [True, False] + DTYPES = [torch.bfloat16, torch.float] + + combinations = product(NUM_TOKENS, HIDDEN_SIZES, ADD_RESIDUAL, DTYPES) + bench_params = list( + map(lambda x: bench_params_t(x[0], x[1], x[2], x[3]), combinations) + ) + return bench_params + + +# Reference impls +def unfused_int8_impl( + rms_norm_layer: RMSNorm, + x: torch.Tensor, + residual: Optional[torch.Tensor], + quant_dtype: torch.dtype, +): + # Norm + torch_out = None + if residual is None: + torch_out = rms_norm_layer.forward_cuda(x, residual) + else: + torch_out, _ = rms_norm_layer.forward_cuda(x, residual) + + # Quant + torch_out, _, _ = ops.scaled_int8_quant(torch_out) + + +def unfused_fp8_impl( + rms_norm_layer: RMSNorm, + x: torch.Tensor, + residual: Optional[torch.Tensor], + quant_dtype: torch.dtype, +): + # Norm + torch_out = None + if residual is None: + torch_out = rms_norm_layer.forward_cuda(x, residual) + else: + torch_out, _ = rms_norm_layer.forward_cuda(x, residual) + + # Quant + torch_out, _ = ops.scaled_fp8_quant(torch_out) + + +def fused_impl( + rms_norm_layer: RMSNorm, # this stores the weights + x: torch.Tensor, + residual: Optional[torch.Tensor], + quant_dtype: torch.dtype, +): + out, _ = ops.rms_norm_dynamic_per_token_quant( + x, rms_norm_layer.weight, 1e-6, quant_dtype, residual=residual + ) + + +# Bench functions +def bench_fn( + rms_norm_layer: RMSNorm, + x: torch.Tensor, + residual: torch.Tensor, + quant_dtype: torch.dtype, + label: str, + sub_label: str, + fn: Callable, + description: str, +) -> TMeasurement: + min_run_time = 1 + + globals = { + "rms_norm_layer": rms_norm_layer, + "x": x, + "residual": residual, + "quant_dtype": quant_dtype, + "fn": fn, + } + return TBenchmark.Timer( + stmt="fn(rms_norm_layer, x, residual, quant_dtype)", + globals=globals, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + +def bench(params: bench_params_t, label: str, sub_label: str) -> Iterable[TMeasurement]: + # Make inputs + layer = RMSNorm(params.hidden_size, 1e-6).to(dtype=params.dtype) + # Make weights + layer.weight.data.normal_(mean=1.0, std=0.1) + # Make inputs + scale = 1 / params.hidden_size + x = ( + torch.randn( + params.num_tokens, params.hidden_size, dtype=params.dtype, device="cuda" + ) + * scale + ) + residual = ( + (torch.randn_like(x) * scale).to(device="cuda") if params.add_residual else None + ) + + timers = [] + + # unfused int8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.int8, + label, + sub_label, + unfused_int8_impl, + "unfused_int8_impl", + ) + ) + + # unfused fp8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.float8_e4m3fn, + label, + sub_label, + unfused_fp8_impl, + "unfused_fp8_impl", + ) + ) + + # fused int8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.int8, + label, + sub_label, + fused_impl, + "fused_int8_impl", + ) + ) + + # fused fp8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.float8_e4m3fn, + label, + sub_label, + fused_impl, + "fused_fp8_impl", + ) + ) + + print_timers(timers) + + return timers + + +# launch bench +# runner +def print_timers(timers: Iterable[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def main(): + torch.set_default_device("cuda") + bench_params = get_bench_params() + + timers = [] + for bp in tqdm(bench_params): + timers.extend(bench(bp, "rms-norm-dynamic-per-token-quant", bp.description())) + print_timers(timers) + + # pickle all the results + timestamp = int(time.time()) + with open(f"rms_norm_dpt_quant-{timestamp}.pkl", "wb") as f: + pkl.dump(timers, f) + + +if __name__ == "__main__": + main() diff --git a/online_apiserver_test/benchmarks/kernels/bench_fp8_gemm.py b/online_apiserver_test/benchmarks/kernels/bench_fp8_gemm.py new file mode 100644 index 0000000000000000000000000000000000000000..920961899038061c13125bdd37f9843b5e4d548a --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/bench_fp8_gemm.py @@ -0,0 +1,159 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import argparse +import copy +import itertools + +import torch +from weight_shapes import WEIGHT_SHAPES + +from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm +from vllm._custom_ops import scaled_fp8_quant as vllm_scaled_fp8_quant +from vllm.triton_utils import triton + +PROVIDER_CFGS = { + "torch-bf16": dict(enabled=True), + "fp8-tensor-w-token-a": dict( + w="tensor", a="token", no_a_quant=False, enabled=False + ), + "fp8-tensor-w-tensor-a": dict( + w="tensor", a="tensor", no_a_quant=False, enabled=True + ), + "fp8-channel-w-token-a": dict( + w="channel", a="token", no_a_quant=False, enabled=True + ), + "fp8-channel-w-tensor-a": dict( + w="channel", a="tensor", no_a_quant=False, enabled=False + ), + "fp8-tensor-w-token-a-noquant": dict( + w="tensor", a="token", no_a_quant=True, enabled=False + ), + "fp8-tensor-w-tensor-a-noquant": dict( + w="tensor", a="tensor", no_a_quant=True, enabled=True + ), + "fp8-channel-w-token-a-noquant": dict( + w="channel", a="token", no_a_quant=True, enabled=True + ), + "fp8-channel-w-tensor-a-noquant": dict( + w="channel", a="tensor", no_a_quant=True, enabled=False + ), +} + +_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]] + + +def _quant_weight_fp8(b: torch.Tensor, w_type: str, device: str): + if w_type == "tensor": + scale_b = torch.ones(1, device=device, dtype=torch.float32) + b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b) + else: + b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, use_per_token_if_dynamic=True) + return b_fp8.t(), scale_b_fp8 + + +def build_fp8_runner(cfg, a, b, dtype, device): + b_fp8, scale_b_fp8 = _quant_weight_fp8(b, cfg["w"], device) + + scale_a_const = ( + torch.ones(1, device=device, dtype=torch.float32) + if cfg["a"] == "tensor" + else None + ) + + if cfg["no_a_quant"]: + if cfg["a"] == "tensor": + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const) + else: + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True) + + def run(): + return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype) + + return run + + if cfg["a"] == "tensor": + + def run(): + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const) + return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype) + + else: + + def run(): + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True) + return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype) + + return run + + +@triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["batch_size"], + x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384], + x_log=False, + line_arg="provider", + line_vals=_enabled, + line_names=_enabled, + ylabel="TFLOP/s (larger is better)", + plot_name="BF16 vs FP8 GEMMs", + args={}, + ) +) +def benchmark(batch_size, provider, N, K): + M = batch_size + device = "cuda" + dtype = torch.bfloat16 + + a = torch.randn((M, K), device=device, dtype=dtype) + b = torch.randn((N, K), device=device, dtype=dtype) + + quantiles = [0.5, 0.2, 0.8] + + if provider == "torch-bf16": + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: torch.nn.functional.linear(a, b), quantiles=quantiles + ) + else: + cfg = PROVIDER_CFGS[provider] + run_quant = build_fp8_runner(cfg, a, b, dtype, device) + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: run_quant(), quantiles=quantiles + ) + + to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3) + return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms) + + +def prepare_shapes(args): + out = [] + for model, tp_size in itertools.product(args.models, args.tp_sizes): + for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]): + KN[tp_dim] //= tp_size + KN.append(model) + out.append(KN) + return out + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--models", + nargs="+", + type=str, + default=["meta-llama/Llama-3.1-8B-Instruct"], + choices=list(WEIGHT_SHAPES.keys()), + ) + parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1]) + args = parser.parse_args() + + for K, N, model in prepare_shapes(args): + print(f"{model}, N={N} K={K}, BF16 vs FP8 GEMMs TFLOP/s:") + benchmark.run( + print_data=True, + show_plots=True, + save_path=f"bench_fp8_res_n{N}_k{K}", + N=N, + K=K, + ) + + print("Benchmark finished!") diff --git a/online_apiserver_test/benchmarks/kernels/bench_int8_gemm.py b/online_apiserver_test/benchmarks/kernels/bench_int8_gemm.py new file mode 100644 index 0000000000000000000000000000000000000000..e9c6d64404d0dc6958aeb675bf6b893623649ffa --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/bench_int8_gemm.py @@ -0,0 +1,169 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import argparse +import copy +import itertools + +import torch +from weight_shapes import WEIGHT_SHAPES + +from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm +from vllm._custom_ops import scaled_int8_quant as vllm_scaled_int8_quant +from vllm.triton_utils import triton + +PROVIDER_CFGS = { + "torch-bf16": dict(enabled=True), + "int8-tensor-w-token-a": dict( + w="tensor", a="token", no_a_quant=False, enabled=False + ), + "int8-tensor-w-tensor-a": dict( + w="tensor", a="tensor", no_a_quant=False, enabled=True + ), + "int8-channel-w-token-a": dict( + w="channel", a="token", no_a_quant=False, enabled=True + ), + "int8-channel-w-tensor-a": dict( + w="channel", a="tensor", no_a_quant=False, enabled=False + ), + "int8-tensor-w-token-a-noquant": dict( + w="tensor", a="token", no_a_quant=True, enabled=False + ), + "int8-tensor-w-tensor-a-noquant": dict( + w="tensor", a="tensor", no_a_quant=True, enabled=True + ), + "int8-channel-w-token-a-noquant": dict( + w="channel", a="token", no_a_quant=True, enabled=True + ), + "int8-channel-w-tensor-a-noquant": dict( + w="channel", a="tensor", no_a_quant=True, enabled=False + ), +} + + +def _quant_weight(b, w_type, device): + if w_type == "tensor": + scale_b = torch.ones(1, device=device, dtype=torch.float32) + b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b, scale_b) + assert scale_b_int8.numel() == 1 + else: # channel + b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b) + assert scale_b_int8.numel() == b.shape[0] + return b_int8.t(), scale_b_int8 + + +def build_int8_runner(cfg, a, b, dtype, device): + # quant before running the kernel + b_int8, scale_b_int8 = _quant_weight(b, cfg["w"], device) + + scale_a_const = None + if cfg["a"] == "tensor": + scale_a_const = torch.ones(1, device=device, dtype=torch.float32) + + # no quant, create activation ahead + if cfg["no_a_quant"]: + if cfg["a"] == "tensor": + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const) + else: # token + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a) + + def run_quant(): + return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype) + + return run_quant + + # dynamic quant, create activation inside + if cfg["a"] == "tensor": + + def run_quant(): + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const) + return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype) + + else: # token + + def run_quant(): + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a) + return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype) + + return run_quant + + +_enabled = [k for k, v in PROVIDER_CFGS.items() if v.get("enabled")] + + +@triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["batch_size"], + x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384], + x_log=False, + line_arg="provider", + line_vals=_enabled, + line_names=[k for k in _enabled], + ylabel="TFLOP/s (larger is better)", + plot_name="BF16 vs INT8 GEMMs", + args={}, + ) +) +def benchmark(batch_size, provider, N, K): + M = batch_size + device = "cuda" + dtype = torch.bfloat16 + a = torch.randn((M, K), device=device, dtype=dtype) + b = torch.randn((N, K), device=device, dtype=dtype) + + quantiles = [0.5, 0.2, 0.8] + + if provider == "torch-bf16": + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: torch.nn.functional.linear(a, b), quantiles=quantiles + ) + else: + cfg = PROVIDER_CFGS[provider] + run_quant = build_int8_runner(cfg, a, b, dtype, device) + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: run_quant(), quantiles=quantiles + ) + + to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3) + return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms) + + +def prepare_shapes(args): + KN_model_names = [] + for model, tp_size in itertools.product(args.models, args.tp_sizes): + for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]): + KN[tp_dim] //= tp_size + KN.append(model) + KN_model_names.append(KN) + return KN_model_names + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--models", + nargs="+", + type=str, + default=["meta-llama/Llama-3.1-8B-Instruct"], + choices=list(WEIGHT_SHAPES.keys()), + help="List of models to benchmark", + ) + parser.add_argument( + "--tp-sizes", + nargs="+", + type=int, + default=[1], + help="List of tensor parallel sizes", + ) + args = parser.parse_args() + + for K, N, model in prepare_shapes(args): + print(f"{model}, N={N} K={K}, BF16 vs INT8 GEMMs TFLOP/s:") + benchmark.run( + print_data=True, + show_plots=True, + save_path=f"bench_int8_res_n{N}_k{K}", + N=N, + K=K, + ) + + print("Benchmark finished!") diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_aqlm.py b/online_apiserver_test/benchmarks/kernels/benchmark_aqlm.py new file mode 100644 index 0000000000000000000000000000000000000000..42de062b08e424619b010f9170725b6785a35f58 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_aqlm.py @@ -0,0 +1,345 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import os +import sys +from typing import Optional + +import torch +import torch.nn.functional as F + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.aqlm import ( + dequantize_weight, + generic_dequantize_gemm, + get_int_dtype, + optimized_dequantize_gemm, +) +from vllm.utils import FlexibleArgumentParser + +os.environ["CUDA_VISIBLE_DEVICES"] = "0" + + +def torch_mult( + # [..., in_features] + input: torch.Tensor, + weights: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, +) -> torch.Tensor: + output = F.linear(input, weights) + return output + + +def dequant_out_scale( + # [..., in_features] + input: torch.Tensor, + # [num_out_groups, num_in_groups, num_codebooks] + codes: torch.IntTensor, + # [num_codebooks, codebook_size, out_group_size, in_group_size] + codebooks: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, + output_partition_sizes: torch.IntTensor, + bias: Optional[torch.Tensor], +) -> torch.Tensor: + weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes) + + if bias is None: + output = F.linear(input, weights, bias) + orig_shape = output.shape + flattened_output = output.view(-1, output.size(-1)) + f_scales = scales.view(-1, scales.shape[0]) + b_scales = f_scales.expand(flattened_output.shape[0], -1) + flattened_output *= b_scales + return flattened_output.view(orig_shape) + else: + b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1]) + weights *= b_scales + return F.linear(input, weights, bias) + + +def dequant_weight_scale( + # [..., in_features] + input: torch.Tensor, + # [num_out_groups, num_in_groups, num_codebooks] + codes: torch.IntTensor, + # [num_codebooks, codebook_size, out_group_size, in_group_size] + codebooks: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, + output_partition_sizes: torch.IntTensor, + bias: Optional[torch.Tensor], +) -> torch.Tensor: + weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes) + + b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1]) + weights *= b_scales + return F.linear(input, weights, bias) + + +def dequant_no_scale( + # [..., in_features] + input: torch.Tensor, + # [num_out_groups, num_in_groups, num_codebooks] + codes: torch.IntTensor, + # [num_codebooks, codebook_size, out_group_size, in_group_size] + codebooks: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, + output_partition_sizes: torch.IntTensor, + bias: Optional[torch.Tensor], +) -> torch.Tensor: + weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes) + + return F.linear(input, weights, bias) + + +# Compare the optimized 1x16 and 2x8 cuda decompression/dequant kernels against +# the generic pytorch version. +# Just visual comparison. +def dequant_test(k: int, parts: torch.Tensor, nbooks: int, bits: int) -> None: + n = int(parts.sum().item()) + + device = torch.device("cuda:0") + + code_range = (1 << bits) // 2 + ingroups = 8 + + codes = torch.randint( + -code_range, + code_range, + size=(n, k // ingroups, nbooks), + dtype=get_int_dtype(bits), + device=device, + ) + + codebooks = torch.randn( + size=(parts.shape[0] * nbooks, 1 << bits, 1, 8), + dtype=torch.float16, + device=device, + ) + + count = 0 + for index in range(16): + for i in range(8): + for book in range(nbooks): + codebooks[book, index, 0, i] = count * (10**book) + count += 1 + + print("codes shape", codes.shape) + + for i in range(16): + for book in range(nbooks): + codes[0, i, book] = i + codes[0, -i, book] = i + + weights = dequantize_weight(codes, codebooks, None) + weights2 = ops.aqlm_dequant(codes, codebooks, parts) + + print("weights shape:", weights.shape) + print("weights2 shape:", weights2.shape) + + print("weights are:", weights) + print("weights2 are:", weights2) + + print("first 128 weights are", weights[0, 0:128].to(torch.int32)) + print("first 128 weights2 are:", weights2[0, 0:128].to(torch.int32)) + + print("last 128 weights are", weights[0, -128:]) + print("last 128 weights2 are:", weights2[0, -128:]) + + +def main(): + parser = FlexibleArgumentParser(description="Benchmark aqlm performance.") + + # Add arguments + parser.add_argument( + "--nbooks", type=int, default=1, help="Number of codebooks (default: 1)" + ) + parser.add_argument( + "--bits", + type=int, + default=16, + help="Number of bits per code element (default: 16)", + ) + parser.add_argument( + "--test", + type=bool, + default=False, + help="Run the decompression/dequant tester rather than benchmarking " + "(default: False)", + ) + + # Parse the arguments + args = parser.parse_args() + + # Extract values + nbooks = args.nbooks + bits = args.bits + + if args.test: + dequant_test(4096, torch.tensor((4096,)), nbooks, bits) + return + + # Otherwise, benchmark. + methods = [ + ops.aqlm_gemm, + dequant_out_scale, + generic_dequantize_gemm, + optimized_dequantize_gemm, + dequant_weight_scale, + torch_mult, + dequant_no_scale, + ] + + filename = f"./aqlm_benchmark_{nbooks}x{bits}.csv" + print(f"writing benchmarks to file {filename}") + with open(filename, "w") as f: + sys.stdout = f + + print("m | k | n | n parts", end="") + for method in methods: + print(f" | {method.__name__.replace('_', ' ')} (µs)", end="") + print("") + + # These are reasonable prefill sizes. + ksandpartions = ( + (4096, (4096, 4096, 4096)), + (4096, (4096,)), + (4096, (11008, 11008)), + (11008, (4096,)), + ) + + # reasonable ranges for m. + for m in [ + 1, + 2, + 4, + 8, + 10, + 12, + 14, + 16, + 24, + 32, + 48, + 52, + 56, + 64, + 96, + 112, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ]: + print(f"{m}", file=sys.__stdout__) + for ksp in ksandpartions: + run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits, methods) + + sys.stdout = sys.__stdout__ + + +def run_grid(m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, methods): + # I didn't see visible improvements from increasing these, but feel free :) + num_warmup_trials = 1 + num_trials = 1 + + num_calls = 100 + + # warmup. + for method in methods: + for _ in range(num_warmup_trials): + run_timing( + num_calls=num_calls, + m=m, + k=k, + parts=parts, + nbooks=nbooks, + bits=bits, + method=method, + ) + + n = parts.sum().item() + print(f"{m} | {k} | {n} | {parts.tolist()}", end="") + + for method in methods: + best_time_us = 1e20 + for _ in range(num_trials): + kernel_dur_ms = run_timing( + num_calls=num_calls, + m=m, + k=k, + parts=parts, + nbooks=nbooks, + bits=bits, + method=method, + ) + + kernel_dur_us = 1000 * kernel_dur_ms + + if kernel_dur_us < best_time_us: + best_time_us = kernel_dur_us + + print(f" | {kernel_dur_us:.0f}", end="") + + print("") + + +def run_timing( + num_calls: int, m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, method +) -> float: + n = int(parts.sum().item()) + + device = torch.device("cuda:0") + + input = torch.randn((1, m, k), dtype=torch.float16, device=device) + + code_range = (1 << bits) // 2 + ingroups = 8 + + codes = torch.randint( + -code_range, + code_range, + size=(n, k // ingroups, nbooks), + dtype=get_int_dtype(bits), + device=device, + ) + + codebooks = torch.randn( + size=(parts.shape[0] * nbooks, 1 << bits, 1, 8), + dtype=torch.float16, + device=device, + ) + + scales = torch.randn(size=(n, 1, 1, 1), dtype=torch.float16, device=device) + + # for comparison to just a pytorch mult. + weights = torch.randn((n, k), dtype=torch.float16, device=device) + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + start_event.record() + + if method is torch_mult: + for i in range(num_calls): + torch_mult(input, weights, scales) + else: + for i in range(num_calls): + method(input, codes, codebooks, scales, parts, None) + + end_event.record() + end_event.synchronize() + + dur_ms = start_event.elapsed_time(end_event) / num_calls + return dur_ms + + +if __name__ == "__main__": + sys.exit(main()) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_bitblas.py b/online_apiserver_test/benchmarks/kernels/benchmark_bitblas.py new file mode 100644 index 0000000000000000000000000000000000000000..97ee060341373022b06c6ce62600d5c23c771c37 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_bitblas.py @@ -0,0 +1,242 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +# Copyright (c) Microsoft Corporation. +# Licensed under the MIT License. + +from vllm.model_executor.layers.quantization.utils.bitblas_utils import ( + MINIMUM_BITBLAS_VERSION, +) + +try: + import bitblas + + if bitblas.__version__ < MINIMUM_BITBLAS_VERSION: + raise ImportError( + "bitblas version is wrong. Please " + f"install bitblas>={MINIMUM_BITBLAS_VERSION}" + ) +except ImportError as e: + bitblas_import_exception = e + raise ValueError( + "Trying to use the bitblas backend, but could not import" + f"with the following error: {bitblas_import_exception}. " + "Please install bitblas through the following command: " + f"`pip install bitblas>={MINIMUM_BITBLAS_VERSION}`" + ) from bitblas_import_exception + +from bitblas import Matmul, MatmulConfig, auto_detect_nvidia_target + +from vllm.utils import FlexibleArgumentParser + +parser = FlexibleArgumentParser( + description="Benchmark BitBLAS int4 on a specific target." +) + +# Add arguments to the parser +parser.add_argument( + "--target", + type=str, + default=auto_detect_nvidia_target(), + help="Specify the target device for benchmarking.", +) +parser.add_argument( + "--group_size", type=int, default=None, help="Group size for grouped quantization." +) +parser.add_argument( + "--A_dtype", + type=str, + default="float16", + choices=["float16", "float32", "float64", "int32", "int8"], + help="Data type of activation A.", +) +parser.add_argument( + "--W_dtype", + type=str, + default="int4", + choices=[ + "float16", + "float32", + "float64", + "int32", + "int8", + "int4", + "int2", + "int1", + "nf4", + "fp4_e2m1", + ], + help="Data type of weight W.", +) +parser.add_argument( + "--accum_dtype", + type=str, + default="float16", + choices=["float16", "int32"], + help="Data type for accumulation.", +) +parser.add_argument( + "--out_dtype", + type=str, + default="float16", + choices=["float16", "float32", "int32", "int8"], + help="Data type for output.", +) +parser.add_argument( + "--layout", + type=str, + default="nt", + choices=["nt", "nn"], + help="Matrix layout, 'nt' for non-transpose A and transpose W.", +) +parser.add_argument( + "--with_bias", action="store_true", help="Include bias in the benchmark." +) +parser.add_argument( + "--with_scaling", + action="store_true", + help="Include scaling factor in the quantization.", +) +parser.add_argument( + "--with_zeros", action="store_true", help="Include zeros in the quantization." +) +parser.add_argument( + "--zeros_mode", + type=str, + default=None, + choices=["original", "rescale", "quantized"], + help="Specify the mode for calculating zeros.", +) + +# Parse the arguments +args = parser.parse_args() + +# Assign arguments to variables +target = args.target +A_dtype = args.A_dtype +W_dtype = args.W_dtype +accum_dtype = args.accum_dtype +out_dtype = args.out_dtype +layout = args.layout +with_bias = args.with_bias +group_size = args.group_size +with_scaling = args.with_scaling +with_zeros = args.with_zeros +zeros_mode = args.zeros_mode + +# Define a list of shared arguments that repeat in every config +shared_args = [ + A_dtype, + W_dtype, + out_dtype, + accum_dtype, + layout, + with_bias, + group_size, + with_scaling, + with_zeros, + zeros_mode, +] + +# Define just the (M, K, N) shapes in a more compact list +shapes = [ + # square test + (1, 16384, 16384), + # BLOOM-176B + (1, 43008, 14336), + (1, 14336, 14336), + (1, 57344, 14336), + (1, 14336, 57344), + # OPT-65B + (1, 9216, 9216), + (1, 36864, 9216), + (1, 9216, 36864), + (1, 22016, 8192), + # LLAMA-70B/65B + (1, 8192, 22016), + (1, 8192, 8192), + (1, 28672, 8192), + (1, 8192, 28672), + # square test + (16384, 16384, 16384), + # BLOOM-176B + (8192, 43008, 14336), + (8192, 14336, 14336), + (8192, 57344, 14336), + (8192, 14336, 57344), + # OPT-65B + (8192, 9216, 9216), + (8192, 36864, 9216), + (8192, 9216, 36864), + (8192, 22016, 8192), + # LLAMA-70B/65B + (8192, 8192, 22016), + (8192, 8192, 8192), + (8192, 28672, 8192), + (8192, 8192, 28672), +] + +# Build test shapes with all the shared arguments +test_shapes = [(MatmulConfig, Matmul, (*shape, *shared_args)) for shape in shapes] + +benchmark_sets = [] +benchmark_sets.extend(test_shapes) + +benchmark_results = {} +for config_class, operator, input_args in benchmark_sets: + config = config_class(*input_args) + matmul = operator(config, target=target, enable_tuning=True) + kernel_latency = matmul.profile_latency() + + print("Time cost is: {:.3f} ms".format(kernel_latency)) + + profile_config = { + f"{operator.__name__}-{'-'.join([str(i) for i in input_args])}": { + "BitBLAS_top20_latency": kernel_latency, + } + } + + benchmark_results.update(profile_config) + +# Define headers for the table +headers = [ + "PrimFunc", + "Input Arguments", + "BitBLAS Top20 Latency", +] + +# Calculate column widths for pretty printing +col_widths = [0, 0, 0] +for config_key, values in benchmark_results.items(): + args_split = config_key.split("-") + func_name = args_split[0] + input_args_str = "-".join(args_split[1:]) + col_widths[0] = max(col_widths[0], len(func_name) + 2, len(headers[0]) + 2) + col_widths[1] = max(col_widths[1], len(input_args_str) + 2, len(headers[1]) + 2) + col_widths[2] = max( + col_widths[2], + len(f"{values['BitBLAS_top20_latency']:.3f} ms") + 2, + len(headers[2]) + 2, + ) + # break only if you want to measure widths from a single example; + # otherwise, let it loop over all items. + +# Print header +for i, header in enumerate(headers): + headers[i] = header.ljust(col_widths[i]) +print("".join(headers)) +print("-" * sum(col_widths)) + +# Print rows +for config_key, values in benchmark_results.items(): + args_split = config_key.split("-") + func_name = args_split[0] + input_args_str = "-".join(args_split[1:]) + row = [ + func_name, + input_args_str, + f"{values['BitBLAS_top20_latency']:.3f} ms", + ] + row_str = "".join( + [str(cell).ljust(col_widths[idx]) for idx, cell in enumerate(row)] + ) + print(row_str) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_cutlass_fp4_moe.py b/online_apiserver_test/benchmarks/kernels/benchmark_cutlass_fp4_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..35c20ee41b9a94c10f1367a45fe391cc15e9ffec --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_cutlass_fp4_moe.py @@ -0,0 +1,490 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +Benchmark the performance of the cutlass_moe_fp4 kernel vs the triton_moe +kernel. The cutlass_moe_fp4 kernel takes in fp4 quantized weights and 16-bit +activations. The triton_moe kernel takes in fp8 weights(tensor scaled to fp8) +and 16-bit activations. +""" + +import nvtx +import torch +import torch.utils.benchmark as benchmark + +from vllm import _custom_ops as ops +from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config +from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4 +from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk +from vllm.scalar_type import scalar_types +from vllm.utils import FlexibleArgumentParser + +WEIGHT_SHAPES_MOE = { + "nvidia/DeepSeek-R1-FP4": [ + [256, 8, 2048, 7168], + ], +} + +DEFAULT_MODELS = [ + "nvidia/DeepSeek-R1-FP4", +] + +DEFAULT_BATCH_SIZES = [4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048] +DEFAULT_TP_SIZES = [1] + +PER_ACT_TOKEN_OPTS = [False] +PER_OUT_CH_OPTS = [False] +FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max() +FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max + + +def to_fp8(tensor: torch.Tensor): + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to( + dtype=torch.float8_e4m3fn + ) + + +def bench_run( + results: list[benchmark.Measurement], + model: str, + num_experts: int, + topk: int, + per_act_token: bool, + per_out_ch: bool, + mkn: tuple[int, int, int], +): + label = "NVFP4 Blockscaled CUTLASS MOE vs FP8 Tensor Scaled Triton" + + sub_label = ( + "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, MKN=({})".format( + model, num_experts, topk, per_act_token, per_out_ch, mkn + ) + ) + + print(f"Testing: {sub_label}") + + (m, k, n) = mkn + + dtype = torch.half + device = "cuda" + a = torch.randn((m, k), device=device, dtype=dtype) / 10 + w1 = torch.randn((num_experts, 2 * n, k), device=device, dtype=dtype) / 10 + w2 = torch.randn((num_experts, k, n), device=device, dtype=dtype) / 10 + + _, a_fp8_scale = ops.scaled_fp8_quant(a) + + w1_fp8q = torch.empty( + (num_experts, 2 * n, k), device=device, dtype=torch.float8_e4m3fn + ) + w2_fp8q = torch.empty((num_experts, k, n), device=device, dtype=torch.float8_e4m3fn) + w1_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32) + w2_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32) + + for expert in range(num_experts): + w1_fp8q[expert], w1_fp8scale[expert] = ops.scaled_fp8_quant(w1[expert]) + w2_fp8q[expert], w2_fp8scale[expert] = ops.scaled_fp8_quant(w2[expert]) + + w1_fp8q_notransp = w1_fp8q.clone() + w2_fp8q_notransp = w2_fp8q.clone() + w1_fp8q = w1_fp8q.transpose(1, 2) + w2_fp8q = w2_fp8q.transpose(1, 2) + + score = torch.randn((m, num_experts), device=device, dtype=dtype) + + topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False) + + quant_blocksize = 16 + w1_blockscale = torch.empty( + (num_experts, 2 * n, k // quant_blocksize), + device=device, + dtype=torch.float8_e4m3fn, + ) + w2_blockscale = torch.empty( + (num_experts, k, n // quant_blocksize), device=device, dtype=torch.float8_e4m3fn + ) + + # n_b_scales = 2 * n if per_out_ch else 1 + # k_b_scales = k if per_out_ch else 1 + w1_fp4 = torch.empty((num_experts, 2 * n, k // 2), device=device, dtype=torch.uint8) + w2_fp4 = torch.empty((num_experts, k, n // 2), device=device, dtype=torch.uint8) + + w1_gs = torch.empty((num_experts,), device=device, dtype=torch.float32) + w2_gs = torch.empty((num_experts,), device=device, dtype=torch.float32) + a1_gs = torch.ones((num_experts,), device=device, dtype=torch.float32) + a2_gs = torch.ones((num_experts,), device=device, dtype=torch.float32) + + for expert in range(num_experts): + w1_e = w1[expert] + w2_e = w2[expert] + w1_amax = torch.abs(w1_e).max().to(torch.float32) + w2_amax = torch.abs(w2_e).max().to(torch.float32) + w1_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w1_amax + w2_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w2_amax + + w1_fp4[expert], w1_blockscale[expert] = ops.scaled_fp4_quant( + w1_e, w1_gs[expert] + ) + + w2_fp4[expert], w2_blockscale[expert] = ops.scaled_fp4_quant( + w2_e, w2_gs[expert] + ) + + def run_triton_moe( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_fp8_scale: torch.Tensor, + num_repeats: int, + ): + for _ in range(num_repeats): + fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_fp8_scale, + ) + + def run_cutlass_moe_fp4( + a: torch.Tensor, + w1_fp4: torch.Tensor, + w2_fp4: torch.Tensor, + w1_blockscale: torch.Tensor, + w2_blockscale: torch.Tensor, + w1_gs: torch.Tensor, + w2_gs: torch.Tensor, + a1_gs: torch.Tensor, + a2_gs: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + m: int, + n: int, + k: int, + e: int, + device: torch.device, + num_repeats: int, + ): + for _ in range(num_repeats): + with nvtx.annotate("cutlass_moe_fp4", color="green"): + cutlass_moe_fp4( + a=a, + a1_gscale=a1_gs, + a2_gscale=a2_gs, + w1_fp4=w1_fp4, + w1_blockscale=w1_blockscale, + w1_alphas=w1_gs, + w2_fp4=w2_fp4, + w2_blockscale=w2_blockscale, + w2_alphas=w2_gs, + topk_weights=topk_weights, + topk_ids=topk_ids, + m=m, + n=n, + k=k, + e=num_experts, + device=device, + ) + + def run_cutlass_from_graph( + a: torch.Tensor, + a1_gscale: torch.Tensor, + w1_fp4: torch.Tensor, + w1_blockscale: torch.Tensor, + w1_alphas: torch.Tensor, + a2_gscale: torch.Tensor, + w2_fp4: torch.Tensor, + w2_blockscale: torch.Tensor, + w2_alphas: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + m: int, + n: int, + k: int, + e: int, + device: torch.device, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return cutlass_moe_fp4( + a=a, + a1_gscale=a1_gs, + w1_fp4=w1_fp4, + w1_blockscale=w1_blockscale, + w1_alphas=w1_alphas, + a2_gscale=a2_gs, + w2_fp4=w2_fp4, + w2_blockscale=w2_blockscale, + w2_alphas=w2_alphas, + topk_weights=topk_weights, + topk_ids=topk_ids, + m=m, + n=n, + k=k, + e=num_experts, + device=device, + ) + + def run_triton_from_graph( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_fp8_scale: torch.Tensor, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_fp8_scale, + ) + + def replay_graph(graph, num_repeats): + for _ in range(num_repeats): + graph.replay() + torch.cuda.synchronize() + + cutlass_stream = torch.cuda.Stream() + cutlass_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(cutlass_graph, stream=cutlass_stream): + run_cutlass_from_graph( + a=a, + a1_gscale=a1_gs, + w1_fp4=w1_fp4, + w1_blockscale=w1_blockscale, + w1_alphas=w1_gs, + a2_gscale=a2_gs, + w2_fp4=w2_fp4, + w2_blockscale=w2_blockscale, + w2_alphas=w2_gs, + topk_weights=topk_weights, + topk_ids=topk_ids, + m=m, + n=n, + k=k, + e=num_experts, + device=device, + ) + torch.cuda.synchronize() + + triton_stream = torch.cuda.Stream() + triton_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(triton_graph, stream=triton_stream): + run_triton_from_graph( + a, + w1_fp8q_notransp, + w2_fp8q_notransp, + topk_weights, + topk_ids, + w1_fp8scale, + w2_fp8scale, + a_fp8_scale, + ) + torch.cuda.synchronize() + + min_run_time = 5 + num_warmup = 5 + num_runs = 25 + + globals = { + # Baseline params + "w1": w1, + "w2": w2, + "score": score, + "topk": topk, + "w1_fp8q_notransp": w1_fp8q_notransp, + "w2_fp8q_notransp": w2_fp8q_notransp, + "w1_fp8scale": w1_fp8scale, + "w2_fp8scale": w2_fp8scale, + "a_fp8_scale": a_fp8_scale, + # Cutlass params + "a": a, + "a1_gscale": a1_gs, + "w1_fp4": w1_fp4, + "w1_blockscale": w1_blockscale, + "w1_alphas": w1_gs, + "a2_gscale": a2_gs, + "w2_fp4": w2_fp4, + "w2_blockscale": w2_blockscale, + "w2_alphas": w2_gs, + "topk_weights": topk_weights, + "topk_ids": topk_ids, + "m": m, + "n": n, + "k": k, + "e": num_experts, + "device": device, + # cuda graph params + "cutlass_graph": cutlass_graph, + "triton_graph": triton_graph, + # Gen params + "num_runs": num_runs, + # Kernels + "run_triton_moe": run_triton_moe, + "run_cutlass_moe_fp4": run_cutlass_moe_fp4, + "replay_graph": replay_graph, + } + + # Warmup + run_triton_moe( + a, + w1_fp8q_notransp, + w2_fp8q_notransp, + topk_weights, + topk_ids, + w1_fp8scale, + w2_fp8scale, + a_fp8_scale, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_triton_moe(a, w1_fp8q_notransp, w2_fp8q_notransp, topk_weights, topk_ids, w1_fp8scale, w2_fp8scale, a_fp8_scale, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(triton_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(triton_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + + run_cutlass_moe_fp4( + a, + w1_fp4, + w2_fp4, + w1_blockscale, + w2_blockscale, + w1_gs, + w2_gs, + a1_gs, + a2_gs, + topk_weights, + topk_ids, + m, + n, + k, + num_experts, + device, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_cutlass_moe_fp4(a, w1_fp4, w2_fp4, w1_blockscale, w2_blockscale, w1_alphas, w2_alphas, a1_gscale, a2_gscale, topk_weights, topk_ids, m, n, k, e, device, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="cutlass_moe_fp4", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(cutlass_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(cutlass_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="cutlass_moe_fp4_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + +def main(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + results: list[benchmark.Measurement] = [] + + for model in args.models: + for tp in args.tp_sizes: + for layer in WEIGHT_SHAPES_MOE[model]: + num_experts = layer[0] + topk = layer[1] + size_k = layer[2] + size_n = layer[3] // tp + + if len(args.limit_k) > 0 and size_k not in args.limit_k: + continue + + if len(args.limit_n) > 0 and size_n not in args.limit_n: + continue + + for per_act_token in PER_ACT_TOKEN_OPTS: + for per_out_ch in PER_OUT_CH_OPTS: + for size_m in args.batch_sizes: + mkn = (size_m, size_k, size_n) + bench_run( + results, + model, + num_experts, + topk, + per_act_token, + per_out_ch, + mkn, + ) + + compare = benchmark.Compare(results) + compare.print() + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark NVFP4 CUTLASS MOE across specified models/shapes/batches" + ) + parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES_MOE.keys(), + ) + parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES) + parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + parser.add_argument("--limit-k", nargs="+", type=int, default=[]) + parser.add_argument("--limit-n", nargs="+", type=int, default=[]) + parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-act-token", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[]) + + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py b/online_apiserver_test/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py new file mode 100644 index 0000000000000000000000000000000000000000..1d4e730f99ae911535caa4cca17122656db06a4b --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py @@ -0,0 +1,383 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import torch +import torch.utils.benchmark as benchmark +from benchmark_shapes import WEIGHT_SHAPES_MOE + +from vllm import _custom_ops as ops +from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config +from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8 +from vllm.model_executor.layers.fused_moe.fused_moe import ( + fused_experts, + fused_topk, +) +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = [ + "nm-testing/Mixtral-8x7B-Instruct-v0.1", + "nm-testing/deepseekv2-lite", + "ibm-granite/granite-3.0-1b-a400m", + "ibm-granite/granite-3.0-3b-a800m", +] +DEFAULT_BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512] +DEFAULT_TP_SIZES = [1] + +PER_ACT_TOKEN_OPTS = [False] +PER_OUT_CH_OPTS = [False] + + +def to_fp8(tensor: torch.Tensor): + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to( + dtype=torch.float8_e4m3fn + ) + + +def bench_run( + results: list[benchmark.Measurement], + model: str, + num_experts: int, + topk: int, + per_act_token: bool, + per_out_ch: bool, + mkn: tuple[int, int, int], +): + label = "Quant Matmul" + + sub_label = ( + "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, MKN=({})".format( + model, num_experts, topk, per_act_token, per_out_ch, mkn + ) + ) + + print(f"Testing: {sub_label}") + + (m, k, n) = mkn + + dtype = torch.half + + a = torch.randn((m, k), device="cuda", dtype=dtype) / 10 + w1 = torch.randn((num_experts, 2 * n, k), device="cuda", dtype=dtype) / 10 + w2 = torch.randn((num_experts, k, n), device="cuda", dtype=dtype) / 10 + + _, a_scale = ops.scaled_fp8_quant(a) + + w1_q = torch.empty( + (num_experts, 2 * n, k), device="cuda", dtype=torch.float8_e4m3fn + ) + w2_q = torch.empty((num_experts, k, n), device="cuda", dtype=torch.float8_e4m3fn) + w1_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32) + w2_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32) + + for expert in range(num_experts): + w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(w1[expert]) + w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(w2[expert]) + + score = torch.randn((m, num_experts), device="cuda", dtype=dtype) + + topk_weights, topk_ids, token_expert_indices = fused_topk( + a, score, topk, renormalize=False + ) + + def run_triton_moe( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_scale: torch.Tensor, + num_repeats: int, + ): + for _ in range(num_repeats): + fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_scale, + ) + + def run_cutlass_moe( + a: torch.Tensor, + a_scale: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + per_act_token: bool, + num_repeats: int, + ): + for _ in range(num_repeats): + cutlass_moe_fp8( + a, + w1, + w2, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + per_act_token, + a1_scale=None, + ) + + def run_cutlass_from_graph( + a: torch.Tensor, + a_scale: torch.Tensor, + w1_q: torch.Tensor, + w2_q: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return cutlass_moe_fp8( + a, + w1_q, + w2_q, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + per_act_token, + a1_scale=None, + ) + + def run_triton_from_graph( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_scale: torch.Tensor, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_scale, + ) + + def replay_graph(graph, num_repeats): + for _ in range(num_repeats): + graph.replay() + torch.cuda.synchronize() + + cutlass_stream = torch.cuda.Stream() + cutlass_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(cutlass_graph, stream=cutlass_stream): + run_cutlass_from_graph( + a, + a_scale, + w1_q, + w2_q, + w1_scale, + w2_scale, + topk_weights, + topk_ids, + ) + torch.cuda.synchronize() + + triton_stream = torch.cuda.Stream() + triton_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(triton_graph, stream=triton_stream): + run_triton_from_graph( + a, + w1_q, + w2_q, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + a_scale, + ) + torch.cuda.synchronize() + + min_run_time = 5 + num_warmup = 5 + num_runs = 25 + + globals = { + # Baseline params + "w1": w1, + "w2": w2, + "score": score, + "topk": topk, + # Cutlass params + "a_scale": a_scale, + "w1_q": w1_q, + "w2_q": w2_q, + "w1_scale": w1_scale, + "w2_scale": w2_scale, + "per_act_token": per_act_token, + # cuda graph params + "cutlass_graph": cutlass_graph, + "triton_graph": triton_graph, + # Gen params + "a": a, + "topk_weights": topk_weights, + "topk_ids": topk_ids, + "num_runs": num_runs, + # Kernels + "run_triton_moe": run_triton_moe, + "run_cutlass_moe": run_cutlass_moe, + "replay_graph": replay_graph, + } + + # Warmup + run_triton_moe( + a, + w1_q, + w2_q, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + a_scale, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_triton_moe(a, w1_q, w2_q, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(triton_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(triton_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + run_cutlass_moe( + a, + a_scale, + w1_q, + w2_q, + w1_scale, + w2_scale, + topk_weights, + topk_ids, + per_act_token, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, per_act_token, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="grouped_gemm_moe", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(cutlass_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(cutlass_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="grouped_gemm_moe_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + +def main(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + results: list[benchmark.Measurement] = [] + + for model in args.models: + for tp in args.tp_sizes: + for layer in WEIGHT_SHAPES_MOE[model]: + num_experts = layer[0] + topk = layer[1] + size_k = layer[2] + size_n = layer[3] // tp + + if len(args.limit_k) > 0 and size_k not in args.limit_k: + continue + + if len(args.limit_n) > 0 and size_n not in args.limit_n: + continue + + for per_act_token in PER_ACT_TOKEN_OPTS: + for per_out_ch in PER_OUT_CH_OPTS: + for size_m in DEFAULT_BATCH_SIZES: + mkn = (size_m, size_k, size_n) + bench_run( + results, + model, + num_experts, + topk, + per_act_token, + per_out_ch, + mkn, + ) + + compare = benchmark.Compare(results) + compare.print() + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark Marlin across specified models/shapes/batches" + ) + parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES_MOE.keys(), + ) + parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES) + parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + parser.add_argument("--limit-k", nargs="+", type=int, default=[]) + parser.add_argument("--limit-n", nargs="+", type=int, default=[]) + parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-act-token", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[]) + + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_layernorm.py b/online_apiserver_test/benchmarks/kernels/benchmark_layernorm.py new file mode 100644 index 0000000000000000000000000000000000000000..69978ec6b23e94ce21d8f3dfb633317b3803e6ef --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_layernorm.py @@ -0,0 +1,93 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import time + +import torch + +from vllm.model_executor.layers.layernorm import RMSNorm +from vllm.platforms import current_platform +from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser + + +@torch.inference_mode() +def main( + num_tokens: int, + hidden_size: int, + add_residual: bool, + dtype: torch.dtype, + seed: int = 0, + do_profile: bool = False, + num_warmup_iters: int = 5, + num_iters: int = 100, +) -> None: + current_platform.seed_everything(seed) + torch.set_default_device("cuda") + + layer = RMSNorm(hidden_size).to(dtype=dtype) + layer.weight.data.normal_(mean=1.0, std=0.1) + scale = 1 / (2 * hidden_size) + x = torch.randn(num_tokens, hidden_size, dtype=dtype) + x *= scale + residual = torch.randn_like(x) * scale if add_residual else None + + def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: + torch.cuda.synchronize() + if profile: + torch.cuda.cudart().cudaProfilerStart() + start_time = time.perf_counter() + + for _ in range(num_iters): + layer(x, residual) + torch.cuda.synchronize() + + end_time = time.perf_counter() + if profile: + torch.cuda.cudart().cudaProfilerStop() + return (end_time - start_time) / num_iters + + # Warmup. + print("Warming up...") + run_benchmark = run_cuda_benchmark + run_benchmark(num_iters=num_warmup_iters, profile=False) + + # Benchmark. + if do_profile: + latency = run_benchmark(num_iters=1, profile=True) + else: + latency = run_benchmark(num_iters=num_iters, profile=False) + print(f"Kernel running time: {latency * 1000000:.3f} us") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser(description="Benchmark the layernorm kernel.") + parser.add_argument("--num-tokens", type=int, default=4096) + parser.add_argument("--hidden-size", type=int, default=8192) + parser.add_argument("--add-residual", action="store_true") + parser.add_argument( + "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half" + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--profile", action="store_true") + parser.add_argument("--num-warmup-iters", type=int, default=5) + parser.add_argument( + "--num-iters", + type=int, + default=100, + help="Number of benchmark iterations. " + "If --profile is set, this number is ignored", + ) + + args = parser.parse_args() + print(args) + + main( + num_tokens=args.num_tokens, + hidden_size=args.hidden_size, + add_residual=args.add_residual, + dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype], + seed=args.seed, + do_profile=args.profile, + num_warmup_iters=args.num_warmup_iters, + num_iters=args.num_iters, + ) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_lora.py b/online_apiserver_test/benchmarks/kernels/benchmark_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..3d38d4b3534e8993abc8aff85712230e7262a740 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_lora.py @@ -0,0 +1,1065 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import json +import pickle +import time +from dataclasses import dataclass +from enum import Enum, auto +from itertools import product +from pathlib import Path +from typing import Any, Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import ArgPool, Bench, CudaGraphBenchParams +from weight_shapes import WEIGHT_SHAPES + +from vllm.triton_utils import HAS_TRITON + +if HAS_TRITON: + from vllm.lora.ops.triton_ops import LoRAKernelMeta, lora_expand, lora_shrink + from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT + +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_TP_SIZES = [1] +DEFAULT_BATCH_SIZES = [ + 1, + 16, + 32, + 64, + 128, + 192, + 256, + 320, + 384, + 448, + 512, + 640, + 768, + 896, + 1024, + 2048, + 3072, + 4096, + 5120, + 6144, + 7168, + 8192, +] +DEFAULT_HIDDEN_SIZES = [1024, 2048, 4096, 8192, 16384] +DEFAULT_LORA_RANKS = [16] +DEFAULT_NUM_LORAS = [1, 2, 3, 4] +DEFAULT_SORT_BY_LORA_IDS = [False, True] +DEFAULT_SEQ_LENGTHS = [1] +DEFAULT_EXPAND_FN_ADD_INPUTS = [True, False] + + +# Utilities +def dtype_to_str(dtype: torch.dtype): + if dtype == torch.float16: + return "f16" + if dtype == torch.bfloat16: + return "bf16" + if dtype == torch.float32: + return "f32" + raise ValueError(f"Unsupported dtype {dtype}") + + +def make_rand_lora_weight_tensor( + k: int, n: int, num_loras: int, dtype: torch.dtype, device: str = "cuda" +) -> torch.Tensor: + # LoRA weights column major + return torch.rand((num_loras, n, k), dtype=dtype).to(device) + + +def make_rand_tensors( + a_shape: tuple[int], + b_shape: tuple[int], + c_shape: tuple[int], + a_dtype: torch.dtype, + b_dtype: torch.dtype, + c_dtype: torch.dtype, + num_slices: int, + device: str = "cuda", +) -> tuple[torch.Tensor, list[torch.Tensor], torch.Tensor]: + """ + Make LoRA input/output matrices. + """ + A = torch.rand(a_shape, dtype=a_dtype).to(device) + + # LoRA weights column major + Bs = [torch.rand(b_shape, dtype=b_dtype).to(device) for _ in range(num_slices)] + + C = torch.zeros(c_shape, dtype=c_dtype).to(device) + return A, Bs, C + + +def make_prompt_lora_mapping( + num_prompts: int, num_active_loras: int, sort_by_lora_id: bool, device: str +) -> torch.Tensor: + """ + All prompts are mapped to a LoRA ID in range [0, num_active_loras). + where 0 refers to first lora, 1 refers to second lora and so on. + """ + assert num_active_loras > 0 + + if not sort_by_lora_id: + return torch.randint(0, num_active_loras, (num_prompts,), dtype=torch.long) + + # Divide LoRAs equally and in order. + part_size = num_prompts // num_active_loras + part_size = max(part_size, 1) + + lora_id = 0 + prompt_lora_mapping = [] + while len(prompt_lora_mapping) < num_prompts: + prompt_lora_mapping.extend([lora_id] * part_size) + lora_id = lora_id + 1 if lora_id + 1 < num_active_loras else lora_id + return torch.tensor( + prompt_lora_mapping[:num_prompts], dtype=torch.long, device=device + ) + + +def make_token_lora_mapping( + num_tokens: int, + num_prompts: int, + prompt_lora_mapping: torch.Tensor, + seq_len_tensor: torch.Tensor, + device: str, +): + """ + Make token_lora_mapping from prompt_lora_mapping and seq_lens_tensor + """ + assert prompt_lora_mapping.shape[0] == num_prompts + + # token to lora index mapping + token_lora_mapping = [0] * num_tokens + current_offset = 0 + for b_id in range(num_prompts): + lora_index = prompt_lora_mapping[b_id].item() + s = current_offset + e = s + seq_len_tensor[b_id].item() + token_lora_mapping[s:e] = [lora_index] * (e - s) + current_offset += seq_len_tensor[b_id].item() + + return torch.tensor(token_lora_mapping, dtype=torch.long, device=device) + + +def ref_group_gemm( + ref_out: torch.Tensor, + input: torch.Tensor, + lora_weights: list[torch.Tensor], + seq_lens_cpu: torch.Tensor, + prompt_lora_mapping_cpu: torch.Tensor, + scaling: float, + add_inputs: Optional[bool], +): + """ + Torch group gemm reference implementation to test correctness of + benchmarking operations. + """ + batches = seq_lens_cpu.size(0) + out_list = [] + current_offset = 0 + for lora_index, b_length in zip(range(batches), seq_lens_cpu): + x = input[current_offset : b_length + current_offset, :] + current_offset += b_length + w = lora_weights[prompt_lora_mapping_cpu[lora_index]] + result = torch.nn.functional.linear(x, w) + result *= scaling + out_list.append(result) + + cat_result = torch.cat(out_list, dim=0) + + if add_inputs: + ref_out += cat_result + else: + ref_out.copy_(cat_result) + + +class OpType(Enum): + """ + LoRA Ops to benchmark and its properties. + """ + + LORA_SHRINK = auto() + LORA_EXPAND = auto() + + @staticmethod + def from_str(s: str) -> "OpType": + if s.lower() == "lora_shrink": + return OpType.LORA_SHRINK + if s.lower() == "lora_expand": + return OpType.LORA_EXPAND + raise ValueError(f"Unrecognized str {s} to convert to OpType") + + def is_shrink_fn(self) -> bool: + return self in [OpType.LORA_SHRINK] + + def is_expand_fn(self) -> bool: + return self in [OpType.LORA_EXPAND] + + def num_slices(self) -> list[int]: + return [1, 2, 3] + + def mkn( + self, batch_size: int, seq_length: int, hidden_size: int, lora_rank: int + ) -> tuple[int, int, int]: + num_tokens = batch_size * seq_length + if self.is_shrink_fn(): + m = num_tokens + k = hidden_size + n = lora_rank + else: + assert self.is_expand_fn() + m = num_tokens + k = lora_rank + n = hidden_size + return m, k, n + + def matmul_dtypes( + self, op_dtype: torch.dtype + ) -> tuple[torch.dtype, torch.dtype, torch.dtype]: + """ + return a type, b type and c type for A x B = C + """ + if self.is_shrink_fn(): + return op_dtype, op_dtype, torch.float32 + else: + assert self.is_expand_fn() + return torch.float32, op_dtype, op_dtype + + def matmul_shapes( + self, + batch_size: int, + seq_length: int, + hidden_size: int, + lora_rank: int, + num_loras: int, + num_slices: int, + ) -> tuple[tuple[int], tuple[int], tuple[int]]: + """ + Given num_slices, return the shapes of the A, B, and C matrices + in A x B = C, for the op_type + """ + m, k, n = self.mkn(batch_size, seq_length, hidden_size, lora_rank) + + b_shape = (num_loras, n, k) # col-major + if self in [OpType.LORA_SHRINK]: + # LoRA shrink kernels support num_slices inherently in the kernel. + return ((m, k), b_shape, (num_slices, m, n)) + if self in [OpType.LORA_EXPAND]: + # LoRA expand kernels support num_slices inherently in the kernel + return ((num_slices, m, k), b_shape, (m, n * num_slices)) + raise ValueError(f"Unrecognized op_type {self}") + + def bench_fn(self) -> Callable: + if self == OpType.LORA_SHRINK: + return lora_shrink + if self == OpType.LORA_EXPAND: + return lora_expand + + raise ValueError(f"Unrecognized optype {self}") + + def run_ref_group_gemm( + self, + output: torch.Tensor, + input: torch.Tensor, + lora_weights: list[torch.Tensor], + **kwargs, + ) -> Callable: + """Each benchmark operation expects the input, lora_weights and outputs + in a slightly different format. Refer to self.matmul_shapes(). + run_ref_group_gemm accounts for those differences in executing a + reference group gemm for correctness testing. + """ + w_dtype = lora_weights[0].dtype + num_slices = len(lora_weights) + if self in [OpType.LORA_SHRINK]: + for slice_idx in range(num_slices): + ref_group_gemm( + ref_out=output[slice_idx, :], + input=input, + lora_weights=lora_weights[slice_idx], + **kwargs, + ) + elif self in [OpType.LORA_EXPAND]: + hidden_size = lora_weights[0].shape[1] + for slice_idx in range(num_slices): + slice_offset = slice_idx * hidden_size + ref_group_gemm( + ref_out=output[:, slice_offset : slice_offset + hidden_size], + input=input[slice_idx].clone().to(dtype=w_dtype), + lora_weights=lora_weights[slice_idx], + **kwargs, + ) + else: + raise ValueError(f"Unrecognized optype {self}") + + +@dataclass +class BenchmarkContext: + """ + LoRA benchmark context + """ + + batch_size: int + hidden_size: int + num_loras: int + num_active_loras: int + lora_rank: int + sort_by_lora_id: bool + dtype: torch.dtype + seq_length: Optional[int] = None + num_slices: Optional[int] = None # num_slices for slice based ops + + def with_seq_length(self, seq_length: int) -> "BenchmarkContext": + ctx = copy.copy(self) + ctx.seq_length = seq_length + return ctx + + def with_num_slices(self, num_slices: int) -> "BenchmarkContext": + ctx = copy.copy(self) + ctx.num_slices = num_slices + return ctx + + def bench_label(self) -> str: + return f"lora-{self.dtype}" + + def bench_sublabel(self, op_type: OpType) -> str: + m, k, n = op_type.mkn( + self.batch_size, self.seq_length, self.hidden_size, self.lora_rank + ) + desc = { + "bs": self.batch_size, + "sl": self.seq_length, + "m": m, + "k": k, + "n": n, + "num_loras": self.num_loras, + "sort_by_lora": self.sort_by_lora_id, + "num_slices": self.num_slices, + } + return json.dumps(desc) + + +@dataclass +class BenchmarkTensors: + """ + Input/Output tensors used for benchmarks + """ + + # matmul tensors + input: torch.Tensor + lora_weights_lst: list[torch.Tensor] + output: torch.Tensor + # LoRA kernel metadata + lora_kernel_meta: LoRAKernelMeta + # Metadata tensors used in testing correctness + seq_lens: torch.Tensor + prompt_lora_mapping: torch.Tensor + + def io_types(self) -> str: + return ( + f"{dtype_to_str(self.input.dtype)}x" + f"{dtype_to_str(self.lora_weights_lst[0].dtype)}=>" + f"{dtype_to_str(self.output.dtype)}" + ) + + @staticmethod + def make( + ctx: BenchmarkContext, op_type: OpType, device: str = "cuda" + ) -> "BenchmarkTensors": + # Make input / output matmul tensors. + a_shape, b_shape, c_shape = op_type.matmul_shapes( + ctx.batch_size, + ctx.seq_length, + ctx.hidden_size, + ctx.lora_rank, + ctx.num_loras, + ctx.num_slices, + ) + a_type, b_type, c_type = op_type.matmul_dtypes(ctx.dtype) + input_tensor, lora_weights, output_tensor = make_rand_tensors( + a_shape, b_shape, c_shape, a_type, b_type, c_type, num_slices=ctx.num_slices + ) + + # Make metadata tensors. + # Keep the metadata tensors in the CPU for further processing if needed. + # The tensors get moved to the GPU before benchmarking. + assert ctx.num_active_loras <= ctx.num_loras + total_tokens = ctx.batch_size * ctx.seq_length + + # Make metadata tensors involved in correctness testing. + # Prepare seq lens tensor + seq_len_tensor = torch.randint( + ctx.seq_length, ctx.seq_length + 1, (ctx.batch_size,) + ) + assert total_tokens == seq_len_tensor.sum() + # Prepare prompt lora indices tensor + prompt_lora_indices_tensor = make_prompt_lora_mapping( + ctx.batch_size, ctx.num_active_loras, ctx.sort_by_lora_id, "cpu" + ) + + # Make LoRAKernelMeta + token_lora_indices_tensor = make_token_lora_mapping( + total_tokens, + ctx.batch_size, + prompt_lora_indices_tensor, + seq_len_tensor, + "cpu", + ) + lora_kernel_meta = LoRAKernelMeta.make( + max_loras=ctx.num_loras, + max_num_tokens=token_lora_indices_tensor.size(0), + device="cpu", + ) + lora_kernel_meta.prepare_tensors(token_lora_mapping=token_lora_indices_tensor) + + return BenchmarkTensors( + input_tensor, + lora_weights, + output_tensor, + lora_kernel_meta, + seq_len_tensor, + prompt_lora_indices_tensor, + ) + + def sanity_check(self) -> None: + """ + Fails asserts when non-conformality is detected. + """ + num_tokens = self.input.shape[-2] + # check metadata tensors + assert torch.sum(self.seq_lens) == num_tokens + num_seqs = self.seq_lens.shape[0] + # assert self.seq_start_loc.shape[0] == num_seqs + assert self.prompt_lora_mapping.shape[0] == num_seqs + assert self.lora_kernel_meta.token_lora_mapping.shape[0] == num_tokens + + def to_device(self, device: str): + """ + Transfer tensors to device if the tensors aren't already on the device + """ + + def to_device(tensor: torch.Tensor): + if tensor.device != device: + tensor = tensor.to(device=device) + return tensor + + self.input = to_device(self.input) + self.output = to_device(self.output) + self.seq_lens = to_device(self.seq_lens) + self.prompt_lora_mapping = to_device(self.prompt_lora_mapping) + for i in range(len(self.lora_weights_lst)): + self.lora_weights_lst[i] = to_device(self.lora_weights_lst[i]) + + # LoRA meta + for field_name in LoRAKernelMeta.__dataclass_fields__: + field = getattr(self.lora_kernel_meta, field_name) + assert isinstance(field, torch.Tensor) + setattr(self.lora_kernel_meta, field_name, to_device(field)) + + def metadata(self) -> tuple[int, int, int]: + """ + Return num_seqs, num_tokens and max_seq_len + """ + num_seqs = self.seq_lens.shape[0] + num_tokens = self.lora_kernel_meta.token_lora_mapping.shape[0] + max_seq_len = torch.max(self.seq_lens).item() + num_slices = len(self.lora_weights_lst) + return num_seqs, num_tokens, max_seq_len, num_slices + + def as_lora_shrink_kwargs(self) -> dict[str, Any]: + self.sanity_check() + self.to_device(self.input.device) + + _, num_tokens, _, num_slices = self.metadata() + + # Sanity check matrix shapes. + i_shape, lw_shape, o_shape = ( + self.input.shape, + self.lora_weights_lst[0].shape, + self.output.shape, + ) + # Expected input shape [num_tokens, hidden_size] + assert len(i_shape) == 2 + assert i_shape[0] == num_tokens + hidden_size = i_shape[1] + # Expected lora weight shape [num_loras, lora_rank, hidden_size] + assert len(lw_shape) == 3 + assert lw_shape[2] == hidden_size + lora_rank = lw_shape[1] + # Expected output shape [num_slices, num_tokens, lora_rank] + assert len(o_shape) == 3 + assert o_shape == (num_slices, num_tokens, lora_rank) + + return { + "inputs": self.input, + "lora_a_weights": self.lora_weights_lst, + "output_tensor": self.output, + "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, + "token_indices_sorted_by_lora_ids": ( + self.lora_kernel_meta.token_indices_sorted_by_lora_ids + ), + "num_tokens_per_lora": self.lora_kernel_meta.num_tokens_per_lora, + "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc, + "lora_ids": self.lora_kernel_meta.active_lora_ids, + "scaling": 1.0, + } + + def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]: + self.sanity_check() + self.to_device(self.input.device) + + _, num_tokens, _, num_slices = self.metadata() + + # Sanity check matrix shapes. + i_shape, lw_shape, o_shape = ( + self.input.shape, + self.lora_weights_lst[0].shape, + self.output.shape, + ) + # Expected input shape : [num_slices, num_tokens, lora_rank] + assert len(i_shape) == 3 + assert i_shape[0] == num_slices + assert i_shape[1] == num_tokens + lora_rank = i_shape[2] + # Expected lora weight shape : [num_lora, hidden_size, lora_rank] + assert len(lw_shape) == 3 + assert lw_shape[2] == lora_rank + hidden_size = lw_shape[1] + # Expected output shape : [num_tokens, hidden_size * num_slices] + assert len(o_shape) == 2 + assert o_shape == (num_tokens, hidden_size * num_slices) + + return { + "inputs": self.input, + "lora_b_weights": self.lora_weights_lst, + "output_tensor": self.output, + "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, + "token_indices_sorted_by_lora_ids": ( + self.lora_kernel_meta.token_indices_sorted_by_lora_ids + ), + "num_tokens_per_lora": self.lora_kernel_meta.num_tokens_per_lora, + "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc, + "lora_ids": self.lora_kernel_meta.active_lora_ids, + "offset_start": 0, + "add_inputs": add_inputs, + } + + def bench_fn_kwargs( + self, op_type: OpType, add_inputs: Optional[bool] = None + ) -> dict[str, Any]: + if op_type.is_shrink_fn(): + assert add_inputs is None + else: + assert add_inputs is not None + + if op_type == OpType.LORA_SHRINK: + return self.as_lora_shrink_kwargs() + if op_type == OpType.LORA_EXPAND: + return self.as_lora_expand_kwargs(add_inputs) + raise ValueError(f"Unrecognized optype {self}") + + def test_correctness( + self, op_type: OpType, expand_fn_add_inputs: Optional[bool] + ) -> bool: + """ + Test correctness of op_type implementation against a grouped gemm + reference implementation. + """ + seq_lens_cpu = self.seq_lens.to(device="cpu") + prompt_lora_mapping_cpu = self.prompt_lora_mapping.to(device="cpu") + ref_output = self.output.clone() + + self.output.zero_() + op_type.bench_fn()(**self.bench_fn_kwargs(op_type, expand_fn_add_inputs)) + + op_type.run_ref_group_gemm( + ref_output, + self.input, + self.lora_weights_lst, + seq_lens_cpu=seq_lens_cpu, + prompt_lora_mapping_cpu=prompt_lora_mapping_cpu, + scaling=1.0, + add_inputs=expand_fn_add_inputs, + ) + + rtol, atol = { + torch.float16: (6e-2, 6e-2), + torch.bfloat16: (6e-2, 6e-2), + torch.float32: (1e-2, 1e-2), + }[self.output.dtype] + + return torch.allclose(ref_output, self.output, rtol=rtol, atol=atol) + + +def bench_optype( + ctx: BenchmarkContext, + arg_pool_size: int, + op_type: OpType, + cuda_graph_nops: Optional[int] = None, + expand_fn_add_inputs: Optional[bool] = None, + test_correctness: bool = False, +) -> TMeasurement: + assert arg_pool_size >= 1 + if op_type.is_shrink_fn(): + assert expand_fn_add_inputs is None + else: + assert expand_fn_add_inputs is not None + + # BenchmarkContext -> BenchmarkTensors + bench_tensors: list[BenchmarkTensors] = [ + BenchmarkTensors.make(ctx, op_type) for _ in range(arg_pool_size) + ] + for bt in bench_tensors: + bt.sanity_check() + + # Test correctness of our implementation. + if test_correctness: + assert all( + [bt.test_correctness(op_type, expand_fn_add_inputs) for bt in bench_tensors] + ) + + # BenchmarkTensors -> dict (kwargs) + kwargs_list = [ + bt.bench_fn_kwargs(op_type, add_inputs=expand_fn_add_inputs) + for bt in bench_tensors + ] + + # Clear LoRA optimization hash-maps. + _LORA_A_PTR_DICT.clear() + _LORA_B_PTR_DICT.clear() + # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are setup + for kwargs in kwargs_list: + op_type.bench_fn()(**kwargs) + torch.cuda.synchronize() + + # Merge into a single kwargs and qualify arguments as ArgPool + kwargs = {k: ArgPool([]) for k in kwargs_list[0]} + for _kwargs in kwargs_list: + for k, v in _kwargs.items(): + kwargs[k].values.append(v) + + describe_args = ( + f"add_inputs={expand_fn_add_inputs}" if expand_fn_add_inputs is not None else "" + ) + description = f"{op_type.name}({describe_args}) ({bench_tensors[0].io_types()})" + + cuda_graph_params = None + if cuda_graph_nops: + cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) + timer = None + with Bench( + cuda_graph_params, + ctx.bench_label(), + ctx.bench_sublabel(op_type), + description, + op_type.bench_fn(), + **kwargs, + ) as bench: + timer = bench.run() + return timer + + +def bench_torch_mm( + ctx: BenchmarkContext, + arg_pool_size: int, + op_type: OpType, + cuda_graph_nops: Optional[int] = None, +) -> TMeasurement: + """ + Benchmark basic torch.mm as a roofline. + + When all the input tokens have the same LoRA ID, the LoRA kernels are just + a matmul. This torch.mm benchmark serves as a roofline for that case. + + input op_type is used in determining the m, k, n dimensions for the matmul. + """ + + batch_size, hidden_size, lora_rank, seq_length, dtype = ( + ctx.batch_size, + ctx.hidden_size, + ctx.lora_rank, + ctx.seq_length, + ctx.dtype, + ) + + m, k, n = op_type.mkn(batch_size, seq_length, hidden_size, lora_rank) + # For a fairer comparison. + n = n * ctx.num_slices + + # Get matmul input and output tensors for A x B = C + As, Bs, Cs = [], [], [] + for _ in range(arg_pool_size): + As.append(torch.rand((m, k), dtype=dtype).to("cuda")) + Bs.append(torch.rand((n, k), dtype=dtype).to("cuda").t()) + Cs.append(torch.rand((m, n), dtype=dtype).to("cuda")) + + # Make torch.mm kwargs + mm_kwargs = {"input": ArgPool(As), "mat2": ArgPool(Bs), "out": ArgPool(Cs)} + + description = ( + f"single-lora roofline using torch.mm ({dtype_to_str(dtype)}" + f"x{dtype_to_str(dtype)}" + f"=>{dtype_to_str(dtype)})" + ) + cuda_graph_params = None + if cuda_graph_nops: + cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) + with Bench( + cuda_graph_params, + ctx.bench_label(), + ctx.bench_sublabel(op_type), + description, + torch.mm, + **mm_kwargs, + ) as bench: + return bench.run() + + +# runner +def use_cuda_graph_recommendation() -> str: + return """ + Triton kernels have a significant launch overhead with + launched directly via python. This overhead is more noticeable + for small the problem sizes. For these cases, it is recommended + to use the script with `--cuda-graph-nops N` to benchmark N + consecutive invocations of the benchmarking operations from + inside a CUDA Graph. Note that the returned measurement is for N + invocations of the operation. + """ + + +def print_timers(timers: list[TMeasurement], args: Optional[argparse.Namespace] = None): + compare = TBenchmark.Compare(timers) + compare.print() + + if args and args.cuda_graph_nops: + print( + f"Note : The timings reported above is for {args.cuda_graph_nops} " + "consecutive invocations of the benchmarking functions. " + f"Please divide by {args.cuda_graph_nops} for single invocation " + "timings." + ) + + print( + "Note on Comparison with torch.mm : The torch.mm numbers are " + "benchmark numbers of a simple matmul emulating the single lora " + "case. It is provided as a roofline for comparing our LoRA Kernel " + "implementations. It is expected that the LoRA kernels will be " + "slower than torch.mm in cases where num_loras is big. But for " + "small num_loras the goal should be to match the torch.mm numbers." + ) + + +def run(args: argparse.Namespace, bench_ctxs: list[BenchmarkContext]): + if args.cuda_graph_nops is not None: + assert args.cuda_graph_nops > 0 + print(f"Benchmarking {args.cuda_graph_nops} invocations inside a CUDA Graph") + else: + print(f"CUDA Graphs not enabled.\n{use_cuda_graph_recommendation()}") + + timers = [] + for bench_ctx in bench_ctxs: + for seq_len in args.seq_lengths: + bench_ops: list[OpType] = args.op_types + seq_len_timers = [] + for bench_op in bench_ops: + for num_slices in bench_op.num_slices(): + _ctx = bench_ctx.with_seq_length(seq_len).with_num_slices( + num_slices + ) + # Benchmark torch.mm as a roofline + seq_len_timers.append( + bench_torch_mm( + _ctx, args.arg_pool_size, bench_op, args.cuda_graph_nops + ) + ) + + # Benchmark bench_op + expand_fn_add_inputs = ( + [None] if bench_op.is_shrink_fn() else args.expand_fn_add_inputs + ) + for add_input_arg in expand_fn_add_inputs: + seq_len_timers.append( + bench_optype( + _ctx, + args.arg_pool_size, + bench_op, + args.cuda_graph_nops, + add_input_arg, + args.test_correctness, + ) + ) + + print_timers(seq_len_timers) + timers.extend(seq_len_timers) + + # Result stdout dump + print("== All Results ====") + print_timers(timers, args) + + if args.output_directory: + # Result file dump + od = Path(args.output_directory) + if not od.exists(): + od.mkdir() + + timestamp = int(time.time()) + pkl_file = od / f"lora_bench-{timestamp}.pkl" + print(f"Writing benchmarks to {pkl_file}") + with open(pkl_file, "wb") as f: + pickle.dump(timers, f) + + +def as_benchmark_contexts( + hidden_sizes: list[int], lora_ranks: list[int], args: argparse.Namespace +) -> list[BenchmarkContext]: + ctxs: list[BenchmarkContext] = [] + for batch_size, hidden_size, lora_rank, num_loras, sort_by_lora_id in product( # noqa + args.batch_sizes, + list(hidden_sizes), + lora_ranks, + args.num_loras, + args.sort_by_lora_id, + ): + ctxs.append( + BenchmarkContext( + batch_size=batch_size, + hidden_size=hidden_size, + lora_rank=lora_rank, + num_loras=num_loras, + num_active_loras=args.num_active_loras + if args.num_active_loras + else num_loras, + # To be filled based on the OpType to benchmark + seq_length=None, + sort_by_lora_id=sort_by_lora_id, + dtype=args.dtype, + # To be filled based on the OpType to benchmark + num_slices=None, + ) + ) + + return ctxs + + +def run_list_bench(args: argparse.Namespace): + print(args) + + print( + "List bench :\n" + f" Hidden Sizes {args.hidden_sizes}" + f" LoRA Ranks {args.lora_ranks}" + ) + + # Get all benchmarking contexts + bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=args.hidden_sizes, lora_ranks=args.lora_ranks, args=args + ) + + run(args, bench_contexts) + + +def run_range_bench(args: argparse.Namespace): + print(args) + + hidden_sizes = list( + range( + args.hidden_sizes_start, + args.hidden_sizes_end + 1, + args.hidden_sizes_increment, + ) + ) + lora_ranks = list( + range(args.lora_ranks_start, args.lora_ranks_end + 1, args.lora_ranks_increment) + ) + + print(f"Range bench :\n Hidden Sizes {hidden_sizes} LoRA Ranks {lora_ranks}") + + # Get all benchmarking contexts + bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=hidden_sizes, lora_ranks=lora_ranks, args=args + ) + + run(args, bench_contexts) + + +def run_model_bench(args: argparse.Namespace): + print(args) + + def hidden_sizes_from_model(model: str, tp_size: int) -> set[int]: + hidden_sizes = set() + for KN, tp_split_dim in WEIGHT_SHAPES[model]: + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + hidden_sizes.add(KN[1]) + return hidden_sizes + + # Get all hidden sizes + hidden_sizes: set[int] = set() + for model_name, tp_size in product(args.models, args.tp_sizes): + hidden_sizes = hidden_sizes.union(hidden_sizes_from_model(model_name, tp_size)) + + print(f"Model bench :\n Hidden Sizes {hidden_sizes} LoRA Ranks {args.lora_ranks}") + + # Get all benchmarking contexts + bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=hidden_sizes, lora_ranks=args.lora_ranks, args=args + ) + + run(args, bench_contexts) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "torch.float16": + return torch.float16 + if dt == "torch.bfloat16": + return torch.bfloat16 + raise ValueError("unsupported dtype") + + def get_bool(s: str) -> bool: + return s.lower() in ["true", "1"] + + def add_common_command_args(p: argparse.ArgumentParser): + p.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['torch.float16', 'torch.bfloat16']", + ) + + p.add_argument( + "--arg-pool-size", + type=int, + default=32, + help="Run profiles with a pool of input/output/meta tensors instead" + "of simply reusing the same tensors for all runs. A bigger arg-pool" + "mitigates hardware caching effects during benchmarking.", + ) + + p.add_argument( + "--cuda-graph-nops", + type=int, + help=( + "when set profiling is done using cudagraph, " + "with the given number of operations in a graph." + "Note that the measurement returned is the time " + "taken for N consecutive executions of the benchmarking " + "functions, where N is the value of this argument." + ), + ) + p.add_argument("--num-loras", nargs="+", type=int, default=DEFAULT_NUM_LORAS) + p.add_argument( + "--num-active-loras", + type=int, + default=None, + help="Active LoRAs. When None, all LoRAs are active", + ) + p.add_argument( + "--sort-by-lora-id", + nargs="+", + type=get_bool, + default=DEFAULT_SORT_BY_LORA_IDS, + ) + p.add_argument( + "--op-types", nargs="+", type=OpType.from_str, default=list(OpType) + ) + p.add_argument( + "--seq-lengths", nargs="+", type=int, default=DEFAULT_SEQ_LENGTHS + ) + p.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + p.add_argument( + "--expand-fn-add-inputs", + nargs="+", + type=get_bool, + default=DEFAULT_EXPAND_FN_ADD_INPUTS, + ) + p.add_argument( + "-o", + "--output-directory", + type=str, + help=( + "Output directory to store a the list of benchmarking" + "TMeasurement objects as a pickle file" + ), + ) + + p.add_argument( + "--test-correctness", + action="store_true", + help=( + "When enabled, the benchmarking functions are tested" + "for correctness before the actual benchmarking" + ), + ) + + parser = FlexibleArgumentParser( + description=f""" +Benchmark LoRA kernels: + {use_cuda_graph_recommendation()} + + list_bench example: + python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 + + model_bench example: + python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 + + range_bench example: + python3 benchmarks/kernels/benchmark_lora.py range_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + + subparsers = parser.add_subparsers(dest="cmd", required=True) + + list_parser = subparsers.add_parser("list_bench") + list_parser.add_argument( + "--hidden-sizes", nargs="+", type=int, default=DEFAULT_HIDDEN_SIZES + ) + list_parser.add_argument( + "--lora-ranks", nargs="+", type=int, default=DEFAULT_LORA_RANKS + ) + add_common_command_args(list_parser) + list_parser.set_defaults(func=run_list_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--hidden-sizes-start", type=int, required=True) + range_parser.add_argument("--hidden-sizes-end", type=int, required=True) + range_parser.add_argument("--hidden-sizes-increment", type=int, required=True) + range_parser.add_argument("--lora-ranks-start", type=int, required=True) + range_parser.add_argument("--lora-ranks-end", type=int, required=True) + range_parser.add_argument("--lora-ranks-increment", type=int, required=True) + add_common_command_args(range_parser) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--lora-ranks", nargs="+", type=int, default=DEFAULT_LORA_RANKS + ) + add_common_command_args(model_parser) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_machete.py b/online_apiserver_test/benchmarks/kernels/benchmark_machete.py new file mode 100644 index 0000000000000000000000000000000000000000..f73d0511e01fc3851847278f8604a2baed8c8f33 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_machete.py @@ -0,0 +1,732 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import itertools +import math +import os +import pickle as pkl +import time +from collections.abc import Iterable +from dataclasses import dataclass +from itertools import product +from typing import Callable, Optional + +import pandas as pd +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from weight_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.marlin_utils import ( + GPTQ_MARLIN_MAX_PARALLEL, + GPTQ_MARLIN_MIN_THREAD_N, + marlin_permute_scales, + marlin_zero_points, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_test import ( + MarlinWorkspace, +) +from vllm.model_executor.layers.quantization.utils.quant_utils import ( + pack_rows, + quantize_weights, +) +from vllm.scalar_type import ScalarType, scalar_types +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = ["meta-llama/Llama-3-8b", "meta-llama/Llama-2-70b-hf"] +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024] +DEFAULT_TP_SIZES = [1] + +NVTX_PROFILE = os.environ.get("NVTX_PROFILE", False) + +if NVTX_PROFILE: + import nvtx + + +def terse_type_name(dt): + return { + torch.bfloat16: "bf16", + torch.float16: "fp16", + torch.int8: "int8", + torch.float8_e4m3fn: "fp8", + torch.float: "float", + torch.int: "int", + }[dt] + + +@dataclass +class BenchmarkTensors: + w_ref: torch.Tensor + a: torch.Tensor + + w_q: torch.Tensor + group_size: Optional[int] + wtype: ScalarType + w_g_s: torch.Tensor + w_g_zp: Optional[torch.Tensor] + w_ch_s: Optional[torch.Tensor] + w_tok_s: Optional[torch.Tensor] + + +@dataclass +class TypeConfig: + act_type: torch.dtype + weight_type: ScalarType + output_type: Optional[torch.dtype] + group_scale_type: Optional[torch.dtype] + group_zero_type: Optional[torch.dtype] + channel_scale_type: Optional[torch.dtype] + token_scale_type: Optional[torch.dtype] + + +def rand_data(shape, dtype=torch.float16, scale=1): + if dtype.is_floating_point: + return (scale * torch.rand(shape, device="cuda") - 0.3).to(dtype) + else: + return torch.randint(-15, 15, shape, dtype=dtype, device="cuda") + + +def quantize_and_pack( + atype: torch.dtype, + w: torch.Tensor, + wtype: ScalarType, + stype: Optional[torch.dtype], + group_size: Optional[int], + zero_points: bool = False, +): + assert wtype.is_integer(), "TODO: support floating point weights" + + w_ref, w_q, w_s, w_zp = quantize_weights( + w, + wtype, + group_size=group_size, + zero_points=zero_points, + # to match how the kernel applies zps + ref_zero_points_after_scales=True, + ) + + w_q = pack_rows(w_q, wtype.size_bits, *w_q.shape) + return w_ref, w_q, w_s, w_zp + + +def create_bench_tensors( + shape: tuple[int, int, int], types: TypeConfig, group_size: Optional[int] +) -> list[BenchmarkTensors]: + m, n, k = shape + + # we want to make sure that weights don't fit into L2 cache between runs so + # we construct enough weights to exceed L2 cache, which is 50mb on a H100 + # so we target total weight size > 2*50mb + num_weights = math.ceil( + 2 * 50 * 1024**2 * 8 / (k * n * types.weight_type.size_bits) + ) + + a = rand_data((m, k), types.act_type, scale=5) + + benchmark_tensors: list[BenchmarkTensors] = [] + for _ in range(num_weights): + w = rand_data((k, n), types.act_type, scale=5) + + if types.group_scale_type is not None: + w = w.to(types.group_scale_type) + if w.dtype.itemsize == 1: + w = w.to(torch.float16) + + w_ref, w_q_packed, w_s, w_zp = quantize_and_pack( + a.dtype, + w, + types.weight_type, + types.group_scale_type, + group_size, + types.group_zero_type is not None, + ) + + if not a.dtype.is_floating_point: + aiinfo = torch.iinfo(a.dtype) + w_ref = w_ref.round().clamp(aiinfo.min, aiinfo.max) + + w_ref = w_ref.to(torch.float32) + + w_ch_s = ( + None + if types.channel_scale_type is None + else rand_data((n,), types.channel_scale_type) + ) + w_tok_s = ( + None + if types.token_scale_type is None + else rand_data((m,), types.token_scale_type) + ) + + benchmark_tensors.append( + BenchmarkTensors( + w_ref=w_ref, + a=a, + w_q=w_q_packed, + wtype=types.weight_type, + w_g_s=w_s, + w_g_zp=w_zp, + group_size=group_size, + w_ch_s=w_ch_s, + w_tok_s=w_tok_s, + ) + ) + + return benchmark_tensors + + +def torch_matmul_f16_create_bench_fn(bt: BenchmarkTensors) -> Callable: + a = bt.a + w = bt.w_ref.to(bt.a.dtype) # use float reference tensor + if a.dtype not in [torch.float16, torch.bfloat16]: + a = a.to(torch.float16) + w = w.to(torch.float16) + return lambda: torch.matmul(a, w) + + +def cutlass_scaled_mm_create_bench_fn(bt: BenchmarkTensors) -> Callable: + if bt.w_ch_s is not None and bt.w_tok_s is not None: + scale_a = bt.w_tok_s.to(torch.float32) + scale_b = bt.w_ch_s.to(torch.float32) + else: + scale_a = torch.tensor(1.0, dtype=torch.float32, device=bt.a.device) + scale_b = torch.tensor(1.0, dtype=torch.float32, device=bt.a.device) + w_col_major = bt.w_ref.to(bt.a.dtype).t().contiguous().t() + return lambda: ops.cutlass_scaled_mm( + bt.a, w_col_major, scale_a, scale_b, out_dtype=torch.float16 + ) + + +def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable: + device = bt.a.device + + workspace = MarlinWorkspace( + bt.w_ref.shape[1], GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL + ) + + if bt.w_g_zp is None: + w_zp = torch.empty(0, dtype=torch.int, device=device) + else: + w_zp = marlin_zero_points( + bt.w_g_zp, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits + ) + + if bt.group_size is None: + w_s = torch.tensor([], device="cuda", dtype=torch.half) + else: + w_s = marlin_permute_scales( + bt.w_g_s, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.group_size + ) + + sort_indices = torch.empty(0, dtype=torch.int, device=device) + g_idx = torch.empty(0, dtype=torch.int, device=device) + w_q = ops.gptq_marlin_repack( + bt.w_q, sort_indices, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits + ) + + if bt.a.dtype.is_floating_point: + assert bt.w_ch_s is None + assert bt.w_tok_s is None + assert bt.group_size is not None + + fn = lambda: ops.gptq_marlin_gemm( + a=bt.a, + c=None, + b_q_weight=w_q, + b_scales=w_s, + global_scale=None, + b_zeros=w_zp, + g_idx=g_idx, + perm=sort_indices, + workspace=workspace.scratch, + b_q_type=bt.wtype, + size_m=bt.a.shape[0], + size_n=bt.w_ref.shape[1], + size_k=bt.w_ref.shape[0], + is_k_full=True, + is_zp_float=False, + ) + else: + assert bt.a.dtype == torch.int8 + assert bt.wtype == scalar_types.uint4b8 + + if bt.w_ch_s is not None: + s_ch = bt.w_ch_s.to(torch.float32) + else: + s_ch = torch.ones(bt.w_ref.shape[1], dtype=torch.float32, device=device) + + if bt.w_tok_s is not None: + s_tok = bt.w_tok_s.to(torch.float32) + else: + s_tok = torch.ones(bt.a.shape[0], dtype=torch.float32, device=device) + + fn = lambda: ops.marlin_qqq_gemm( + a=bt.a, + b_q_weight=w_q, + s_group=w_s, + s_tok=s_tok, + s_ch=s_ch, + workspace=workspace.scratch, + size_m=bt.a.shape[0], + size_n=bt.w_ref.shape[1], + size_k=bt.w_ref.shape[0], + ) + + return fn + + +def machete_create_bench_fn( + bt: BenchmarkTensors, out_type=torch.dtype, schedule=None +) -> Callable: + w_q = bt.w_q.t().contiguous().t() # make col major + w_q = ops.machete_prepack_B( + w_q, bt.a.dtype, bt.wtype, None if bt.w_g_s is None else bt.w_g_s.dtype + ) + + w_g_zp = bt.w_g_zp + if w_g_zp is not None: + w_g_zp = -1 * bt.w_g_s * (w_g_zp.to(bt.w_g_s.dtype)) + + return lambda: ops.machete_mm( + a=bt.a, + b_q=w_q, + b_type=bt.wtype, + b_group_scales=bt.w_g_s, + b_group_zeros=w_g_zp, + b_group_size=bt.group_size, + b_channel_scales=bt.w_ch_s, + a_token_scales=bt.w_tok_s, + out_type=out_type, + schedule=schedule, + ) + + +# impl + +# bench + + +def bench_fns(label: str, sub_label: str, description: str, fns: list[Callable]): + min_run_time = 1 if not NVTX_PROFILE else 0.1 + res = TBenchmark.Timer( + stmt=""" + for fn in fns: + fn() + """, + globals={"fns": fns}, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + if NVTX_PROFILE: + with ( + nvtx.annotate("mm-bench"), + nvtx.annotate(f"{label}|{sub_label}|{description}"), + ): + fns[0]() + + return res + + +_SWEEP_SCHEDULES_RESULTS: Optional[pd.DataFrame] = None +_SWEEP_SCHEDULES_RESULTS_CSV: Optional[str] = None + + +def bench( + types: TypeConfig, + group_size: int, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + sweep_schedules: bool = True, +) -> list[TMeasurement]: + benchmark_tensors = create_bench_tensors((m, n, k), types, group_size) + sub_label += f", L={len(benchmark_tensors)}" + + name_type_string = f"W{types.weight_type}" + f"-A{terse_type_name(types.act_type)}" + if types.group_scale_type is not None: + name_type_string += f"-GS{terse_type_name(types.group_scale_type)}" + if types.group_zero_type is not None: + name_type_string += f"-GZ{terse_type_name(types.group_zero_type)}" + if group_size is not None: + name_type_string += f"-G{group_size}" + if types.channel_scale_type is not None: + name_type_string += f"-CS{terse_type_name(types.channel_scale_type)}" + if types.token_scale_type is not None: + name_type_string += f"-TS{terse_type_name(types.token_scale_type)}" + + timers = [] + # pytorch impl + timers.append( + bench_fns( + label, + sub_label, + "torch.matmul (fp16)", + [torch_matmul_f16_create_bench_fn(bt) for bt in benchmark_tensors], + ) + ) + + if types.act_type == torch.int8 or types.act_type == torch.float8_e4m3fn: + timers.append( + bench_fns( + label, + sub_label, + f"cutlass_scaled_mm ({terse_type_name(types.act_type)})", + [cutlass_scaled_mm_create_bench_fn(bt) for bt in benchmark_tensors], + ) + ) + + if types.act_type != torch.float8_e4m3fn: + timers.append( + bench_fns( + label, + sub_label, + f"marlin ({name_type_string})", + [marlin_create_bench_fn(bt) for bt in benchmark_tensors], + ) + ) + + # machete + timers.append( + bench_fns( + label, + sub_label, + f"machete ({name_type_string})", + [ + machete_create_bench_fn(bt, out_type=types.output_type) + for bt in benchmark_tensors + ], + ) + ) + + if sweep_schedules: + global _SWEEP_SCHEDULES_RESULTS + + print("Finding best schedule for machete") + best = None + best_schedule = None + schedules = ops.machete_supported_schedules( + a_type=types.act_type, + b_type=types.weight_type, + group_scales_type=types.group_scale_type, + group_zeros_type=types.group_zero_type, + token_scales_type=types.token_scale_type, + channel_scales_type=types.channel_scale_type, + out_type=types.output_type, + ) + + if schedules is None or len(schedules) == 0: + raise ValueError("No schedules found to sweep") + + for schedule in reversed(schedules): + schedule_M = int(schedule.split("_")[0].split("x")[1]) + + # Prune known bad schedules + if schedule_M >= 2 * max(m, 16) or schedule_M < m // 4: + continue + + res = bench_fns( + label, + sub_label, + "machete_best", + [ + machete_create_bench_fn( + bt, out_type=types.output_type, schedule=schedule + ) + for bt in benchmark_tensors + ], + ) + + results_row = { + "M": m, + "K": k, + "N": n, + "group_size": group_size, + "schedule": schedule, + "median": res.median, + } + if _SWEEP_SCHEDULES_RESULTS is None: + _SWEEP_SCHEDULES_RESULTS = pd.DataFrame(columns=results_row.keys()) + _SWEEP_SCHEDULES_RESULTS.loc[len(_SWEEP_SCHEDULES_RESULTS)] = results_row + + print(f" {res.median:5.5} ", schedule) + if not best or res.median < best.median: + best = res + best_schedule = schedule + print("Best schedule:", best_schedule) + timers.append(best) + + return timers + + +# runner +def print_timers(timers: list[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def run(args, MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]: + types = TypeConfig( + act_type=args.act_type, + weight_type=scalar_types.uint4b8 + if args.group_zero_type is None + else scalar_types.uint4, + output_type=args.out_type, + group_scale_type=args.group_scale_type, + group_zero_type=args.group_zero_type, + channel_scale_type=args.channel_scale_type, + token_scale_type=args.token_scale_type, + ) + + results: list[TMeasurement] = [] + for m, k, n in MKNs: + timers = bench( + types, + args.group_size, + m, + k, + n, + f"{args.act_type}-gemm", + f"MKN=({m}x{k}x{n})", + sweep_schedules=args.sweep_schedules, + ) + print_timers(timers) + results.extend(timers) + + return results + + +# output makers +def make_output( + data: list[TMeasurement], + MKNs: Iterable[tuple[int, int, int]], + base_description: str, + timestamp=None, +): + print(f"== All Results {base_description} ====") + print_timers(data) + + # pickle all the results + timestamp = int(time.time()) if timestamp is None else timestamp + with open(f"{base_description}-{timestamp}.pkl", "wb") as f: + pkl.dump(data, f) + + +# argparse runners + + +def run_square_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment)) + MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes)) + data = run(args.dtype, args.sweep_schedules, MKNs) + + make_output(data, MKNs, f"square_bench-{args.dtype}") + + +def run_range_bench(args): + m_start, k_start, n_start = (int(x) for x in args.dim_start.split(",")) + m_end, k_end, n_end = (int(x) for x in args.dim_end.split(",")) + m_increment, k_increment, n_increment = ( + int(x) for x in args.dim_increment.split(",") + ) + Ms = list(range(m_start, m_end + 1, m_increment)) + Ks = list(range(k_start, k_end + 1, k_increment)) + Ns = list(range(n_start, n_end + 1, n_increment)) + MKNs = list(product(Ms, Ks, Ns)) + + data = run(args.dtype, args.sweep_schedules, MKNs) + + make_output(data, MKNs, f"range_bench-{args.dtype}") + + +def run_model_bench(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]: + KNs = [] + for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]): + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + KNs.append(KN) + return KNs + + model_bench_data = [] + models_tps = list(itertools.product(args.models, args.tp_sizes)) + for model, tp_size in models_tps: + Ms = args.batch_sizes + KNs = model_shapes(model, tp_size) + MKNs = [] + for m in Ms: + for k, n in KNs: + MKNs.append((m, k, n)) + + data = run(args, MKNs) + model_bench_data.append(data) + + type_string = f"{args.act_type}" + + # Print all results + for data, model_tp in zip(model_bench_data, models_tps): + model, tp_size = model_tp + print(f"== Results {type_string} {model}-TP{tp_size} ====") + print_timers(data) + + timestr = time.strftime("%Y%m%d-%H%M%S") + + all_results = [] + for d in model_bench_data: + all_results.extend(d) + + # pickle all data + with open(f"model_bench-{type_string}-{timestr}.pkl", "wb") as f: + args_dict = vars(args) + args_dict.pop("func") + pkl.dump( + { + "args": args_dict, + "results": all_results, + }, + f, + ) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + return { + "bfloat16": torch.bfloat16, + "float16": torch.float16, + "int8": torch.int8, + "float8_e4m3fn": torch.float8_e4m3fn, + "int": torch.int, + "float": torch.float, + }[dt] + + class ToTorchDtype(argparse.Action): + def __call__(self, parser, namespace, values, option_string=None): + setattr(namespace, self.dest, to_torch_dtype(values)) + + parser = FlexibleArgumentParser( + description=""" +Benchmark Machete GEMM. + + To run square GEMMs: + python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 square_bench --dim-start 128 --dim-end 512 --dim-increment 64 + + To run constant N and K and sweep M: + python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384 + + To run dimensions from a model: + python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1 + + Output: + - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs. + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + parser.add_argument( + "--act-type", + action=ToTorchDtype, + required=True, + choices=["bfloat16", "float16", "int8", "float8_e4m3fn"], + ) + parser.add_argument( + "--group-scale-type", + action=ToTorchDtype, + choices=["bfloat16", "float16"], + ) + parser.add_argument( + "--group-zero-type", + type=to_torch_dtype, + choices=["bfloat16", "float16"], + ) + parser.add_argument( + "--channel-scale-type", + action=ToTorchDtype, + choices=["float"], + ) + parser.add_argument( + "--token-scale-type", + action=ToTorchDtype, + choices=["float"], + ) + parser.add_argument( + "--out-type", + action=ToTorchDtype, + choices=["bfloat16", "float16"], + ) + parser.add_argument( + "--group-size", + type=int, + help="Available options are ['None', '-1', '128'], default=128", + default=128, + ) + parser.add_argument( + "--sweep-schedules", + action="store_true", + help="Run a sweep over all supported schedules", + ) + parser.add_argument( + "--sweep-csv-out", + help="CSV to store sweep results", + default="sch_sweep_results.csv", + ) + subparsers = parser.add_subparsers(dest="cmd", required=True) + + square_parser = subparsers.add_parser("square_bench") + square_parser.add_argument("--dim-start", type=int, required=True) + square_parser.add_argument("--dim-end", type=int, required=True) + square_parser.add_argument("--dim-increment", type=int, required=True) + square_parser.set_defaults(func=run_square_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument( + "--dim-start", + type=str, + required=True, + help="Start value for M,K,N as common separated list", + ) + range_parser.add_argument( + "--dim-end", + type=str, + required=True, + help="End value (inclusive) for M,K,N as common separated list", + ) + range_parser.add_argument( + "--dim-increment", + type=str, + required=True, + help="Increment value for M,K,N as common separated list", + ) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + + _SWEEP_SCHEDULES_RESULTS_CSV = args.sweep_csv_out + args.func(args) + + if _SWEEP_SCHEDULES_RESULTS is not None: + _SWEEP_SCHEDULES_RESULTS.to_csv(_SWEEP_SCHEDULES_RESULTS_CSV) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_marlin.py b/online_apiserver_test/benchmarks/kernels/benchmark_marlin.py new file mode 100644 index 0000000000000000000000000000000000000000..34cc45e94d76d930013c9d60ca8c5f78dd0b7254 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_marlin.py @@ -0,0 +1,413 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import torch +import torch.utils.benchmark as benchmark +from benchmark_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.gptq_marlin_24 import ( + GPTQ_MARLIN_24_MAX_PARALLEL, + GPTQ_MARLIN_24_MIN_THREAD_N, + GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, + GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES, +) +from vllm.model_executor.layers.quantization.utils.allspark_utils import ( + ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD, + ALLSPARK_SUPPORTED_QUANT_TYPES, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils import ( + GPTQ_MARLIN_MAX_PARALLEL, + GPTQ_MARLIN_MIN_THREAD_N, + MARLIN_SUPPORTED_GROUP_SIZES, + query_marlin_supported_quant_types, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import ( + FP4_MARLIN_SUPPORTED_GROUP_SIZES, + rand_marlin_weight_fp4_like, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import ( + marlin_quant_fp8_torch, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_test import ( + MarlinWorkspace, + awq_marlin_quantize, + marlin_quantize, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import ( + marlin_24_quantize, +) +from vllm.model_executor.layers.quantization.utils.quant_utils import ( + gptq_pack, + gptq_quantize_weights, + quantize_weights, + sort_weights, +) +from vllm.scalar_type import ScalarType, scalar_types +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"] +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192] + +ACT_ORDER_OPTS = [False, True] +K_FULL_OPTS = [False, True] + + +def bench_run( + results: list[benchmark.Measurement], + model: str, + act_order: bool, + is_k_full: bool, + quant_type: ScalarType, + group_size: int, + size_m: int, + size_k: int, + size_n: int, +): + label = "Quant Matmul" + sub_label = "{}, act={} k_full={}, q={}, g={}, MKN=({}x{}x{})".format( + model, act_order, is_k_full, str(quant_type), group_size, size_m, size_k, size_n + ) + print(f"Testing: {sub_label}") + + a = torch.randn(size_m, size_k).to(torch.half).cuda() + b = torch.rand(size_k, size_n).to(torch.half).cuda() + has_zp = quant_type in [scalar_types.uint4, scalar_types.uint8] + if act_order and (group_size == -1 or group_size == size_k or has_zp): + return + if size_k % group_size != 0: + return + + marlin_24_supported = ( + quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES + and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES + ) + repack_supported = ( + quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES + and group_size in MARLIN_SUPPORTED_GROUP_SIZES + ) + allspark_supported = ( + quant_type in ALLSPARK_SUPPORTED_QUANT_TYPES + and group_size == -1 + and not act_order + and is_k_full + ) + + def gen_marlin_params(): + # Marlin quant + marlin_g_idx = marlin_sort_indices = marlin_zp = marlin_s2 = None + if quant_type == scalar_types.float4_e2m1f: + if group_size != 16 or act_order: + return + marlin_w_ref, marlin_q_w, marlin_s, marlin_s2 = rand_marlin_weight_fp4_like( + b.T, group_size + ) + elif quant_type == scalar_types.float8_e4m3fn: + if group_size not in [-1, 128] or act_order: + return + marlin_w_ref, marlin_q_w, marlin_s = marlin_quant_fp8_torch(b.T, group_size) + elif group_size == 16: + return + elif has_zp: + marlin_w_ref, marlin_q_w, marlin_s, marlin_zp = awq_marlin_quantize( + b, quant_type, group_size + ) + else: + marlin_w_ref, marlin_q_w, marlin_s, marlin_g_idx, marlin_sort_indices, _ = ( + marlin_quantize(b, quant_type, group_size, act_order) + ) + return ( + marlin_w_ref, + marlin_q_w, + marlin_s, + marlin_s2, + marlin_zp, + marlin_g_idx, + marlin_sort_indices, + ) + + def gen_marlin_24_params(): + marlin_24_w_ref = marlin_24_q_w_comp = marlin_24_meta = marlin_24_s = None + if marlin_24_supported: + (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) = ( + marlin_24_quantize(b, quant_type, group_size) + ) + return (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) + + def gen_repack_params(): + q_w_gptq = None + repack_sort_indices = None + if repack_supported: + (w_ref, q_w, s, g_idx, rand_perm) = gptq_quantize_weights( + b, quant_type, group_size, act_order + ) + q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n) + + # For act_order, sort the "weights" and "g_idx" + # so that group ids are increasing + repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device) + if act_order: + (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx) + return q_w_gptq, repack_sort_indices + + def gen_allspark_params(): + qw_reorder = s_reorder = zp_reorder = sm_count = sm_version = ( + CUBLAS_M_THRESHOLD + ) = None + nonlocal allspark_supported + if allspark_supported: + properties = torch.cuda.get_device_properties(b.device.index) + sm_count = properties.multi_processor_count + sm_version = properties.major * 10 + properties.minor + + supported_arch = sm_version >= 80 and sm_version < 90 + allspark_supported = allspark_supported and supported_arch + if supported_arch: + w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size, has_zp) + qw = qw.to(torch.uint8) + + qw_reorder, s_reorder, zp_reorder = ops.allspark_repack_weight( + qw, s, zp, has_zp + ) + CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD + return ( + qw_reorder, + s_reorder, + zp_reorder, + sm_count, + sm_version, + CUBLAS_M_THRESHOLD, + ) + + ( + marlin_w_ref, + marlin_q_w, + marlin_s, + marlin_s2, + marlin_zp, + marlin_g_idx, + marlin_sort_indices, + ) = gen_marlin_params() + marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s = ( + gen_marlin_24_params() + ) + q_w_gptq, repack_sort_indices = gen_repack_params() + qw_reorder, s_reorder, zp_reorder, sm_count, sm_version, CUBLAS_M_THRESHOLD = ( + gen_allspark_params() + ) + + # Prepare + marlin_workspace = MarlinWorkspace( + size_n, GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL + ) + marlin_24_workspace = MarlinWorkspace( + size_n, GPTQ_MARLIN_24_MIN_THREAD_N, GPTQ_MARLIN_24_MAX_PARALLEL + ) + + globals = { + # Gen params + "quant_type": quant_type, + "group_size": group_size, + "size_m": size_m, + "size_n": size_n, + "size_k": size_k, + "a": a, + # Marlin params + "marlin_w_ref": marlin_w_ref, + "marlin_q_w": marlin_q_w, + "marlin_s": marlin_s, + "marlin_s2": marlin_s2, + "marlin_zp": marlin_zp, + "marlin_g_idx": marlin_g_idx, + "marlin_sort_indices": marlin_sort_indices, + "marlin_workspace": marlin_workspace, + "is_k_full": is_k_full, + # Marlin_24 params + "marlin_24_w_ref": marlin_24_w_ref, + "marlin_24_q_w_comp": marlin_24_q_w_comp, + "marlin_24_meta": marlin_24_meta, + "marlin_24_s": marlin_24_s, + "marlin_24_workspace": marlin_24_workspace, + # GPTQ params + "q_w_gptq": q_w_gptq, + "repack_sort_indices": repack_sort_indices, + # AllSpark W8A16 params + "qw_reorder": qw_reorder, + "s_reorder": s_reorder, + "zp_reorder": zp_reorder, + "sm_count": sm_count, + "sm_version": sm_version, + "CUBLAS_M_THRESHOLD": CUBLAS_M_THRESHOLD, + # Kernels + "gptq_marlin_gemm": ops.gptq_marlin_gemm, + "gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm, + "gptq_marlin_repack": ops.gptq_marlin_repack, + "allspark_w8a16_gemm": ops.allspark_w8a16_gemm, + } + + min_run_time = 1 + + # Warmup pytorch + for _ in range(5): + torch.matmul(a, marlin_w_ref) + + results.append( + benchmark.Timer( + stmt="torch.matmul(a, marlin_w_ref)", + globals=globals, + label=label, + sub_label=sub_label, + description="pytorch_gemm", + ).blocked_autorange(min_run_time=min_run_time) + ) + + results.append( + benchmark.Timer( + stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False, False)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_gemm", + ).blocked_autorange(min_run_time=min_run_time) + ) + + results.append( + benchmark.Timer( + stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True, False)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_gemm_fp32", + ).blocked_autorange(min_run_time=min_run_time) + ) + + if marlin_24_supported: + results.append( + benchmark.Timer( + stmt="output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, quant_type, size_m, size_n, size_k)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_24_gemm", + ).blocked_autorange(min_run_time=min_run_time) + ) + + if repack_supported: + results.append( + benchmark.Timer( + stmt="q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_repack", + ).blocked_autorange(min_run_time=min_run_time) + ) + + if allspark_supported: + results.append( + benchmark.Timer( + stmt="output = allspark_w8a16_gemm(a, qw_reorder, s_reorder, zp_reorder, size_n, group_size, sm_count, sm_version, CUBLAS_M_THRESHOLD, False, True)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="allspark_w8a16_gemm_fp32", + ).blocked_autorange(min_run_time=min_run_time) + ) + + +def main(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + results: list[benchmark.Measurement] = [] + + for model in args.models: + for layer in WEIGHT_SHAPES[model]: + size_k = layer[0] + size_n = layer[1] + + if len(args.limit_k) > 0 and size_k not in args.limit_k: + continue + + if len(args.limit_n) > 0 and size_n not in args.limit_n: + continue + + for act_order in ACT_ORDER_OPTS: + if ( + len(args.limit_act_order) > 0 + and act_order not in args.limit_act_order + ): + continue + + for is_k_full in K_FULL_OPTS: + if ( + len(args.limit_k_full) > 0 + and is_k_full not in args.limit_k_full + ): + continue + + for quant_type in query_marlin_supported_quant_types(): + if ( + len(args.limit_num_bits) > 0 + and quant_type.size_bits not in args.limit_num_bits + ): + continue + + for group_size in ( + MARLIN_SUPPORTED_GROUP_SIZES + + FP4_MARLIN_SUPPORTED_GROUP_SIZES + ): + if ( + len(args.limit_group_size) > 0 + and group_size not in args.limit_group_size + ): + continue + + # For act_order, the group_size must be less than + # size_k + if act_order and (group_size == size_k or group_size == -1): + continue + + for size_m in args.batch_sizes: + bench_run( + results, + model, + act_order, + is_k_full, + quant_type, + group_size, + size_m, + size_k, + size_n, + ) + + compare = benchmark.Compare(results) + compare.print() + + +# For quick benchmarking use: +# python benchmark_marlin.py --batch-sizes 1 16 32 --limit-k 4096 --limit-n 4096 --limit-group-size 128 --limit-num-bits 4 --limit-act-order 0 --limit-k-full 1 # noqa E501 +# +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark Marlin across specified models/shapes/batches" + ) + parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + parser.add_argument("--limit-k", nargs="+", type=int, default=[]) + parser.add_argument("--limit-n", nargs="+", type=int, default=[]) + parser.add_argument("--limit-group-size", nargs="+", type=int, default=[]) + parser.add_argument("--limit-num-bits", nargs="+", type=int, default=[]) + parser.add_argument("--limit-act-order", nargs="+", type=int, default=[]) + parser.add_argument("--limit-k-full", nargs="+", type=int, default=[]) + + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_moe.py b/online_apiserver_test/benchmarks/kernels/benchmark_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..07af58d81c68331597f1e63c0c62dc814b41025c --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_moe.py @@ -0,0 +1,737 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import json +import time +from contextlib import nullcontext +from datetime import datetime +from itertools import product +from typing import Any, TypedDict + +import ray +import torch +from ray.experimental.tqdm_ray import tqdm + +from vllm.model_executor.layers.fused_moe.fused_moe import * +from vllm.platforms import current_platform +from vllm.transformers_utils.config import get_config +from vllm.triton_utils import triton +from vllm.utils import FlexibleArgumentParser + +FP8_DTYPE = current_platform.fp8_dtype() + + +class BenchmarkConfig(TypedDict): + BLOCK_SIZE_M: int + BLOCK_SIZE_N: int + BLOCK_SIZE_K: int + GROUP_SIZE_M: int + num_warps: int + num_stages: int + + +def benchmark_config( + config: BenchmarkConfig, + num_tokens: int, + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + num_iters: int = 100, + block_quant_shape: list[int] = None, + use_deep_gemm: bool = False, +) -> float: + init_dtype = torch.float16 if use_fp8_w8a8 else dtype + x = torch.randn(num_tokens, hidden_size, dtype=dtype) + if use_int8_w8a16: + w1 = torch.randint( + -127, + 127, + ( + num_experts, + shard_intermediate_size, + hidden_size, + ), + dtype=torch.int8, + ) + w2 = torch.randint( + -127, + 127, + ( + num_experts, + hidden_size, + shard_intermediate_size // 2, + ), + dtype=torch.int8, + ) + else: + w1 = torch.randn( + num_experts, shard_intermediate_size, hidden_size, dtype=init_dtype + ) + w2 = torch.randn( + num_experts, hidden_size, shard_intermediate_size // 2, dtype=init_dtype + ) + gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32) + + w1_scale = None + w2_scale = None + a1_scale = None + a2_scale = None + if use_int8_w8a16: + w1_scale = torch.randn( + (num_experts, 2 * shard_intermediate_size), dtype=torch.float32 + ) + w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32) + if use_fp8_w8a8: + if block_quant_shape: + block_n, block_k = block_quant_shape[0], block_quant_shape[1] + E = num_experts + N = shard_intermediate_size // 2 + K = hidden_size + factor_for_scale = 1e-2 + n_tiles_w1 = (2 * N + block_n - 1) // block_n + n_tiles_w2 = (K + block_n - 1) // block_n + k_tiles_w1 = (K + block_k - 1) // block_k + k_tiles_w2 = (N + block_k - 1) // block_k + w1_scale = ( + torch.rand((E, n_tiles_w1, k_tiles_w1), dtype=torch.float32) + * factor_for_scale + ) + w2_scale = ( + torch.rand((E, n_tiles_w2, k_tiles_w2), dtype=torch.float32) + * factor_for_scale + ) + else: + w1_scale = torch.randn(num_experts, dtype=torch.float32) + w2_scale = torch.randn(num_experts, dtype=torch.float32) + + a1_scale = torch.randn(1, dtype=torch.float32) + a2_scale = torch.randn(1, dtype=torch.float32) + + w1 = w1.to(FP8_DTYPE) + w2 = w2.to(FP8_DTYPE) + + input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32) + + def prepare(i: int): + input_gating.copy_(gating_output[i]) + + def run(): + from vllm.model_executor.layers.fused_moe import override_config + + with override_config(config): + if use_deep_gemm: + topk_weights, topk_ids, token_expert_indices = fused_topk( + x, input_gating, topk, False + ) + return fused_experts( + x, + w1, + w2, + topk_weights, + topk_ids, + inplace=True, + use_fp8_w8a8=use_fp8_w8a8, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_quant_shape, + allow_deep_gemm=True, + ) + else: + fused_moe( + x, + w1, + w2, + input_gating, + topk, + renormalize=True, + inplace=True, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_quant_shape, + ) + + # JIT compilation & warmup + run() + torch.cuda.synchronize() + + # Capture 10 invocations with CUDA graph + graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(graph): + for _ in range(10): + run() + torch.cuda.synchronize() + + # Warmup + for _ in range(5): + graph.replay() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + prepare(i) + torch.cuda.synchronize() + + start_event.record() + graph.replay() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + graph.reset() + return avg + + +def get_rocm_tuning_space(use_fp16): + block_mn_range = [16, 32, 64, 128, 256] + block_k_range = [16, 32, 64, 128, 256] + if not use_fp16: + block_k_range.remove(16) # BLOCK_K=16 not supported for fp8 + num_warps_range = [1, 2, 4, 8] + group_m_range = [1, 4, 8, 16, 32] + num_stage_range = [2] + waves_per_eu_range = [0] + matrix_instr_nonkdim_range = [16, 32] if use_fp16 else [] + kpack_range = [1, 2] if use_fp16 else [] + + param_ranges = { + "BLOCK_SIZE_M": block_mn_range, + "BLOCK_SIZE_N": block_mn_range, + "BLOCK_SIZE_K": block_k_range, + "GROUP_SIZE_M": group_m_range, + "num_warps": num_warps_range, + "num_stages": num_stage_range, + "waves_per_eu": waves_per_eu_range, + } + if use_fp16: + param_ranges["matrix_instr_nonkdim"] = matrix_instr_nonkdim_range + param_ranges["kpack"] = kpack_range + + return param_ranges + + +def get_configs_compute_bound(use_fp16, block_quant_shape) -> list[dict[str, int]]: + configs: list[BenchmarkConfig] = [] + + if current_platform.is_rocm(): + param_ranges = get_rocm_tuning_space(use_fp16) + else: + # Reduced search space for faster tuning. + # TODO(woosuk): Increase the search space and use a performance model to + # prune the search space. + block_m_range = [16, 32, 64, 128, 256] + block_n_range = [32, 64, 128, 256] + block_k_range = [64, 128, 256] + num_warps_range = [4, 8] + group_m_range = [1, 16, 32, 64] + num_stage_range = [2, 3, 4, 5] + + param_ranges = { + "BLOCK_SIZE_M": block_m_range, + "BLOCK_SIZE_N": block_n_range, + "BLOCK_SIZE_K": block_k_range, + "GROUP_SIZE_M": group_m_range, + "num_warps": num_warps_range, + "num_stages": num_stage_range, + } + + keys, values = zip(*param_ranges.items()) + for config_values in product(*values): + config = dict(zip(keys, config_values)) + configs.append(config) + + # Remove configs that are not compatible with fp8 block quantization + # BLOCK_SIZE_K must be a multiple of block_k + # BLOCK_SIZE_N must be a multiple of block_n + if block_quant_shape is not None and not use_fp16: + block_n, block_k = block_quant_shape[0], block_quant_shape[1] + for config in configs[:]: + if ( + config["BLOCK_SIZE_K"] % block_k != 0 + or config["BLOCK_SIZE_N"] % block_n != 0 + ): + configs.remove(config) + return configs + + +def prune_rocm_search_space( + num_tokens, shard_intermediate_size, hidden_size, search_space, is_fp16, topk +): + N1, K1 = shard_intermediate_size, hidden_size + N2, K2 = hidden_size, shard_intermediate_size // 2 + pruned_space_1 = prune_rocm_configs( + num_tokens * topk, N1, K1, search_space, is_fp16 + ) + pruned_space_2 = prune_rocm_configs( + num_tokens * topk, N2, K2, search_space, is_fp16 + ) + search_space = merge_unique_dicts(pruned_space_1, pruned_space_2) + return search_space + + +# The following code is inspired by ROCm/Triton GEMM tuning script: +# https://github.com/ROCm/triton/blob/triton-mlir/scripts/amd/gemm/tune_gemm.py#L89 +def prune_rocm_configs(M, N, K, configs, is_fp16=True): + pruned_configs = [] + elemBytes_a = 2 if is_fp16 else 1 + elemBytes_b = 2 if is_fp16 else 1 + + mfma = 16 if M < 32 or N < 32 else 32 + + # TODO (zhanglx): figure out the boundary between large and small gemms + large_gemm = False + if M >= 2048 and N >= 2048: + large_gemm = True + + for config in configs: + BLOCK_SIZE_M = config.get("BLOCK_SIZE_M") + BLOCK_SIZE_N = config.get("BLOCK_SIZE_N") + BLOCK_SIZE_K = config.get("BLOCK_SIZE_K") + num_warps = config.get("num_warps") + + if is_fp16: + matrix_instr_nonkdim = config.get("matrix_instr_nonkdim") + if matrix_instr_nonkdim > mfma: + continue + if mfma == 4 and BLOCK_SIZE_K < 64: + continue + # some layouts could not work properly in case + # number elements per thread is less 1 + if BLOCK_SIZE_M * BLOCK_SIZE_N < 64: + continue + SPLIT_K = config.get("SPLIT_K", 1) + GROUP_M = config.get("GROUP_SIZE_M") + if is_fp16: + if ( + matrix_instr_nonkdim > BLOCK_SIZE_M + or matrix_instr_nonkdim > BLOCK_SIZE_N + ): + continue + if matrix_instr_nonkdim >= M and matrix_instr_nonkdim != BLOCK_SIZE_M: + continue + if matrix_instr_nonkdim >= N and matrix_instr_nonkdim != BLOCK_SIZE_N: + continue + # Skip BLOCK_SIZE that is too large compare to M/N + # unless BLOCK_SIZE is already small enough + if M * 2 < BLOCK_SIZE_M and BLOCK_SIZE_M != 16: + continue + if N * 2 < BLOCK_SIZE_N and BLOCK_SIZE_N != 16: + continue + # skip large split_k when not necessary + if SPLIT_K != 1 and not need_split_k(M, N, K): + continue + # skip split_k that leads to EVEN_K = false + leap = SPLIT_K * BLOCK_SIZE_K + modv = K % leap + if modv != 0: + continue + # skip large GROUP_M + if GROUP_M * BLOCK_SIZE_M > M and GROUP_M != 1: + continue + # out of shared memory resource + # TODO (zhanglx): This does not consider the LDS usage in the epilogue + LDS = ( + BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a + + BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b + ) + if LDS > 65536: + continue + # Skip small block sizes and num_warps for large gemm + # For fp16 and f8, we want to only use BLOCK_SIZE >= 64 + if large_gemm: + if BLOCK_SIZE_M < 64 or BLOCK_SIZE_N < 64: + continue + if BLOCK_SIZE_K < 64: + continue + if num_warps < 4: + continue + + pruned_configs.append(config) + + return pruned_configs + + +def need_split_k(SIZE_M, SIZE_N, SIZE_K): + return (SIZE_M < 64 or SIZE_N < 64) and SIZE_K > 1024 + + +def merge_unique_dicts(list1, list2): + result = [] + combined_list = list1.copy() + combined_list.extend(list2) + for dictionary in combined_list: + if dictionary not in result: + result.append(dictionary) + return result + + +@ray.remote(num_gpus=1) +class BenchmarkWorker: + def __init__(self, seed: int) -> None: + torch.set_default_device("cuda") + current_platform.seed_everything(seed) + self.seed = seed + # Get the device ID to allocate tensors and kernels + # on the respective GPU. This is required for Ray to work + # correctly with multi-GPU tuning on the ROCm platform. + self.device_id = int(ray.get_gpu_ids()[0]) + + def benchmark( + self, + num_tokens: int, + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + block_quant_shape: list[int] = None, + use_deep_gemm: bool = False, + ) -> tuple[dict[str, int], float]: + current_platform.seed_everything(self.seed) + dtype_str = get_config_dtype_str( + dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8 + ) + # NOTE(woosuk): The current naming convention uses w2.shape[2], which + # is the intermediate size after silu_and_mul. + op_config = get_moe_configs( + num_experts, shard_intermediate_size // 2, dtype_str + ) + if op_config is None: + config = get_default_config( + num_tokens, + num_experts, + shard_intermediate_size, + hidden_size, + topk, + dtype_str, + is_marlin=False, + ) + else: + config = op_config[min(op_config.keys(), key=lambda x: abs(x - num_tokens))] + kernel_time = benchmark_config( + config, + num_tokens, + num_experts, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=100, + block_quant_shape=block_quant_shape, + use_deep_gemm=use_deep_gemm, + ) + return config, kernel_time + + def tune( + self, + num_tokens: int, + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + search_space: list[dict[str, int]], + block_quant_shape: list[int], + use_deep_gemm: bool, + ) -> dict[str, int]: + best_config = None + best_time = float("inf") + if current_platform.is_rocm(): + is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16) + search_space = prune_rocm_search_space( + num_tokens, + shard_intermediate_size, + hidden_size, + search_space, + is_fp16, + topk, + ) + + need_device_guard = False + if current_platform.is_rocm(): + visible_device = os.environ.get("ROCR_VISIBLE_DEVICES", None) + if visible_device != f"{self.device_id}": + need_device_guard = True + + with torch.cuda.device(self.device_id) if need_device_guard else nullcontext(): + for config in tqdm(search_space): + try: + kernel_time = benchmark_config( + config, + num_tokens, + num_experts, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=20, + block_quant_shape=block_quant_shape, + use_deep_gemm=use_deep_gemm, + ) + except triton.runtime.autotuner.OutOfResources: + # Some configurations may be invalid and fail to compile. + continue + + if kernel_time < best_time: + best_time = kernel_time + best_config = config + now = datetime.now() + print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}") + assert best_config is not None + return best_config + + +def sort_config(config: BenchmarkConfig) -> BenchmarkConfig: + return { + "BLOCK_SIZE_M": config["BLOCK_SIZE_M"], + "BLOCK_SIZE_N": config["BLOCK_SIZE_N"], + "BLOCK_SIZE_K": config["BLOCK_SIZE_K"], + "GROUP_SIZE_M": config["GROUP_SIZE_M"], + "num_warps": config["num_warps"], + "num_stages": config["num_stages"], + **( + {"waves_per_eu": config["waves_per_eu"]} if "waves_per_eu" in config else {} + ), + **( + {"matrix_instr_nonkdim": config["matrix_instr_nonkdim"]} + if "matrix_instr_nonkdim" in config + else {} + ), + **({"kpack": config["kpack"]} if "kpack" in config else {}), + } + + +def save_configs( + configs: dict[int, BenchmarkConfig], + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + block_quant_shape: list[int], +) -> None: + dtype_str = get_config_dtype_str( + dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8 + ) + + # NOTE(woosuk): The current naming convention uses w2.shape[2], which + # is the intermediate size after silu_and_mul. + filename = get_config_file_name( + num_experts, shard_intermediate_size // 2, dtype_str, block_quant_shape + ) + + print(f"Writing best config to {filename}...") + with open(filename, "w") as f: + json.dump(configs, f, indent=4) + f.write("\n") + + +def get_weight_block_size_safety(config, default_value=None): + quantization_config = getattr(config, "quantization_config", {}) + if isinstance(quantization_config, dict): + return quantization_config.get("weight_block_size", default_value) + return default_value + + +def main(args: argparse.Namespace): + print(args) + + config = get_config(model=args.model, trust_remote_code=args.trust_remote_code) + if args.model_prefix: + config = getattr(config, args.model_prefix) + + if config.architectures[0] == "DbrxForCausalLM": + E = config.ffn_config.moe_num_experts + topk = config.ffn_config.moe_top_k + intermediate_size = config.ffn_config.ffn_hidden_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + elif config.architectures[0] == "JambaForCausalLM": + E = config.num_experts + topk = config.num_experts_per_tok + intermediate_size = config.intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + elif config.architectures[0] in ("DeepseekV3ForCausalLM", "DeepseekV2ForCausalLM"): + E = config.n_routed_experts + topk = config.num_experts_per_tok + intermediate_size = config.moe_intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + elif config.architectures[0] in ("Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"): + E = config.num_experts + topk = config.num_experts_per_tok + intermediate_size = config.moe_intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + else: + # Support for llama4 + config = config.get_text_config() + # Default: Mixtral. + E = config.num_local_experts + topk = config.num_experts_per_tok + intermediate_size = config.intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + + hidden_size = config.hidden_size + dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype + use_fp8_w8a8 = args.dtype == "fp8_w8a8" + use_int8_w8a16 = args.dtype == "int8_w8a16" + block_quant_shape = get_weight_block_size_safety(config) + + if args.batch_size is None: + batch_sizes = [ + 1, + 2, + 4, + 8, + 16, + 24, + 32, + 48, + 64, + 96, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ] + else: + batch_sizes = args.batch_size + + use_deep_gemm = bool(args.use_deep_gemm) + + if current_platform.is_rocm() and "HIP_VISIBLE_DEVICES" in os.environ: + # Ray will set ROCR_VISIBLE_DEVICES for device visibility + logger.warning( + "Ray uses ROCR_VISIBLE_DEVICES to control device accessibility." + "Replacing HIP_VISIBLE_DEVICES with ROCR_VISIBLE_DEVICES." + ) + val = os.environ["HIP_VISIBLE_DEVICES"] + os.environ["ROCR_VISIBLE_DEVICES"] = val + del os.environ["HIP_VISIBLE_DEVICES"] + + ray.init() + num_gpus = int(ray.available_resources()["GPU"]) + workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)] + + def _distribute(method: str, inputs: list[Any]) -> list[Any]: + outputs = [] + worker_idx = 0 + for input_args in inputs: + worker = workers[worker_idx] + worker_method = getattr(worker, method) + output = worker_method.remote(*input_args) + outputs.append(output) + worker_idx = (worker_idx + 1) % num_gpus + return ray.get(outputs) + + if args.tune: + is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16) + search_space = get_configs_compute_bound(is_fp16, block_quant_shape) + print(f"Start tuning over {len(search_space)} configurations...") + + start = time.time() + configs = _distribute( + "tune", + [ + ( + batch_size, + E, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + search_space, + block_quant_shape, + use_deep_gemm, + ) + for batch_size in batch_sizes + ], + ) + best_configs = { + M: sort_config(config) for M, config in zip(batch_sizes, configs) + } + save_configs( + best_configs, + E, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + block_quant_shape, + ) + end = time.time() + print(f"Tuning took {end - start:.2f} seconds") + else: + outputs = _distribute( + "benchmark", + [ + ( + batch_size, + E, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + block_quant_shape, + use_deep_gemm, + ) + for batch_size in batch_sizes + ], + ) + + for batch_size, (config, kernel_time) in zip(batch_sizes, outputs): + print(f"Batch size: {batch_size}, config: {config}") + print(f"Kernel time: {kernel_time:.2f} us") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser() + parser.add_argument( + "--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1" + ) + parser.add_argument( + "--tp-size", "-tp", "--tensor-parallel-size", type=int, default=2 + ) + parser.add_argument( + "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto" + ) + parser.add_argument("--use-deep-gemm", action="store_true") + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--batch-size", type=int, nargs="+", required=False) + parser.add_argument("--tune", action="store_true") + parser.add_argument("--trust-remote-code", action="store_true") + parser.add_argument("--model-prefix", type=str, required=False) + args = parser.parse_args() + + main(args) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_moe_align_block_size.py b/online_apiserver_test/benchmarks/kernels/benchmark_moe_align_block_size.py new file mode 100644 index 0000000000000000000000000000000000000000..5170ac09dc42af8fac0f15e48dc5d348b728f04a --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_moe_align_block_size.py @@ -0,0 +1,159 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import argparse +import itertools + +import torch + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.fused_moe.moe_align_block_size import ( + moe_align_block_size_triton, +) +from vllm.triton_utils import triton + + +def get_topk_ids(num_tokens: int, num_experts: int, topk: int) -> torch.Tensor: + return torch.stack( + [ + torch.randperm(num_experts, dtype=torch.int32, device="cuda")[:topk] + for _ in range(num_tokens) + ] + ) + + +def check_correctness(num_tokens, num_experts=256, block_size=256, topk=8): + """ + Verifies vllm vs. Triton + """ + topk_ids = get_topk_ids(num_tokens, num_experts, topk) + + # 1. malloc space for triton and vllm + # malloc enough space (max_num_tokens_padded) for the sorted ids + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + sorted_ids_triton = torch.empty( + (max_num_tokens_padded,), dtype=torch.int32, device="cuda" + ) + sorted_ids_triton.fill_(topk_ids.numel()) # fill with sentinel value + expert_ids_triton = torch.zeros( + (max_num_tokens_padded // block_size,), dtype=torch.int32, device="cuda" + ) + num_tokens_post_pad_triton = torch.empty((1,), dtype=torch.int32, device="cuda") + + sorted_ids_vllm = torch.empty_like(sorted_ids_triton) + sorted_ids_vllm.fill_(topk_ids.numel()) + expert_ids_vllm = torch.zeros_like(expert_ids_triton) + num_tokens_post_pad_vllm = torch.empty_like(num_tokens_post_pad_triton) + + # 2. run implementations + moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids_triton, + expert_ids_triton, + num_tokens_post_pad_triton, + ) + + ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids_vllm, + expert_ids_vllm, + num_tokens_post_pad_vllm, + ) + print(f"✅ VLLM implementation works with {num_experts} experts!") + + # 3. compare results + if torch.allclose(expert_ids_triton, expert_ids_vllm) and torch.allclose( + num_tokens_post_pad_triton, num_tokens_post_pad_vllm + ): + print("✅ Triton and VLLM implementations match.") + else: + print("❌ Triton and VLLM implementations DO NOT match.") + print("Triton expert_ids:", expert_ids_triton) + print("VLLM expert_ids:", expert_ids_vllm) + print("Triton num_tokens_post_pad:", num_tokens_post_pad_triton) + print("VLLM num_tokens_post_pad:", num_tokens_post_pad_vllm) + + +# test configurations +num_tokens_range = [1, 16, 256, 4096] +num_experts_range = [16, 64, 224, 256, 280, 512] +topk_range = [1, 2, 8] +configs = list(itertools.product(num_tokens_range, num_experts_range, topk_range)) + + +@triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["num_tokens", "num_experts", "topk"], + x_vals=configs, + line_arg="provider", + line_vals=["vllm", "triton"], # "triton" + line_names=["VLLM", "Triton"], # "Triton" + plot_name="moe-align-block-size-performance", + args={}, + ) +) +def benchmark(num_tokens, num_experts, topk, provider): + """Benchmark function for Triton.""" + block_size = 256 + topk_ids = get_topk_ids(num_tokens, num_experts, topk) + + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + sorted_ids = torch.empty((max_num_tokens_padded,), dtype=torch.int32, device="cuda") + sorted_ids.fill_(topk_ids.numel()) + max_num_m_blocks = max_num_tokens_padded // block_size + expert_ids = torch.empty((max_num_m_blocks,), dtype=torch.int32, device="cuda") + num_tokens_post_pad = torch.empty((1,), dtype=torch.int32, device="cuda") + + quantiles = [0.5, 0.2, 0.8] + + if provider == "vllm": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids.clone(), + expert_ids.clone(), + num_tokens_post_pad.clone(), + ), + quantiles=quantiles, + ) + elif provider == "triton": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids.clone(), + expert_ids.clone(), + num_tokens_post_pad.clone(), + ), + quantiles=quantiles, + ) + + return 1000 * ms, 1000 * max_ms, 1000 * min_ms + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--num_experts", + type=int, + default=64, + choices=[8, 16, 32, 64, 128, 256], + ) + parser.add_argument( + "--topk", + type=int, + default=8, + choices=[2, 4, 8], + help="Top-k value for correctness check.", + ) + args = parser.parse_args() + + print("Running correctness check...") + check_correctness(num_tokens=1024, num_experts=args.num_experts, topk=args.topk) + benchmark.run(print_data=True, show_plots=True) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_moe_permute_unpermute.py b/online_apiserver_test/benchmarks/kernels/benchmark_moe_permute_unpermute.py new file mode 100644 index 0000000000000000000000000000000000000000..dba1f3943b96c370acf1262273695dbdb4097711 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_moe_permute_unpermute.py @@ -0,0 +1,417 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +from typing import Any, TypedDict + +import ray +import torch +from transformers import AutoConfig + +from vllm.model_executor.layers.fused_moe.deep_gemm_moe import ( + _moe_permute, + _moe_unpermute_and_reduce, +) +from vllm.model_executor.layers.fused_moe.fused_moe import * +from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import * +from vllm.model_executor.layers.fused_moe.utils import _fp8_quantize +from vllm.platforms import current_platform +from vllm.utils import FlexibleArgumentParser + +FP8_DTYPE = current_platform.fp8_dtype() + + +class BenchmarkConfig(TypedDict): + BLOCK_SIZE_M: int + BLOCK_SIZE_N: int + BLOCK_SIZE_K: int + GROUP_SIZE_M: int + num_warps: int + num_stages: int + + +def benchmark_permute( + num_tokens: int, + num_experts: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + num_iters: int = 100, + use_customized_permute: bool = False, +) -> float: + # init_dtype = torch.float16 if use_fp8_w8a8 else dtype + hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype) + # output_hidden_states = torch.empty_like(hidden_states) + if use_fp8_w8a8: + align_block_size = 128 # deepgemm needs 128 m aligned block + qhidden_states, scale = _fp8_quantize(hidden_states, None, None) + else: + align_block_size = None + qhidden_states = hidden_states + + gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32) + + input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32) + topk_weights, topk_ids, token_expert_indices = fused_topk( + qhidden_states, input_gating, topk, False + ) + + def prepare(i: int): + input_gating.copy_(gating_output[i]) + + def run(): + if use_customized_permute: + (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = ( + moe_permute( + qhidden_states, + topk_weights=topk_weights, + topk_ids=topk_ids, + token_expert_indices=token_expert_indices, + topk=topk, + n_expert=num_experts, + n_local_expert=num_experts, + expert_map=None, + align_block_size=align_block_size, + ) + ) + else: + ( + permuted_hidden_states, + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) = _moe_permute( + qhidden_states, None, topk_ids, num_experts, None, align_block_size + ) + + # JIT compilation & warmup + run() + torch.cuda.synchronize() + + # Capture 10 invocations with CUDA graph + graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(graph): + for _ in range(10): + run() + torch.cuda.synchronize() + + # Warmup + for _ in range(5): + graph.replay() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + prepare(i) + torch.cuda.synchronize() + + start_event.record() + graph.replay() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + graph.reset() + return avg + + +def benchmark_unpermute( + num_tokens: int, + num_experts: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + num_iters: int = 100, + use_customized_permute: bool = False, +) -> float: + # init_dtype = torch.float16 if use_fp8_w8a8 else dtype + hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype) + output_hidden_states = torch.empty_like(hidden_states) + if use_fp8_w8a8: + align_block_size = 128 # deepgemm needs 128 m aligned block + qhidden_states, scale = _fp8_quantize(hidden_states, None, None) + else: + align_block_size = None + qhidden_states = hidden_states + + input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32) + + topk_weights, topk_ids, token_expert_indices = fused_topk( + qhidden_states, input_gating, topk, False + ) + + def prepare(): + if use_customized_permute: + (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = ( + moe_permute( + qhidden_states, + topk_weights=topk_weights, + topk_ids=topk_ids, + token_expert_indices=token_expert_indices, + topk=topk, + n_expert=num_experts, + n_local_expert=num_experts, + expert_map=None, + align_block_size=align_block_size, + ) + ) + # convert to fp16/bf16 as gemm output + return ( + permuted_hidden_states.to(dtype), + first_token_off, + inv_perm_idx, + m_indices, + ) + else: + ( + permuted_qhidden_states, + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) = _moe_permute( + qhidden_states, None, topk_ids, num_experts, None, align_block_size + ) + # convert to fp16/bf16 as gemm output + return ( + permuted_qhidden_states.to(dtype), + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) + + def run(input: tuple): + if use_customized_permute: + (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = input + moe_unpermute( + permuted_hidden_states, + topk_weights, + topk_ids, + inv_perm_idx, + first_token_off, + topk, + num_experts, + num_experts, + ) + else: + ( + permuted_hidden_states, + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) = input + _moe_unpermute_and_reduce( + output_hidden_states, permuted_hidden_states, inv_perm, topk_weights + ) + + # JIT compilation & warmup + input = prepare() + run(input) + torch.cuda.synchronize() + + # Capture 10 invocations with CUDA graph + graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(graph): + for _ in range(10): + run(input) + torch.cuda.synchronize() + + # Warmup + for _ in range(5): + graph.replay() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + torch.cuda.synchronize() + start_event.record() + graph.replay() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + graph.reset() + return avg + + +@ray.remote(num_gpus=1) +class BenchmarkWorker: + def __init__(self, seed: int) -> None: + torch.set_default_device("cuda") + current_platform.seed_everything(seed) + self.seed = seed + # Get the device ID to allocate tensors and kernels + # on the respective GPU. This is required for Ray to work + # correctly with multi-GPU tuning on the ROCm platform. + self.device_id = int(ray.get_gpu_ids()[0]) + + def benchmark( + self, + num_tokens: int, + num_experts: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + use_customized_permute: bool = False, + ) -> tuple[dict[str, int], float]: + current_platform.seed_everything(self.seed) + + permute_time = benchmark_permute( + num_tokens, + num_experts, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=100, + use_customized_permute=use_customized_permute, + ) + unpermute_time = benchmark_unpermute( + num_tokens, + num_experts, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=100, + use_customized_permute=use_customized_permute, + ) + return permute_time, unpermute_time + + +def get_weight_block_size_safety(config, default_value=None): + quantization_config = getattr(config, "quantization_config", {}) + if isinstance(quantization_config, dict): + return quantization_config.get("weight_block_size", default_value) + return default_value + + +def main(args: argparse.Namespace): + print(args) + + config = AutoConfig.from_pretrained( + args.model, trust_remote_code=args.trust_remote_code + ) + if config.architectures[0] == "DbrxForCausalLM": + E = config.ffn_config.moe_num_experts + topk = config.ffn_config.moe_top_k + elif config.architectures[0] == "JambaForCausalLM": + E = config.num_experts + topk = config.num_experts_per_tok + elif ( + config.architectures[0] == "DeepseekV3ForCausalLM" + or config.architectures[0] == "DeepseekV2ForCausalLM" + ): + E = config.n_routed_experts + topk = config.num_experts_per_tok + elif config.architectures[0] in ["Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"]: + E = config.num_experts + topk = config.num_experts_per_tok + + else: + # Support for llama4 + config = config.get_text_config() + # Default: Mixtral. + E = config.num_local_experts + topk = config.num_experts_per_tok + + hidden_size = config.hidden_size + dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype + use_fp8_w8a8 = args.dtype == "fp8_w8a8" + use_int8_w8a16 = args.dtype == "int8_w8a16" + use_customized_permute = args.use_customized_permute + + if args.batch_size is None: + batch_sizes = [ + 1, + 2, + 4, + 8, + 16, + 24, + 32, + 48, + 64, + 96, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ] + else: + batch_sizes = [args.batch_size] + + ray.init() + num_gpus = int(ray.available_resources()["GPU"]) + workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)] + + def _distribute(method: str, inputs: list[Any]) -> list[Any]: + outputs = [] + worker_idx = 0 + for input_args in inputs: + worker = workers[worker_idx] + worker_method = getattr(worker, method) + output = worker_method.remote(*input_args) + outputs.append(output) + worker_idx = (worker_idx + 1) % num_gpus + return ray.get(outputs) + + outputs = _distribute( + "benchmark", + [ + ( + batch_size, + E, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + use_customized_permute, + ) + for batch_size in batch_sizes + ], + ) + + for batch_size, (permute, unpermute) in zip(batch_sizes, outputs): + print(f"Batch size: {batch_size}") + print(f"Permute time: {permute:.2f} us") + print(f"Unpermute time: {unpermute:.2f} us") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser() + parser.add_argument( + "--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1" + ) + parser.add_argument( + "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto" + ) + parser.add_argument("--use-customized-permute", action="store_true") + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--batch-size", type=int, required=False) + parser.add_argument("--trust-remote-code", action="store_true") + args = parser.parse_args() + + main(args) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_paged_attention.py b/online_apiserver_test/benchmarks/kernels/benchmark_paged_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..7e0376c18ecc79690ca49e3995258c5e749bbac7 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_paged_attention.py @@ -0,0 +1,251 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import random +import time +from typing import Optional + +import torch + +from vllm import _custom_ops as ops +from vllm.logger import init_logger +from vllm.platforms import current_platform +from vllm.utils import ( + STR_DTYPE_TO_TORCH_DTYPE, + FlexibleArgumentParser, + create_kv_caches_with_random, +) + +logger = init_logger(__name__) + +NUM_BLOCKS = 128 * 1024 +PARTITION_SIZE = 512 +PARTITION_SIZE_ROCM = 256 + + +@torch.inference_mode() +def main( + version: str, + num_seqs: int, + seq_len: int, + num_query_heads: int, + num_kv_heads: int, + head_size: int, + use_alibi: bool, + block_size: int, + dtype: torch.dtype, + seed: int, + do_profile: bool, + device: str = "cuda", + kv_cache_dtype: Optional[str] = None, +) -> None: + current_platform.seed_everything(seed) + + scale = float(1.0 / (head_size**0.5)) + query = torch.empty( + num_seqs, num_query_heads, head_size, dtype=dtype, device=device + ) + query.uniform_(-scale, scale) + + assert num_query_heads % num_kv_heads == 0 + alibi_slopes = None + if use_alibi: + alibi_slopes = torch.randn(num_query_heads, dtype=torch.float, device=device) + + seq_lens = [seq_len for _ in range(num_seqs)] + max_seq_len = max(seq_lens) + seq_lens = torch.tensor(seq_lens, dtype=torch.int, device=device) + + # Create the block tables. + max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size + block_tables_lst: list[list[int]] = [] + for _ in range(num_seqs): + block_table = [ + random.randint(0, NUM_BLOCKS - 1) for _ in range(max_num_blocks_per_seq) + ] + block_tables_lst.append(block_table) + + block_tables = torch.tensor(block_tables_lst, dtype=torch.int, device=device) + + # Create the KV cache. + key_caches, value_caches = create_kv_caches_with_random( + NUM_BLOCKS, + block_size, + 1, + num_kv_heads, + head_size, + kv_cache_dtype, + dtype, + device=device, + ) + key_cache, value_cache = key_caches[0], value_caches[0] + + # Prepare for the paged attention kernel. + output = torch.empty_like(query) + if version == "v2": + if current_platform.is_rocm(): + global PARTITION_SIZE + if not args.custom_paged_attn and not current_platform.is_navi(): + PARTITION_SIZE = 1024 + else: + PARTITION_SIZE = PARTITION_SIZE_ROCM + num_partitions = (max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE + tmp_output = torch.empty( + size=(num_seqs, num_query_heads, num_partitions, head_size), + dtype=output.dtype, + device=output.device, + ) + exp_sums = torch.empty( + size=(num_seqs, num_query_heads, num_partitions), + dtype=torch.float32, + device=output.device, + ) + max_logits = torch.empty_like(exp_sums) + + def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: + torch.cuda.synchronize() + if profile: + torch.cuda.cudart().cudaProfilerStart() + start_time = time.perf_counter() + + # Using default kv_scale + k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device) + + for _ in range(num_iters): + if version == "v1": + ops.paged_attention_v1( + output, + query, + key_cache, + value_cache, + num_kv_heads, + scale, + block_tables, + seq_lens, + block_size, + max_seq_len, + alibi_slopes, + kv_cache_dtype, + k_scale, + v_scale, + ) + elif version == "v2": + if not args.custom_paged_attn: + ops.paged_attention_v2( + output, + exp_sums, + max_logits, + tmp_output, + query, + key_cache, + value_cache, + num_kv_heads, + scale, + block_tables, + seq_lens, + block_size, + max_seq_len, + alibi_slopes, + kv_cache_dtype, + k_scale, + v_scale, + ) + else: + ops.paged_attention_rocm( + output, + exp_sums, + max_logits, + tmp_output, + query, + key_cache, + value_cache, + num_kv_heads, + scale, + block_tables, + seq_lens, + None, + block_size, + max_seq_len, + alibi_slopes, + kv_cache_dtype, + k_scale, + v_scale, + ) + else: + raise ValueError(f"Invalid version: {version}") + torch.cuda.synchronize() + + end_time = time.perf_counter() + if profile: + torch.cuda.cudart().cudaProfilerStop() + return (end_time - start_time) / num_iters + + # Warmup. + print("Warming up...") + run_benchmark = run_cuda_benchmark + run_benchmark(num_iters=3, profile=False) + + # Benchmark. + if do_profile: + latency = run_benchmark(num_iters=1, profile=True) + else: + latency = run_benchmark(num_iters=100, profile=False) + print(f"Kernel running time: {latency * 1000000:.3f} us") + + +if __name__ == "__main__": + logger.warning( + "This script benchmarks the paged attention kernel. " + "By default this is no longer used in vLLM inference." + ) + + parser = FlexibleArgumentParser(description="Benchmark the paged attention kernel.") + parser.add_argument("--version", type=str, choices=["v1", "v2"], default="v2") + parser.add_argument("--batch-size", type=int, default=8) + parser.add_argument("--seq-len", type=int, default=4096) + parser.add_argument("--num-query-heads", type=int, default=64) + parser.add_argument("--num-kv-heads", type=int, default=8) + parser.add_argument( + "--head-size", + type=int, + choices=[64, 80, 96, 112, 120, 128, 192, 256], + default=128, + ) + parser.add_argument("--block-size", type=int, choices=[16, 32], default=16) + parser.add_argument("--use-alibi", action="store_true") + parser.add_argument( + "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half" + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--profile", action="store_true") + parser.add_argument( + "--kv-cache-dtype", + type=str, + choices=["auto", "fp8", "fp8_e5m2", "fp8_e4m3"], + default="auto", + help="Data type for kv cache storage. If 'auto', will use model " + "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. " + "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)", + ) + parser.add_argument( + "--custom-paged-attn", action="store_true", help="Use custom paged attention" + ) + args = parser.parse_args() + print(args) + + if args.num_query_heads % args.num_kv_heads != 0: + raise ValueError("num_query_heads must be divisible by num_kv_heads") + main( + version=args.version, + num_seqs=args.batch_size, + seq_len=args.seq_len, + num_query_heads=args.num_query_heads, + num_kv_heads=args.num_kv_heads, + head_size=args.head_size, + block_size=args.block_size, + use_alibi=args.use_alibi, + dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype], + seed=args.seed, + do_profile=args.profile, + kv_cache_dtype=args.kv_cache_dtype, + ) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_quant.py b/online_apiserver_test/benchmarks/kernels/benchmark_quant.py new file mode 100644 index 0000000000000000000000000000000000000000..6ab26f5f1adf73ccdc68c0d5f20d622d0c09b7ba --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_quant.py @@ -0,0 +1,108 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import time + +import torch + +from vllm import _custom_ops as ops +from vllm.platforms import current_platform +from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser + + +@torch.inference_mode() +def main( + num_tokens: int, + hidden_size: int, + static_scale: bool, + quant_dtype: torch.dtype, + dtype: torch.dtype, + seed: int = 0, + do_profile: bool = False, + num_warmup_iters: int = 5, + num_iters: int = 100, +) -> None: + current_platform.seed_everything(seed) + torch.set_default_device("cuda") + + x = torch.randn(num_tokens, hidden_size, dtype=dtype) + scale = torch.randn(1, 1, dtype=torch.float32) if static_scale else None + + def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: + torch.cuda.synchronize() + if profile: + torch.cuda.cudart().cudaProfilerStart() + start_time = time.perf_counter() + + for _ in range(num_iters): + if quant_dtype == torch.int8: + ops.scaled_int8_quant(x, scale) + else: + ops.scaled_fp8_quant(x, scale) + torch.cuda.synchronize() + + end_time = time.perf_counter() + if profile: + torch.cuda.cudart().cudaProfilerStop() + return (end_time - start_time) / num_iters + + # Warmup. + print("Warming up...") + run_benchmark = run_cuda_benchmark + run_benchmark(num_iters=num_warmup_iters, profile=False) + + # Benchmark. + if do_profile: + latency = run_benchmark(num_iters=1, profile=True) + else: + latency = run_benchmark(num_iters=num_iters, profile=False) + print(f"Kernel running time: {latency * 1000000:.3f} us") + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "int8": + return torch.int8 + if dt == "fp8": + return torch.float8_e4m3fn + raise ValueError(f"Unsupported dtype: {dt}") + + parser = FlexibleArgumentParser( + description="Benchmark the quantization (fp8 or int8) kernel." + ) + parser.add_argument("--num-tokens", type=int, default=4096) + parser.add_argument("--hidden-size", type=int, default=8192) + parser.add_argument("--static-scale", action="store_true") + parser.add_argument( + "--quant-dtype", type=str, choices=["fp8", "int8"], default="int8" + ) + parser.add_argument( + "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half" + ) + + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--profile", action="store_true") + parser.add_argument("--num-warmup-iters", type=int, default=5) + parser.add_argument( + "--num-iters", + type=int, + default=100, + help="Number of benchmark iterations. " + "If --profile is set, this number is ignored", + ) + + args = parser.parse_args() + print(args) + + main( + num_tokens=args.num_tokens, + hidden_size=args.hidden_size, + static_scale=args.static_scale, + quant_dtype=to_torch_dtype(args.quant_dtype), + dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype], + seed=args.seed, + do_profile=args.profile, + num_warmup_iters=args.num_warmup_iters, + num_iters=args.num_iters, + ) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_rmsnorm.py b/online_apiserver_test/benchmarks/kernels/benchmark_rmsnorm.py new file mode 100644 index 0000000000000000000000000000000000000000..4cf633a81358d80202fa383d7312dac4a818ea62 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_rmsnorm.py @@ -0,0 +1,256 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import itertools +from typing import Optional, Union + +import torch +from flashinfer.norm import fused_add_rmsnorm, rmsnorm +from torch import nn + +from vllm import _custom_ops as vllm_ops +from vllm.triton_utils import triton + + +class HuggingFaceRMSNorm(nn.Module): + def __init__(self, hidden_size: int, eps: float = 1e-6) -> None: + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward( + self, + x: torch.Tensor, + residual: Optional[torch.Tensor] = None, + ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]: + orig_dtype = x.dtype + x = x.to(torch.float32) + if residual is not None: + x = x + residual.to(torch.float32) + residual = x.to(orig_dtype) + + variance = x.pow(2).mean(dim=-1, keepdim=True) + x = x * torch.rsqrt(variance + self.variance_epsilon) + x = x.to(orig_dtype) * self.weight + if residual is None: + return x + else: + return x, residual + + +def rmsnorm_naive( + x: torch.Tensor, + weight: torch.Tensor, + residual: Optional[torch.Tensor] = None, + eps: float = 1e-6, +): + naive_norm = HuggingFaceRMSNorm(x.shape[-1], eps=eps) + naive_norm.weight = nn.Parameter(weight) + naive_norm = naive_norm.to(x.device) + + orig_shape = x.shape + x = x.view(-1, x.shape[-1]) + if residual is not None: + residual = residual.view(-1, residual.shape[-1]) + + output = naive_norm(x, residual) + + if isinstance(output, tuple): + output = (output[0].view(orig_shape), output[1].view(orig_shape)) + else: + output = output.view(orig_shape) + return output + + +def rmsnorm_flashinfer( + x: torch.Tensor, + weight: torch.Tensor, + residual: Optional[torch.Tensor] = None, + eps: float = 1e-6, +): + orig_shape = x.shape + x = x.view(-1, x.shape[-1]) + if residual is not None: + residual = residual.view(-1, residual.shape[-1]) + + if residual is not None: + fused_add_rmsnorm(x, residual, weight, eps) + output = (x, residual) + else: + output = rmsnorm(x, weight, eps) + + if isinstance(output, tuple): + output = (output[0].view(orig_shape), output[1].view(orig_shape)) + else: + output = output.view(orig_shape) + return output + + +def rmsnorm_vllm( + x: torch.Tensor, + weight: torch.Tensor, + residual: Optional[torch.Tensor] = None, + eps: float = 1e-6, +): + orig_shape = x.shape + x = x.view(-1, x.shape[-1]) + if residual is not None: + residual = residual.view(-1, residual.shape[-1]) + + if residual is not None: + vllm_ops.fused_add_rms_norm(x, residual, weight, eps) + output = (x, residual) + else: + out = torch.empty_like(x) + vllm_ops.rms_norm(out, x, weight, eps) + output = out + + if isinstance(output, tuple): + output = (output[0].view(orig_shape), output[1].view(orig_shape)) + else: + output = output.view(orig_shape) + return output + + +def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True): + dtype = torch.bfloat16 + x = torch.randn(batch_size, seq_len, hidden_size, dtype=dtype, device="cuda") + weight = torch.ones(hidden_size, dtype=dtype, device="cuda") + residual = torch.randn_like(x) if use_residual else None + + output_naive = rmsnorm_naive( + x.clone(), weight, residual.clone() if residual is not None else None + ) + output_flashinfer = rmsnorm_flashinfer( + x.clone(), weight, residual.clone() if residual is not None else None + ) + output_vllm = rmsnorm_vllm( + x.clone(), weight, residual.clone() if residual is not None else None + ) + + if use_residual: + output_naive = output_naive[0] + output_flashinfer = output_flashinfer[0] + output_vllm = output_vllm[0] + + print(f"Naive output={output_naive}") + print(f"FlashInfer output={output_flashinfer}") + print(f"vLLM output={output_vllm}") + + if torch.allclose( + output_naive, output_flashinfer, atol=1e-2, rtol=1e-2 + ) and torch.allclose(output_naive, output_vllm, atol=1e-2, rtol=1e-2): + print("✅ All implementations match") + else: + print("❌ Implementations differ") + + +batch_size_range = [2**i for i in range(0, 7, 2)] +seq_length_range = [2**i for i in range(6, 11, 1)] +head_num_range = [32, 48] +configs = list(itertools.product(head_num_range, batch_size_range, seq_length_range)) + + +def get_benchmark(use_residual): + @triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["head_num", "batch_size", "seq_len"], + x_vals=[list(_) for _ in configs], + line_arg="provider", + line_vals=["huggingface", "flashinfer", "vllm"], + line_names=["HuggingFace", "FlashInfer", "vLLM"], + styles=[("blue", "-"), ("green", "-"), ("red", "-")], + ylabel="us", + plot_name=f"rmsnorm-perf-{'with' if use_residual else 'without'}-residual", + args={}, + ) + ) + def benchmark(head_num, batch_size, seq_len, provider): + dtype = torch.bfloat16 + hidden_size = head_num * 128 # assuming head_dim = 128 + + x = torch.randn(batch_size, seq_len, hidden_size, dtype=dtype, device="cuda") + weight = torch.ones(hidden_size, dtype=dtype, device="cuda") + residual = torch.randn_like(x) if use_residual else None + + quantiles = [0.5, 0.2, 0.8] + + if provider == "huggingface": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: rmsnorm_naive( + x.clone(), + weight, + residual.clone() if residual is not None else None, + ), + quantiles=quantiles, + ) + elif provider == "flashinfer": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: rmsnorm_flashinfer( + x.clone(), + weight, + residual.clone() if residual is not None else None, + ), + quantiles=quantiles, + ) + else: + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: rmsnorm_vllm( + x.clone(), + weight, + residual.clone() if residual is not None else None, + ), + quantiles=quantiles, + ) + + return 1000 * ms, 1000 * max_ms, 1000 * min_ms + + return benchmark + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument( + "--batch-size", + type=int, + default=4, + help="Batch size", + ) + parser.add_argument( + "--seq-len", + type=int, + default=128, + help="Sequence length", + ) + parser.add_argument( + "--hidden-size", + type=int, + default=4096, + help="Hidden size (2nd dimension) of the sequence", + ) + parser.add_argument( + "--use-residual", action="store_true", help="Whether to use residual connection" + ) + parser.add_argument( + "--save-path", + type=str, + default="./configs/rmsnorm/", + help="Path to save rmsnorm benchmark results", + ) + + args = parser.parse_args() + + # Run correctness test + calculate_diff( + batch_size=args.batch_size, + seq_len=args.seq_len, + hidden_size=args.hidden_size, + use_residual=args.use_residual, + ) + + # Get the benchmark function with proper use_residual setting + benchmark = get_benchmark(args.use_residual) + # Run performance benchmark + benchmark.run(print_data=True, save_path=args.save_path) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_rope.py b/online_apiserver_test/benchmarks/kernels/benchmark_rope.py new file mode 100644 index 0000000000000000000000000000000000000000..b81baf17a8c674edd72fe9a6397a0b027a8d33b9 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_rope.py @@ -0,0 +1,133 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +from itertools import accumulate +from typing import Optional + +import nvtx +import torch + +from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope +from vllm.platforms import current_platform +from vllm.utils import FlexibleArgumentParser + + +def benchmark_rope_kernels_multi_lora( + is_neox_style: bool, + batch_size: int, + seq_len: int, + num_heads: int, + head_size: int, + rotary_dim: Optional[int], + dtype: torch.dtype, + seed: int, + device: str, + max_position: int = 8192, + base: float = 10000, +) -> None: + current_platform.seed_everything(seed) + torch.set_default_device(device) + if rotary_dim is None: + rotary_dim = head_size + # silulating serving 4 LoRAs + scaling_factors = [1, 2, 4, 8] + # batched RoPE can take multiple scaling factors + batched_rope = get_rope( + head_size, + rotary_dim, + max_position, + base, + is_neox_style, + {"rope_type": "linear", "factor": tuple(scaling_factors)}, + ) + # non-batched RoPE takes only one scaling factor, we create multiple + # instances to simulate the same behavior + non_batched_ropes: list[RotaryEmbedding] = [] + for scaling_factor in scaling_factors: + non_batched_ropes.append( + get_rope( + head_size, + rotary_dim, + max_position, + base, + is_neox_style, + {"rope_type": "linear", "factor": (scaling_factor,)}, + ) + ) + + positions = torch.randint(0, max_position, (batch_size, seq_len)) + query = torch.randn(batch_size, seq_len, num_heads * head_size, dtype=dtype) + key = torch.randn_like(query) + + # create query offsets for batched RoPE, we concat multiple kv cache + # together and each query needs to find the right kv cache of its type + offset_map = torch.tensor( + list( + accumulate( + [0] + + [ + max_position * scaling_factor * 2 + for scaling_factor in scaling_factors[:-1] + ] + ) + ) + ) + query_types = torch.randint( + 0, len(scaling_factors), (batch_size, seq_len), device=device + ) + # map query types to offsets + query_offsets = offset_map[query_types] + # the kernel takes flattened offsets + flatten_offsets = query_offsets.flatten() + + # batched queries of the same type together for non-batched RoPE + queries = [query[query_types == i] for i in range(len(scaling_factors))] + keys = [key[query_types == i] for i in range(len(scaling_factors))] + packed_qkr = zip(queries, keys, non_batched_ropes) + # synchronize before start timing + torch.cuda.synchronize() + with nvtx.annotate("non-batched", color="yellow"): + for q, k, r in packed_qkr: + r.forward(positions, q, k) + torch.cuda.synchronize() + with nvtx.annotate("batched", color="green"): + batched_rope.forward(positions, query, key, flatten_offsets) + torch.cuda.synchronize() + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark the rotary embedding kernels." + ) + parser.add_argument("--is-neox-style", type=bool, default=True) + parser.add_argument("--batch-size", type=int, default=16) + parser.add_argument("--seq-len", type=int, default=512) + parser.add_argument("--num-heads", type=int, default=8) + parser.add_argument( + "--head-size", + type=int, + choices=[64, 80, 96, 112, 120, 128, 192, 256], + default=128, + ) + parser.add_argument("--rotary-dim", type=int, choices=[16, 32], default=32) + parser.add_argument( + "--dtype", type=str, choices=["bfloat16", "float"], default="float" + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument( + "--device", type=str, choices=["cuda:0", "cuda:1"], default="cuda:0" + ) + args = parser.parse_args() + print(args) + + benchmark_rope_kernels_multi_lora( + is_neox_style=args.is_neox_style, + batch_size=args.batch_size, + seq_len=args.seq_len, + num_heads=args.num_heads, + head_size=args.head_size, + rotary_dim=args.rotary_dim, + dtype=getattr(torch, args.dtype), + seed=args.seed, + device=args.device, + ) diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_shapes.py b/online_apiserver_test/benchmarks/kernels/benchmark_shapes.py new file mode 100644 index 0000000000000000000000000000000000000000..18c459c31d3f84d1ca7b7e31d00d50b521177e96 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_shapes.py @@ -0,0 +1,94 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +WEIGHT_SHAPES = { + "ideal": [[4 * 256 * 32, 256 * 32]], + "mistralai/Mistral-7B-v0.1/TP1": [ + [4096, 6144], + [4096, 4096], + [4096, 28672], + [14336, 4096], + ], + "mistralai/Mistral-7B-v0.1/TP2": [ + [4096, 3072], + [2048, 4096], + [4096, 14336], + [7168, 4096], + ], + "mistralai/Mistral-7B-v0.1/TP4": [ + [4096, 1536], + [1024, 4096], + [4096, 7168], + [3584, 4096], + ], + "meta-llama/Llama-2-7b-hf/TP1": [ + [4096, 12288], + [4096, 4096], + [4096, 22016], + [11008, 4096], + ], + "meta-llama/Llama-2-7b-hf/TP2": [ + [4096, 6144], + [2048, 4096], + [4096, 11008], + [5504, 4096], + ], + "meta-llama/Llama-2-7b-hf/TP4": [ + [4096, 3072], + [1024, 4096], + [4096, 5504], + [2752, 4096], + ], + "meta-llama/Llama-2-13b-hf/TP1": [ + [5120, 15360], + [5120, 5120], + [5120, 27648], + [13824, 5120], + ], + "meta-llama/Llama-2-13b-hf/TP2": [ + [5120, 7680], + [2560, 5120], + [5120, 13824], + [6912, 5120], + ], + "meta-llama/Llama-2-13b-hf/TP4": [ + [5120, 3840], + [1280, 5120], + [5120, 6912], + [3456, 5120], + ], + "meta-llama/Llama-2-70b-hf/TP1": [ + [8192, 10240], + [8192, 8192], + [8192, 57344], + [28672, 8192], + ], + "meta-llama/Llama-2-70b-hf/TP2": [ + [8192, 5120], + [4096, 8192], + [8192, 28672], + [14336, 8192], + ], + "meta-llama/Llama-2-70b-hf/TP4": [ + [8192, 2560], + [2048, 8192], + [8192, 14336], + [7168, 8192], + ], +} + +WEIGHT_SHAPES_MOE = { + "nm-testing/Mixtral-8x7B-Instruct-v0.1": [ + [8, 2, 4096, 28672], + [8, 2, 14336, 4096], + ], + "nm-testing/deepseekv2-lite": [ + [64, 6, 2048, 1408], + ], + "ibm-granite/granite-3.0-1b-a400m": [ + [32, 8, 1024, 1024], + ], + "ibm-granite/granite-3.0-3b-a800m": [ + [40, 8, 1024, 1536], + ], +} diff --git a/online_apiserver_test/benchmarks/kernels/benchmark_w8a8_block_fp8.py b/online_apiserver_test/benchmarks/kernels/benchmark_w8a8_block_fp8.py new file mode 100644 index 0000000000000000000000000000000000000000..4fcdbadd65ecd366f51ce5d9053cc700854dcb71 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/benchmark_w8a8_block_fp8.py @@ -0,0 +1,414 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +# Adapted from sglang quantization/tuning_block_wise_kernel.py + +import argparse +import json +import multiprocessing as mp +import os +import time +from datetime import datetime +from typing import Any + +import torch +import tqdm +import triton + +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + _w8a8_block_fp8_matmul, +) +from vllm.platforms import current_platform +from vllm.utils import FlexibleArgumentParser + +mp.set_start_method("spawn", force=True) + +assert current_platform.is_cuda(), ( + "Only support tune w8a8 block fp8 kernel on CUDA device." +) + +DTYPE_MAP = { + "float32": torch.float32, + "float16": torch.float16, + "half": torch.half, + "bfloat16": torch.bfloat16, +} + + +def w8a8_block_matmul( + A: torch.Tensor, + B: torch.Tensor, + As: torch.Tensor, + Bs: torch.Tensor, + block_size: list[int], + config: dict[str, Any], + output_dtype: torch.dtype = torch.float16, +) -> torch.Tensor: + """This function performs matrix multiplication with + block-wise quantization. + + It takes two input tensors `A` and `B` with scales `As` and `Bs`. + The output is returned in the specified `output_dtype`. + + Args: + A: The input tensor, e.g., activation. + B: The input tensor, e.g., weight. + As: The per-token-group quantization scale for `A`. + Bs: The per-block quantization scale for `B`. + block_size: The block size for per-block quantization. + It should be 2-dim, e.g., [128, 128]. + output_dytpe: The dtype of the returned tensor. + + Returns: + torch.Tensor: The result of matmul. + """ + assert len(block_size) == 2 + block_n, block_k = block_size[0], block_size[1] + + assert A.shape[-1] == B.shape[-1] + assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous() + assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1] + M = A.numel() // A.shape[-1] + + assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2 + N, K = B.shape + assert triton.cdiv(N, block_n) == Bs.shape[0] + assert triton.cdiv(K, block_k) == Bs.shape[1] + + C_shape = A.shape[:-1] + (N,) + C = A.new_empty(C_shape, dtype=output_dtype) + + def grid(META): + return ( + triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]), + ) + + if A.dtype == torch.float8_e4m3fn: + kernel = _w8a8_block_fp8_matmul + else: + raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.") + + kernel[grid]( + A, + B, + C, + As, + Bs, + M, + N, + K, + block_n, + block_k, + A.stride(-2), + A.stride(-1), + B.stride(1), + B.stride(0), + C.stride(-2), + C.stride(-1), + As.stride(-2), + As.stride(-1), + Bs.stride(1), + Bs.stride(0), + **config, + ) + + return C + + +def get_configs_compute_bound(): + configs = [] + for num_stages in [2, 3, 4, 5]: + for block_m in [16, 32, 64, 128, 256]: + for block_k in [64, 128]: + for block_n in [32, 64, 128, 256]: + for num_warps in [4, 8]: + for group_size in [1, 16, 32, 64]: + configs.append( + { + "BLOCK_SIZE_M": block_m, + "BLOCK_SIZE_N": block_n, + "BLOCK_SIZE_K": block_k, + "GROUP_SIZE_M": group_size, + "num_warps": num_warps, + "num_stages": num_stages, + } + ) + return configs + + +def get_weight_shapes(tp_size): + # NOTE(HandH1998): The weight shapes only works for DeepSeek-V3. + # Modify them, if you tune for another different model. + # cannot TP + total = [ + (512 + 64, 7168), + ((128 + 64) * 128, 7168), + (128 * (128 + 128), 512), + (7168, 16384), + (7168, 18432), + ] + # N can TP + n_tp = [ + (18432 * 2, 7168), + ((128 + 64) * 128, 7168), + (128 * (128 + 128), 512), + (24576, 1536), + (12288, 7168), + (4096, 7168), + ] + # K can TP + k_tp = [(7168, 18432), (7168, 16384), (7168, 2048)] + + weight_shapes = [] + for t in total: + weight_shapes.append(t) + for n_t in n_tp: + new_t = (n_t[0] // tp_size, n_t[1]) + weight_shapes.append(new_t) + for k_t in k_tp: + new_t = (k_t[0], k_t[1] // tp_size) + weight_shapes.append(new_t) + return weight_shapes + + +def benchmark_config( + A, B, As, Bs, block_size, config, out_dtype=torch.float16, num_iters=10 +): + def run(): + w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype) + + torch.cuda.synchronize() + # JIT complication & warmup + for _ in range(5): + run() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + torch.cuda.synchronize() + start_event.record() + run() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + return avg + + +def tune(M, N, K, block_size, out_dtype, search_space, input_type): + factor_for_scale = 1e-2 + + if input_type == "fp8": + fp8_info = torch.finfo(torch.float8_e4m3fn) + fp8_max, fp8_min = fp8_info.max, fp8_info.min + + A_fp32 = ( + (torch.rand(M, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max + ) + A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) + + B_fp32 = ( + (torch.rand(N, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max + ) + B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) + else: + raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.") + + block_n, block_k = block_size[0], block_size[1] + n_tiles = (N + block_n - 1) // block_n + k_tiles = (K + block_k - 1) // block_k + + As = torch.rand(M, k_tiles, dtype=torch.float32, device="cuda") * factor_for_scale + Bs = ( + torch.rand(n_tiles, k_tiles, dtype=torch.float32, device="cuda") + * factor_for_scale + ) + + best_config = None + best_time = float("inf") + for config in tqdm(search_space): + try: + kernel_time = benchmark_config( + A, + B, + As, + Bs, + block_size, + config, + out_dtype, + num_iters=10, + ) + except triton.runtime.autotuner.OutOfResources: + # Some configurations may be invalid and fail to compile. + continue + + if kernel_time < best_time: + best_time = kernel_time + best_config = config + now = datetime.now() + print(f"{now.ctime()}] Completed tuning for batch_size={M}") + assert best_config is not None + return best_config + + +def save_configs( + N, + K, + block_n, + block_k, + configs, + save_path, + input_type="fp8", +) -> None: + os.makedirs(save_path, exist_ok=True) + device_name = current_platform.get_device_name().replace(" ", "_") + json_file_name = ( + f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8," + f"block_shape=[{block_n},{block_k}].json" + ) + + config_file_path = os.path.join(save_path, json_file_name) + print(f"Writing best config to {config_file_path}...") + + with open(config_file_path, "w") as f: + json.dump(configs, f, indent=4) + f.write("\n") + + +def tune_on_gpu(args_dict): + """Run tuning on a specific GPU.""" + gpu_id = args_dict["gpu_id"] + batch_sizes = args_dict["batch_sizes"] + weight_shapes = args_dict["weight_shapes"] + args = args_dict["args"] + + torch.cuda.set_device(gpu_id) + print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}") + + block_n = args.block_n + block_k = args.block_k + out_dtype = DTYPE_MAP[args.out_dtype] + save_path = args.save_path + input_type = args.input_type + + search_space = get_configs_compute_bound() + search_space = [ + config for config in search_space if block_k % config["BLOCK_SIZE_K"] == 0 + ] + + start = time.time() + for shape in tqdm(weight_shapes, desc=f"GPU {gpu_id} - Shapes"): + N, K = shape[0], shape[1] + print(f"[GPU {gpu_id}] Tune for weight shape of `N: {N}, K: {K}`") + benchmark_results = [ + tune( + batch_size, + N, + K, + [block_n, block_k], + out_dtype, + search_space, + input_type, + ) + for batch_size in tqdm(batch_sizes, desc=f"GPU {gpu_id} - Batch sizes") + ] + best_configs = {M: config for M, config in zip(batch_sizes, benchmark_results)} + save_configs(N, K, block_n, block_k, best_configs, save_path, input_type) + + end = time.time() + print(f"Tuning on GPU {gpu_id} took {end - start:.2f} seconds") + + +def distribute_batch_sizes(batch_sizes, num_gpus): + """Distribute batch sizes across available GPUs.""" + batches_per_gpu = [] + for i in range(num_gpus): + start_idx = i * len(batch_sizes) // num_gpus + end_idx = (i + 1) * len(batch_sizes) // num_gpus + batches_per_gpu.append(batch_sizes[start_idx:end_idx]) + return batches_per_gpu + + +def main(args): + print(args) + num_gpus = torch.cuda.device_count() + if num_gpus == 0: + raise RuntimeError("No GPU available for tuning") + print(f"Found {num_gpus} GPUs for parallel tuning") + + torch.cuda.init() + + if args.batch_size is None: + batch_sizes = [ + 1, + 2, + 4, + 8, + 16, + 24, + 32, + 48, + 64, + 96, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ] + else: + batch_sizes = [args.batch_size] + num_gpus = 1 # If only one batch size, use only one GPU + + weight_shapes = get_weight_shapes(args.tp_size) + + batches_per_gpu = distribute_batch_sizes(batch_sizes, num_gpus) + + process_args = [] + for gpu_id in range(num_gpus): + process_args.append( + { + "gpu_id": gpu_id, + "batch_sizes": batches_per_gpu[gpu_id], + "weight_shapes": weight_shapes, # Each GPU processes all weight shapes + "args": args, + } + ) + + ctx = mp.get_context("spawn") + with ctx.Pool(num_gpus) as pool: + pool.map(tune_on_gpu, process_args) + + print("Multi-GPU tuning completed") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description=""" +Tune triton w8a8 block fp8 for DeepSeek-V3/DeepSeek-R1: + python3 benchmark_w8a8_block_fp8.py --tp-size 8 --input-type fp8 +Then copy to model_executor/layers/quantization/utils/configs + """, + formatter_class=argparse.RawTextHelpFormatter, + ) + + parser.add_argument("--tp-size", "-tp", type=int, default=8) + parser.add_argument("--input-type", type=str, choices=["fp8"], default="fp8") + parser.add_argument( + "--out-dtype", + type=str, + choices=["float32", "float16", "bfloat16", "half"], + default="float16", + ) + parser.add_argument("--block-n", type=int, default=128) + parser.add_argument("--block-k", type=int, default=128) + parser.add_argument("--batch-size", type=int, required=False) + parser.add_argument("--save-path", type=str, default="./") + args = parser.parse_args() + + main(args) diff --git a/online_apiserver_test/benchmarks/kernels/deepgemm/README.md b/online_apiserver_test/benchmarks/kernels/deepgemm/README.md new file mode 100644 index 0000000000000000000000000000000000000000..917e814010f894cd812e42849feb5822ad332d04 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/deepgemm/README.md @@ -0,0 +1,129 @@ +# DeepSeek DeepGEMM Kernels Benchmark + +This directory includes benchmarks between DeepSeek's DeepGEMM block fp8 kernels against vLLM's existing triton and CUTLASS-based kernels. + +Currently this just includes dense GEMMs and only works on Hopper GPUs. + +## Setup + +You need to install vLLM in your usual fashion, then install DeepGEMM from source in its own directory: + +``` +git clone --recursive https://github.com/deepseek-ai/DeepGEMM +cd DeepGEMM +python setup.py install +uv pip install -e . +``` + +## Usage + +``` +python benchmark_fp8_block_dense_gemm.py +INFO 02-26 21:55:13 [__init__.py:207] Automatically detected platform cuda. +===== STARTING FP8 GEMM BENCHMARK ===== +PyTorch version: 2.5.1+cu124 +CUDA version: 12.4 +Triton version: 3.1.0 +Using device: NVIDIA H100 80GB HBM3 +WARNING 02-26 21:55:15 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +INFO 02-26 21:55:15 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel. +WARNING 02-26 21:55:16 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=18432,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +WARNING 02-26 21:55:17 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel. +INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel. + +===== PERFORMANCE COMPARISON ===== + +DeepGEMM Implementation: ++------+-------+-------+-----------+--------+--------+ +| m | n | k | Time (μs) | TFLOPS | GB/s | ++------+-------+-------+-----------+--------+--------+ +| 8 | 4096 | 7168 | 102.9 | 4.6 | 286.4 | +| 8 | 7168 | 18432 | 70.8 | 29.8 | 1868.8 | +| 8 | 18432 | 7168 | 69.3 | 30.5 | 1911.8 | +| 64 | 4096 | 7168 | 69.1 | 54.4 | 439.0 | +| 64 | 7168 | 18432 | 69.4 | 243.6 | 1933.6 | +| 64 | 18432 | 7168 | 70.4 | 240.3 | 1917.2 | +| 64 | 24576 | 1536 | 70.1 | 68.9 | 584.6 | +| 64 | 32768 | 512 | 68.4 | 31.4 | 307.1 | +| 64 | 7168 | 16384 | 69.5 | 216.3 | 1718.5 | +| 128 | 4096 | 7168 | 141.1 | 53.3 | 222.1 | +| 128 | 7168 | 18432 | 71.9 | 470.5 | 1896.1 | +| 128 | 18432 | 7168 | 69.3 | 488.2 | 1988.2 | +| 1024 | 4096 | 7168 | 89.7 | 670.1 | 502.5 | +| 1024 | 18432 | 7168 | 279.0 | 969.8 | 635.2 | +| 2048 | 4096 | 7168 | 175.1 | 687.0 | 347.4 | +| 4096 | 4096 | 7168 | 335.4 | 717.0 | 275.1 | ++------+-------+-------+-----------+--------+--------+ + +vLLM Triton Implementation: ++------+-------+-------+-----------+--------+--------+--------------+ +| m | n | k | Time (μs) | TFLOPS | GB/s | vs DeepGEMM | ++------+-------+-------+-----------+--------+--------+--------------+ +| 8 | 4096 | 7168 | 74.0 | 6.3 | 398.2 | 1.39x faster | +| 8 | 7168 | 18432 | 89.6 | 23.6 | 1478.1 | 0.79x slower | +| 8 | 18432 | 7168 | 113.2 | 18.7 | 1170.4 | 0.61x slower | +| 64 | 4096 | 7168 | 79.4 | 47.3 | 382.2 | 0.87x slower | +| 64 | 7168 | 18432 | 98.5 | 171.7 | 1363.0 | 0.70x slower | +| 64 | 18432 | 7168 | 119.5 | 141.5 | 1129.4 | 0.59x slower | +| 64 | 24576 | 1536 | 37.6 | 128.4 | 1089.7 | 1.86x faster | +| 64 | 32768 | 512 | 38.7 | 55.5 | 542.6 | 1.77x faster | +| 64 | 7168 | 16384 | 86.1 | 174.5 | 1386.4 | 0.81x slower | +| 128 | 4096 | 7168 | 90.7 | 82.9 | 345.4 | 1.56x faster | +| 128 | 7168 | 18432 | 144.0 | 234.9 | 946.9 | 0.50x slower | +| 128 | 18432 | 7168 | 229.5 | 147.4 | 600.1 | 0.30x slower | +| 1024 | 4096 | 7168 | 242.3 | 248.2 | 186.1 | 0.37x slower | +| 1024 | 18432 | 7168 | 897.8 | 301.4 | 197.4 | 0.31x slower | +| 2048 | 4096 | 7168 | 463.0 | 259.7 | 131.4 | 0.38x slower | +| 4096 | 4096 | 7168 | 901.8 | 266.7 | 102.3 | 0.37x slower | ++------+-------+-------+-----------+--------+--------+--------------+ + +vLLM CUTLASS Implementation: ++------+-------+-------+-----------+--------+--------+--------------+--------------+ +| m | n | k | Time (μs) | TFLOPS | GB/s | vs DeepGEMM | vs Triton | ++------+-------+-------+-----------+--------+--------+--------------+--------------+ +| 8 | 4096 | 7168 | 34.6 | 13.6 | 852.3 | 2.98x faster | 2.14x faster | +| 8 | 7168 | 18432 | 78.9 | 26.8 | 1677.3 | 0.90x slower | 1.13x faster | +| 8 | 18432 | 7168 | 81.2 | 26.0 | 1631.1 | 0.85x slower | 1.39x faster | +| 64 | 4096 | 7168 | 36.9 | 101.9 | 822.9 | 1.87x faster | 2.15x faster | +| 64 | 7168 | 18432 | 87.4 | 193.4 | 1535.2 | 0.79x slower | 1.13x faster | +| 64 | 18432 | 7168 | 85.0 | 199.0 | 1587.6 | 0.83x slower | 1.41x faster | +| 64 | 24576 | 1536 | 28.0 | 172.8 | 1465.8 | 2.51x faster | 1.35x faster | +| 64 | 32768 | 512 | 28.8 | 74.5 | 728.5 | 2.37x faster | 1.34x faster | +| 64 | 7168 | 16384 | 77.9 | 193.0 | 1532.8 | 0.89x slower | 1.11x faster | +| 128 | 4096 | 7168 | 39.1 | 192.4 | 802.0 | 3.61x faster | 2.32x faster | +| 128 | 7168 | 18432 | 93.7 | 360.8 | 1454.2 | 0.77x slower | 1.54x faster | +| 128 | 18432 | 7168 | 85.7 | 394.8 | 1608.0 | 0.81x slower | 2.68x faster | +| 1024 | 4096 | 7168 | 99.7 | 603.1 | 452.2 | 0.90x slower | 2.43x faster | +| 1024 | 18432 | 7168 | 331.3 | 816.7 | 534.9 | 0.84x slower | 2.71x faster | +| 2048 | 4096 | 7168 | 198.3 | 606.6 | 306.7 | 0.88x slower | 2.34x faster | +| 4096 | 4096 | 7168 | 392.2 | 613.2 | 235.3 | 0.86x slower | 2.30x faster | ++------+-------+-------+-----------+--------+--------+--------------+--------------+ + +===== AVERAGE PERFORMANCE ===== ++----------------+------------+----------+---------------+ +| Implementation | Avg TFLOPS | Avg GB/s | Avg Time (ms) | ++----------------+------------+----------+---------------+ +| DeepGEMM | 310.98 | 1052.10 | 0.11 | +| vLLM Triton | 144.30 | 715.60 | 0.23 | +| vLLM CUTLASS | 286.78 | 1076.67 | 0.11 | ++----------------+------------+----------+---------------+ + +===== AVERAGE SPEEDUPS ===== ++-----------------------------+--------------+ +| Comparison | Speedup | ++-----------------------------+--------------+ +| DeepGEMM vs vLLM Triton | 1.71x faster | +| DeepGEMM vs vLLM CUTLASS | 0.94x slower | +| vLLM CUTLASS vs vLLM Triton | 1.84x faster | ++-----------------------------+--------------+ + +===== ACCURACY COMPARISON ===== ++----------------+-----------------------+ +| Implementation | Avg Diff vs Reference | ++----------------+-----------------------+ +| DeepGEMM | 0.000684 | +| vLLM Triton | 0.000684 | +| vLLM CUTLASS | 0.000684 | ++----------------+-----------------------+ +``` diff --git a/online_apiserver_test/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py b/online_apiserver_test/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py new file mode 100644 index 0000000000000000000000000000000000000000..43c54d56ca8c1be027bec0d5f54dd3e53a294771 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py @@ -0,0 +1,456 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +# fmt: off +# ruff: noqa: E501 +import time + +# Import DeepGEMM functions +import deep_gemm +import torch +from deep_gemm import calc_diff, ceil_div, get_col_major_tma_aligned_tensor + +# Import vLLM functions +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + per_token_group_quant_fp8, + w8a8_block_fp8_matmul, +) +from vllm.triton_utils import triton + + +# Copied from +# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L9 +def per_token_cast_to_fp8( + x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """Convert tensor to FP8 format with per-token scaling.""" + assert x.dim() == 2 and x.size(1) % 128 == 0 + m, n = x.shape + x_view = x.view(m, -1, 128) + x_amax = x_view.abs().float().amax(dim=2).view(m, -1).clamp(1e-4) + return (x_view * (448.0 / x_amax.unsqueeze(2))).to( + torch.float8_e4m3fn).view(m, n), (x_amax / 448.0).view(m, -1) + + +# Copied from +# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L17 +def per_block_cast_to_fp8( + x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """Convert tensor to FP8 format with per-block scaling.""" + assert x.dim() == 2 + m, n = x.shape + x_padded = torch.zeros((ceil_div(m, 128) * 128, ceil_div(n, 128) * 128), + dtype=x.dtype, + device=x.device) + x_padded[:m, :n] = x + x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128) + x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4) + x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn) + return x_scaled.view_as(x_padded)[:m, :n].contiguous(), ( + x_amax / 448.0).view(x_view.size(0), x_view.size(2)) + + +def benchmark_shape(m: int, + n: int, + k: int, + warmup: int = 100, + repeat: int = 10000, + verbose: bool = False) -> dict: + """Benchmark all implementations for a specific (m, n, k) shape.""" + if verbose: + print(f"\n=== Benchmarking shape: m={m}, n={n}, k={k} ===") + + # Create test tensors + A = torch.randn((m, k), device='cuda', dtype=torch.bfloat16) + B = torch.randn((n, k), device='cuda', dtype=torch.bfloat16) + + # Reference result in BF16 + torch.cuda.synchronize() + C_ref = A @ B.t() + + # Pre-quantize B for all implementations + # (weights can be pre-quantized offline) + B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B) + B_vllm, B_scale_vllm = per_block_cast_to_fp8(B) + + # Block size configuration + block_size = [128, 128] + + # Pre-quantize A for all implementations + A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A) + A_scale_deepgemm = get_col_major_tma_aligned_tensor(A_scale_deepgemm) + C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16) + A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1]) + A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8( + A, block_size[1], column_major_scales=True) + + # === DeepGEMM Implementation === + def deepgemm_gemm(): + deep_gemm.gemm_fp8_fp8_bf16_nt((A_deepgemm, A_scale_deepgemm), + (B_deepgemm, B_scale_deepgemm), + C_deepgemm) + return C_deepgemm + + # === vLLM Triton Implementation === + def vllm_triton_gemm(): + return w8a8_block_fp8_matmul(A_vllm, + B_vllm, + A_scale_vllm, + B_scale_vllm, + block_size, + output_dtype=torch.bfloat16) + + # === vLLM CUTLASS Implementation === + def vllm_cutlass_gemm(): + return ops.cutlass_scaled_mm(A_vllm_cutlass, + B_vllm.T, + scale_a=A_scale_vllm_cutlass, + scale_b=B_scale_vllm.T, + out_dtype=torch.bfloat16) + + # Run correctness check first + if verbose: + print("Running correctness check...") + C_deepgemm = deepgemm_gemm() + C_vllm_triton = vllm_triton_gemm() + C_vllm_cutlass = vllm_cutlass_gemm() + + deepgemm_diff = calc_diff(C_deepgemm, C_ref) + vllm_triton_diff = calc_diff(C_vllm_triton, C_ref) + vllm_cutlass_diff = calc_diff(C_vllm_cutlass, C_ref) + + if verbose: + print(f"DeepGEMM vs Reference difference: {deepgemm_diff:.6f}") + print(f"vLLM Triton vs Reference difference: {vllm_triton_diff:.6f}") + print(f"vLLM CUTLASS vs Reference difference: {vllm_cutlass_diff:.6f}") + print("vLLM Triton vs DeepGEMM difference: " + f"{calc_diff(C_vllm_triton, C_deepgemm):.6f}") + print("vLLM CUTLASS vs DeepGEMM difference: " + f"{calc_diff(C_vllm_cutlass, C_deepgemm):.6f}") + + # Benchmark implementations + implementations = { + "DeepGEMM": deepgemm_gemm, + "vLLM Triton": vllm_triton_gemm, + "vLLM CUTLASS": vllm_cutlass_gemm + } + + benchmark_results = { + "shape": { + "m": m, + "n": n, + "k": k + }, + "implementations": {} + } + + for name, func in implementations.items(): + # Warmup + for _ in range(warmup): + func() + torch.cuda.synchronize() + + # Timing loop + torch.cuda.synchronize() + start = time.time() + for _ in range(repeat): + func() + torch.cuda.synchronize() + end = time.time() + + # Calculate timing and TFLOPS + avg_time_ms = (end - start) / repeat * 1000 + avg_time_us = avg_time_ms * 1000 + tflops = 2 * m * n * k / (avg_time_ms * 1e-3) / 1e12 + gb_s = (m * k + k * n + m * n * 2) / 1e9 / (avg_time_ms * 1e-3) + + benchmark_results["implementations"][name] = { + "time_ms": avg_time_ms, + "time_us": avg_time_us, + "tflops": tflops, + "gb_s": gb_s, + "diff": { + "DeepGEMM": + 0.0 if name == "DeepGEMM" else calc_diff(func(), C_deepgemm), + "Reference": + deepgemm_diff if name == "DeepGEMM" else + (vllm_triton_diff + if name == "vLLM Triton" else vllm_cutlass_diff) + } + } + + if verbose: + print( + f"{name}: {avg_time_ms:.3f} ms, {tflops:.2f} TFLOPS, {gb_s:.2f} GB/s" + ) + + # Calculate speedups + baseline = benchmark_results["implementations"]["DeepGEMM"]["time_ms"] + for name, data in benchmark_results["implementations"].items(): + if name != "DeepGEMM": + speedup = baseline / data["time_ms"] + benchmark_results["implementations"][name][ + "speedup_vs_deepgemm"] = speedup + if verbose: + print(f"DeepGEMM is {1/speedup:.2f}x " + f"{'faster' if 1/speedup > 1 else 'slower'} than {name}") + + vllm_triton_time = benchmark_results["implementations"]["vLLM Triton"][ + "time_ms"] + vllm_cutlass_time = benchmark_results["implementations"]["vLLM CUTLASS"][ + "time_ms"] + cutlass_vs_triton = vllm_triton_time / vllm_cutlass_time + benchmark_results["implementations"]["vLLM CUTLASS"][ + "speedup_vs_triton"] = cutlass_vs_triton + if verbose: + print( + f"vLLM CUTLASS is {cutlass_vs_triton:.2f}x " + f"{'faster' if cutlass_vs_triton > 1 else 'slower'} than vLLM Triton" + ) + + return benchmark_results + + +def format_table_row(values, widths): + """Format a row with specified column widths.""" + return "| " + " | ".join(f"{val:{w}}" + for val, w in zip(values, widths)) + " |" + + +def print_table(headers, rows, title=None): + """Print a table with headers and rows.""" + if title: + print(f"\n{title}") + + # Calculate column widths based on headers and data + widths = [ + max(len(str(h)), max(len(str(row[i])) for row in rows)) + for i, h in enumerate(headers) + ] + + # Create separator line + separator = "+-" + "-+-".join("-" * w for w in widths) + "-+" + + # Print table + print(separator) + print(format_table_row(headers, widths)) + print(separator) + for row in rows: + print(format_table_row(row, widths)) + print(separator) + + +def format_speedup(value): + """Format speedup value with indicator if it's faster or slower.""" + return f"{value:.2f}x {'faster' if value > 1.0 else 'slower'}" + + +def run_benchmarks(verbose: bool = False): + """Run benchmarks for a set of common shapes.""" + print("===== STARTING FP8 GEMM BENCHMARK =====") + + # Make sure we're using the GPU + if not torch.cuda.is_available(): + print("CUDA not available! Tests require GPU.") + return + + # Print system information + print(f"PyTorch version: {torch.__version__}") + print(f"CUDA version: {torch.version.cuda}") + print(f"Triton version: {triton.__version__}") + print(f"Using device: {torch.cuda.get_device_name()}") + + # Enable TF32 for better performance + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + + # Set seeds for reproducibility + torch.manual_seed(42) + torch.cuda.manual_seed(42) + + # Define benchmark shapes (m, n, k) + shapes = [ + (8, 4096, 7168), + (8, 7168, 18432), + (8, 18432, 7168), + (64, 4096, 7168), + (64, 7168, 18432), + (64, 18432, 7168), + (64, 24576, 1536), + (64, 32768, 512), + (64, 7168, 16384), + (128, 4096, 7168), + (128, 7168, 18432), + (128, 18432, 7168), + (1024, 4096, 7168), + (1024, 18432, 7168), + (2048, 4096, 7168), + (4096, 4096, 7168), + ] + shapes = [ + # (64, 2112, 7168), + (64, 24576, 1536), + (64, 32768, 512), + (64, 7168, 16384), + (64, 4096, 7168), + (64, 7168, 2048), + # (128, 2112, 7168), + (128, 24576, 1536), + (128, 32768, 512), + (128, 7168, 16384), + (128, 4096, 7168), + (128, 7168, 2048), + # (4096, 2112, 7168), + (4096, 24576, 1536), + (4096, 32768, 512), + (4096, 7168, 16384), + (4096, 4096, 7168), + (4096, 7168, 2048), + ] + + all_results = [] + for m, n, k in shapes: + result = benchmark_shape(m, n, k, verbose=verbose) + all_results.append(result) + + # Print results in a nicely formatted table + print("\n===== PERFORMANCE COMPARISON =====") + + # Print DeepGEMM table + deepgemm_headers = ["m", "n", "k", "Time (μs)", "TFLOPS", "GB/s"] + deepgemm_rows = [] + for result in all_results: + shape = result["shape"] + impl_data = result["implementations"]["DeepGEMM"] + deepgemm_rows.append([ + shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}", + f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}" + ]) + + print_table(deepgemm_headers, + deepgemm_rows, + title="DeepGEMM Implementation:") + + # Print vLLM Triton table + triton_headers = [ + "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM" + ] + triton_rows = [] + for result in all_results: + shape = result["shape"] + impl_data = result["implementations"]["vLLM Triton"] + speedup = impl_data.get("speedup_vs_deepgemm", 1.0) + triton_rows.append([ + shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}", + f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}", + format_speedup(speedup) + ]) + + print_table(triton_headers, + triton_rows, + title="vLLM Triton Implementation:") + + # Print vLLM CUTLASS table + cutlass_headers = [ + "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM", + "vs Triton" + ] + cutlass_rows = [] + for result in all_results: + shape = result["shape"] + impl_data = result["implementations"]["vLLM CUTLASS"] + vs_deepgemm = impl_data.get("speedup_vs_deepgemm", 1.0) + vs_triton = impl_data.get("speedup_vs_triton", 1.0) + cutlass_rows.append([ + shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}", + f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}", + format_speedup(vs_deepgemm), + format_speedup(vs_triton) + ]) + + print_table(cutlass_headers, + cutlass_rows, + title="vLLM CUTLASS Implementation:") + + # Calculate and print averages + print("\n===== AVERAGE PERFORMANCE =====") + + implementations = ["DeepGEMM", "vLLM Triton", "vLLM CUTLASS"] + avg_metrics = { + impl: { + "tflops": 0, + "gb_s": 0, + "time_ms": 0 + } + for impl in implementations + } + + for result in all_results: + for impl in implementations: + impl_data = result["implementations"][impl] + avg_metrics[impl]["tflops"] += impl_data["tflops"] + avg_metrics[impl]["gb_s"] += impl_data["gb_s"] + avg_metrics[impl]["time_ms"] += impl_data["time_ms"] + + num_shapes = len(all_results) + avg_headers = ["Implementation", "Avg TFLOPS", "Avg GB/s", "Avg Time (ms)"] + avg_rows = [] + + for impl in implementations: + avg_tflops = avg_metrics[impl]["tflops"] / num_shapes + avg_mem_bw = avg_metrics[impl]["gb_s"] / num_shapes + avg_time = avg_metrics[impl]["time_ms"] / num_shapes + avg_rows.append([ + impl, f"{avg_tflops:.2f}", f"{avg_mem_bw:.2f}", f"{avg_time:.2f}" + ]) + + print_table(avg_headers, avg_rows) + + # Calculate average speedups + avg_speedups = { + "DeepGEMM vs vLLM Triton": 0, + "DeepGEMM vs vLLM CUTLASS": 0, + "vLLM CUTLASS vs vLLM Triton": 0 + } + + for result in all_results: + deepgemm_time = result["implementations"]["DeepGEMM"]["time_ms"] + vllm_triton_time = result["implementations"]["vLLM Triton"]["time_ms"] + vllm_cutlass_time = result["implementations"]["vLLM CUTLASS"][ + "time_ms"] + + avg_speedups[ + "DeepGEMM vs vLLM Triton"] += vllm_triton_time / deepgemm_time + avg_speedups[ + "DeepGEMM vs vLLM CUTLASS"] += vllm_cutlass_time / deepgemm_time + avg_speedups[ + "vLLM CUTLASS vs vLLM Triton"] += vllm_triton_time / vllm_cutlass_time + + print("\n===== AVERAGE SPEEDUPS =====") + speedup_headers = ["Comparison", "Speedup"] + speedup_rows = [] + for comparison, total in avg_speedups.items(): + avg_speedup = total / num_shapes + status = "faster" if avg_speedup > 1 else "slower" + speedup_rows.append([comparison, f"{avg_speedup:.2f}x {status}"]) + + print_table(speedup_headers, speedup_rows) + + # Average accuracy comparison + print("\n===== ACCURACY COMPARISON =====") + avg_diff = {impl: 0 for impl in implementations} + + for result in all_results: + for impl in implementations: + avg_diff[impl] += result["implementations"][impl]["diff"][ + "Reference"] + + diff_headers = ["Implementation", "Avg Diff vs Reference"] + diff_rows = [] + for impl in implementations: + diff_rows.append([impl, f"{avg_diff[impl] / num_shapes:.6f}"]) + + print_table(diff_headers, diff_rows) + + +if __name__ == "__main__": + run_benchmarks(verbose=False) diff --git a/online_apiserver_test/benchmarks/kernels/graph_machete_bench.py b/online_apiserver_test/benchmarks/kernels/graph_machete_bench.py new file mode 100644 index 0000000000000000000000000000000000000000..9a4da0ef5a85d0df2178d5e990ad177e3f76bf37 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/graph_machete_bench.py @@ -0,0 +1,64 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import math +import pickle +from collections import defaultdict + +import matplotlib.pyplot as plt +import pandas as pd +import regex as re +import seaborn as sns +from torch.utils.benchmark import Measurement as TMeasurement + +from vllm.utils import FlexibleArgumentParser + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark the latency of processing a single batch of " + "requests till completion." + ) + parser.add_argument("filename", type=str) + + args = parser.parse_args() + + with open(args.filename, "rb") as f: + data = pickle.load(f) + raw_results: list[TMeasurement] = data["results"] + + results = defaultdict(lambda: list()) + for v in raw_results: + result = re.search(r"MKN=\(\d+x(\d+x\d+)\)", v.task_spec.sub_label) + if result is not None: + KN = result.group(1) + else: + raise Exception("MKN not found") + result = re.search(r"MKN=\((\d+)x\d+x\d+\)", v.task_spec.sub_label) + if result is not None: + M = result.group(1) + else: + raise Exception("MKN not found") + + kernel = v.task_spec.description + results[KN].append({"kernel": kernel, "batch_size": M, "median": v.median}) + + rows = int(math.ceil(len(results) / 2)) + fig, axs = plt.subplots(rows, 2, figsize=(12, 5 * rows)) + axs = axs.flatten() + for axs_idx, (shape, data) in enumerate(results.items()): + plt.sca(axs[axs_idx]) + df = pd.DataFrame(data) + sns.lineplot( + data=df, + x="batch_size", + y="median", + hue="kernel", + style="kernel", + markers=True, + dashes=False, + palette="Dark2", + ) + plt.title(f"Shape: {shape}") + plt.ylabel("time (median, s)") + plt.tight_layout() + plt.savefig("graph_machete_bench.pdf") diff --git a/online_apiserver_test/benchmarks/kernels/requirements.txt b/online_apiserver_test/benchmarks/kernels/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..1411a4a0b5ab886adfb744e685d150151ab10023 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/requirements.txt @@ -0,0 +1 @@ +pandas \ No newline at end of file diff --git a/online_apiserver_test/benchmarks/kernels/utils.py b/online_apiserver_test/benchmarks/kernels/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..4bbb36bb4359259944310adb8e1047a02f126dea --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/utils.py @@ -0,0 +1,214 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import dataclasses +from collections.abc import Iterable +from typing import Any, Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement + + +@dataclasses.dataclass +class CudaGraphBenchParams: + num_ops_in_cuda_graph: int + + +@dataclasses.dataclass +class ArgPool: + """ + When some argument of the benchmarking function is annotated with this type, + the benchmarking class (BenchMM) will collapse the argument to a pick a + single value from the given list of values, during function invocation. + For every invocation during a benchmarking run, it will choose a + different value from the list. + """ + + values: Iterable[Any] + + def __getitem__(self, index): + return self.values[index] + + +class Bench: + class ArgsIterator: + def __init__(self, args_list, kwargs_list): + assert len(args_list) == len(kwargs_list) + self.args_list = args_list + self.kwargs_list = kwargs_list + self.n = len(self.args_list) + self.idx = 0 + + def __next__(self): + while True: + yield (self.args_list[self.idx], self.kwargs_list[self.idx]) + self.idx += 1 + self.idx = self.idx % self.n + + def reset(self): + self.idx = 0 + + @property + def n_args(self): + return self.n + + def __init__( + self, + cuda_graph_params: Optional[CudaGraphBenchParams], + label: str, + sub_label: str, + description: str, + fn: Callable, + *args, + **kwargs, + ): + self.cuda_graph_params = cuda_graph_params + self.use_cuda_graph = self.cuda_graph_params is not None + self.label = label + self.sub_label = sub_label + self.description = description + self.fn = fn + + # Process args + self._args = args + self._kwargs = kwargs + self.args_list, self.kwargs_list = self.collapse_argpool(*args, **kwargs) + self.args_iterator = self.ArgsIterator(self.args_list, self.kwargs_list) + + # Cudagraph runner + self.g = None + if self.use_cuda_graph: + self.g = self.get_cuda_graph_runner() + + # benchmark run params + self.min_run_time = 1 + + def collapse_argpool(self, *args, **kwargs): + argpool_args = [arg for arg in args if isinstance(arg, ArgPool)] + [ + arg for arg in kwargs.values() if isinstance(arg, ArgPool) + ] + if len(argpool_args) == 0: + return [args], [kwargs] + + # Make sure all argpools are of the same size + argpool_size = len(argpool_args[0].values) + assert all([argpool_size == len(arg.values) for arg in argpool_args]) + + # create copies of the args + args_list = [] + kwargs_list = [] + for _ in range(argpool_size): + args_list.append(args) + kwargs_list.append(kwargs.copy()) + + for i in range(argpool_size): + # collapse args; Just pick the ith value + args_list[i] = tuple( + [arg[i] if isinstance(arg, ArgPool) else arg for arg in args_list[i]] + ) + + # collapse kwargs + kwargs_i = kwargs_list[i] + arg_pool_keys = [k for k, v in kwargs_i.items() if isinstance(v, ArgPool)] + for k in arg_pool_keys: + # again just pick the ith value + kwargs_i[k] = kwargs_i[k][i] + kwargs_list[i] = kwargs_i + + return args_list, kwargs_list + + def get_cuda_graph_runner(self): + assert self.use_cuda_graph + assert self.args_iterator is not None + + num_graph_ops = self.cuda_graph_params.num_ops_in_cuda_graph + + # warmup + args_it = self.args_iterator.__next__() + for _ in range(2): + args, kwargs = next(args_it) + self.fn(*args, **kwargs) + + self.args_iterator.reset() + args_it = self.args_iterator.__next__() + stream = torch.cuda.Stream() + with torch.cuda.stream(stream): + g = torch.cuda.CUDAGraph() + with torch.cuda.graph(g): + for _ in range(num_graph_ops): + args, kwargs = next(args_it) + self.fn(*args, **kwargs) + return g + + def run_cudagrah(self) -> TMeasurement: + assert self.use_cuda_graph + globals = {"g": self.g} + + return TBenchmark.Timer( + stmt="g.replay()", + globals=globals, + label=( + f"{self.label}" + f" | cugraph {self.cuda_graph_params.num_ops_in_cuda_graph} ops" + ), + sub_label=self.sub_label, + description=self.description, + ).blocked_autorange(min_run_time=self.min_run_time) + + def run_eager(self) -> TMeasurement: + setup = None + stmt = None + globals = None + + has_arg_pool = self.args_iterator.n_args > 1 + if has_arg_pool: + setup = """ + args_iterator.reset() + args_it = args_iterator.__next__() + """ + stmt = """ + args, kwargs = next(args_it) + fn(*args, **kwargs) + """ + globals = {"fn": self.fn, "args_iterator": self.args_iterator} + else: + # no arg pool. Just use the args and kwargs directly + self.args_iterator.reset() + args_it = self.args_iterator.__next__() + args, kwargs = next(args_it) + + setup = "" + stmt = """ + fn(*args, **kwargs) + """ + globals = {"fn": self.fn, "args": args, "kwargs": kwargs} + + return TBenchmark.Timer( + stmt=stmt, + setup=setup, + globals=globals, + label=self.label, + sub_label=self.sub_label, + description=self.description, + ).blocked_autorange(min_run_time=self.min_run_time) + + def run(self) -> TMeasurement: + timer = None + if self.use_cuda_graph: # noqa SIM108 + timer = self.run_cudagrah() + else: + timer = self.run_eager() + if not timer.meets_confidence() or timer.has_warnings: + print("Doesn't meet confidence - re-running bench ...") + return self.run() + return timer + + def __enter__(self): + return self + + def __exit__(self, exc_type, exc_value, traceback): + if exc_type: + print(f"exc type {exc_type}") + print(f"exc value {exc_value}") + print(f"exc traceback {traceback}") diff --git a/online_apiserver_test/benchmarks/kernels/weight_shapes.py b/online_apiserver_test/benchmarks/kernels/weight_shapes.py new file mode 100644 index 0000000000000000000000000000000000000000..a27f02394afbdfa4f3a352a599db349a2601bea5 --- /dev/null +++ b/online_apiserver_test/benchmarks/kernels/weight_shapes.py @@ -0,0 +1,98 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +# Weight Shapes are in the format +# ([K, N], TP_SPLIT_DIM) +# Example: +# A shape of ([14336, 4096], 0) indicates the following GEMM shape, +# - TP1 : K = 14336, N = 4096 +# - TP2 : K = 7168, N = 4096 +# A shape of ([4096, 6144], 1) indicates the following GEMM shape, +# - TP1 : K = 4096, N = 6144 +# - TP4 : K = 4096, N = 1536 + +# TP1 shapes +WEIGHT_SHAPES = { + "mistralai/Mistral-7B-v0.1": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-7b-hf": [ + ([4096, 12288], 1), + ([4096, 4096], 0), + ([4096, 22016], 1), + ([11008, 4096], 0), + ], + "meta-llama/Llama-3-8b": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-13b-hf": [ + ([5120, 15360], 1), + ([5120, 5120], 0), + ([5120, 27648], 1), + ([13824, 5120], 0), + ], + "meta-llama/Llama-2-70b-hf": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 57344], 1), + ([28672, 8192], 0), + ], + "meta-llama/Llama-3.1-405b-hf": [ + ([16384, 18432], 1), + ([16384, 16384], 0), + ([16384, 106496], 1), + ([53248, 16384], 0), + ], + "meta-llama/Llama-3.1-8B-Instruct": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-3.3-70B-Instruct": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 57344], 1), + ([28672, 8192], 0), + ], + "mistralai/Mistral-Large-Instruct-2407": [ + ([12288, 14336], 1), + ([12288, 12288], 0), + ([12288, 57344], 1), + ([28672, 12288], 0), + ], + "Qwen/Qwen2.5-7B-Instruct": [ + ([3584, 4608], 1), + ([3584, 3584], 0), + ([3584, 37888], 1), + ([18944, 3584], 0), + ], + "Qwen/Qwen2.5-32B-Instruct": [ + ([5120, 7168], 1), + ([5120, 5120], 0), + ([5120, 55296], 1), + ([27648, 5120], 0), + ], + "Qwen/Qwen2.5-72B-Instruct": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 59136], 1), + ([29568, 8192], 0), + ], + "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": [ + ([2048, 3072], 1), + ([2048, 4096], 1), + ([2048, 2048], 0), + ([2048, 576], 0), + ([2048, 21888], 1), + ([10944, 2048], 0), + ([2048, 2816], 1), + ([1408, 2048], 0), + ], +} diff --git a/online_apiserver_test/benchmarks/overheads/benchmark_hashing.py b/online_apiserver_test/benchmarks/overheads/benchmark_hashing.py new file mode 100644 index 0000000000000000000000000000000000000000..0957a9c65f06c912ae73588e017602af72ae9772 --- /dev/null +++ b/online_apiserver_test/benchmarks/overheads/benchmark_hashing.py @@ -0,0 +1,64 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import cProfile +import pstats + +from vllm import LLM, SamplingParams +from vllm.utils import FlexibleArgumentParser + +# A very long prompt, total number of tokens is about 15k. +LONG_PROMPT = ["You are an expert in large language models, aren't you?"] * 1000 +LONG_PROMPT = " ".join(LONG_PROMPT) + + +def main(args): + llm = LLM( + model=args.model, + enforce_eager=True, + enable_prefix_caching=True, + tensor_parallel_size=args.tensor_parallel_size, + ) + + sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len) + profiler = cProfile.Profile() + + print("------warm up------") + for i in range(3): + output = llm.generate(LONG_PROMPT, sampling_params) + print(output[0].outputs[0].text) + + print("------start generating------") + for i in range(3): + profiler.runctx( + "llm.generate(LONG_PROMPT, sampling_params)", globals(), locals() + ) + + # analyze the runtime of hashing function + stats = pstats.Stats(profiler) + stats.sort_stats("cumulative") + total_time = 0 + total_calls = 0 + for func in stats.stats: + if "hash_of_block" in func[2]: + total_time = stats.stats[func][3] + total_calls = stats.stats[func][0] + percentage = (total_time / stats.total_tt) * 100 + print( + f"Hashing took {total_time:.2f} seconds,{percentage:.2f}% of the total runtime." + ) + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark the performance of hashing function in" + "automatic prefix caching." + ) + parser.add_argument("--model", type=str, default="lmsys/longchat-7b-16k") + parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1) + parser.add_argument("--output-len", type=int, default=10) + parser.add_argument( + "--enable-prefix-caching", action="store_true", help="enable prefix caching" + ) + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test/benchmarks/pyproject.toml b/online_apiserver_test/benchmarks/pyproject.toml new file mode 100644 index 0000000000000000000000000000000000000000..65b1e09a247e2d74e31b616c38a8544de75ad525 --- /dev/null +++ b/online_apiserver_test/benchmarks/pyproject.toml @@ -0,0 +1,49 @@ +# This local pyproject file is part of the migration from yapf to ruff format. +# It uses the same core rules as the main pyproject.toml file, but with the +# following differences: +# - ruff line length is overridden to 88 +# - deprecated typing ignores (UP006, UP035) have been removed + +[tool.ruff] +line-length = 88 + +[tool.ruff.lint.per-file-ignores] +"vllm/third_party/**" = ["ALL"] +"vllm/version.py" = ["F401"] +"vllm/_version.py" = ["ALL"] + +[tool.ruff.lint] +select = [ + # pycodestyle + "E", + # Pyflakes + "F", + # pyupgrade + "UP", + # flake8-bugbear + "B", + # flake8-simplify + "SIM", + # isort + "I", + # flake8-logging-format + "G", +] +ignore = [ + # star imports + "F405", "F403", + # lambda expression assignment + "E731", + # Loop control variable not used within loop body + "B007", + # f-string format + "UP032", + # Can remove once 3.10+ is the minimum Python version + "UP007", +] + +[tool.ruff.lint.isort] +known-first-party = ["vllm"] + +[tool.ruff.format] +docstring-code-format = true \ No newline at end of file diff --git a/online_apiserver_test/benchmarks/run_structured_output_benchmark.sh b/online_apiserver_test/benchmarks/run_structured_output_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..b043ab83e4608e7b27e2f94e6ec24a05f7c474aa --- /dev/null +++ b/online_apiserver_test/benchmarks/run_structured_output_benchmark.sh @@ -0,0 +1,129 @@ +#!/bin/bash + +# default values +MODEL=${MODEL:-"Qwen/Qwen2.5-7B-Instruct"} +BACKEND=${BACKEND:-"vllm"} +DATASET=${DATASET:-"xgrammar_bench"} +SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" +OUTPUT_DIR=${OUTPUT_DIR:-"$SCRIPT_DIR/structured_output_benchmark_results"} +PORT=${PORT:-8000} +STRUCTURED_OUTPUT_RATIO=${STRUCTURED_OUTPUT_RATIO:-1} +TOTAL_SECONDS=${TOTAL_SECONDS:-90} +MAX_NEW_TOKENS=${MAX_NEW_TOKENS:-300} +TOKENIZER_MODE=${TOKENIZER_MODE:-"auto"} + +usage() { + echo "Usage: $0 [options]" + echo "Options:" + echo " --model MODEL Model to benchmark (default: $MODEL)" + echo " --backend BACKEND Backend to use (default: $BACKEND)" + echo " --dataset DATASET Dataset to use (default: $DATASET)" + echo " --max-new-tokens N Maximum number of tokens to generate (default: $MAX_NEW_TOKENS)" + echo " --output-dir DIR Output directory for results (default: $OUTPUT_DIR)" + echo " --port PORT Port to use (default: $PORT)" + echo " --structured-output-ratio N Ratio of structured outputs (default: $STRUCTURED_OUTPUT_RATIO)" + echo " --tokenizer-mode MODE Tokenizer mode to use (default: $TOKENIZER_MODE)" + echo " --total-seconds N Total seconds to run the benchmark (default: $TOTAL_SECONDS)" + echo " -h, --help Show this help message and exit" + exit 0 +} + +# parse command line arguments +while [[ $# -gt 0 ]]; do + case $1 in + --model) + MODEL="$2" + shift 2 + ;; + --backend) + BACKEND="$2" + shift 2 + ;; + --dataset) + DATASET="$2" + shift 2 + ;; + --max-new-tokens) + MAX_NEW_TOKENS="$2" + shift 2 + ;; + --output-dir) + OUTPUT_DIR="$2" + shift 2 + ;; + --port) + PORT="$2" + shift 2 + ;; + --structured-output-ratio) + STRUCTURED_OUTPUT_RATIO="$2" + shift 2 + ;; + --tokenizer-mode) + TOKENIZER_MODE="$2" + shift 2 + ;; + --total-seconds) + TOTAL_SECONDS="$2" + shift 2 + ;; + -h|--help) + usage + ;; + *) + echo "Unknown argument: $1\n" + usage + ;; + esac +done + +# Create output directory if it doesn't exist +mkdir -p "$OUTPUT_DIR" + +# Define QPS values to test +QPS_VALUES=(25 20 15 10 5 1) + +# Common parameters +COMMON_PARAMS="--backend $BACKEND \ + --model $MODEL \ + --dataset $DATASET \ + --structured-output-ratio $STRUCTURED_OUTPUT_RATIO \ + --save-results \ + --result-dir $OUTPUT_DIR \ + --output-len $MAX_NEW_TOKENS \ + --port $PORT \ + --tokenizer-mode $TOKENIZER_MODE" + +echo "Starting structured output benchmark with model: $MODEL" +echo "Backend: $BACKEND" +echo "Dataset: $DATASET" +echo "Results will be saved to: $OUTPUT_DIR" +echo "----------------------------------------" + +# Run benchmarks with different QPS values +for qps in "${QPS_VALUES[@]}"; do + echo "Running benchmark with QPS: $qps" + + # Get git hash and branch for the filename + GIT_HASH=$(git rev-parse --short HEAD 2>/dev/null || echo "unknown") + GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "unknown") + + # Construct filename for this run + FILENAME="${BACKEND}_${qps}qps_$(basename $MODEL)_${DATASET}_${GIT_HASH}.json" + + NUM_PROMPTS=$(echo "$TOTAL_SECONDS * $qps" | bc) + NUM_PROMPTS=${NUM_PROMPTS%.*} # Remove fractional part + echo "Running benchmark with $NUM_PROMPTS prompts" + + # Run the benchmark + python "$SCRIPT_DIR/benchmark_serving_structured_output.py" $COMMON_PARAMS \ + --request-rate $qps \ + --result-filename "$FILENAME" \ + --num-prompts $NUM_PROMPTS + + echo "Completed benchmark with QPS: $qps" + echo "----------------------------------------" +done + +echo "All benchmarks completed!" +echo "Results saved to: $OUTPUT_DIR" diff --git a/online_apiserver_test/benchmarks/sonnet.txt b/online_apiserver_test/benchmarks/sonnet.txt new file mode 100644 index 0000000000000000000000000000000000000000..34c444e8ce8e2dc701ec80931401c57014ae0bd1 --- /dev/null +++ b/online_apiserver_test/benchmarks/sonnet.txt @@ -0,0 +1,518 @@ +FROM fairest creatures we desire increase, +That thereby beauty's rose might never die, +But as the riper should by time decease, +His tender heir might bear his memory: +But thou, contracted to thine own bright eyes, +Feed'st thy light'st flame with self-substantial fuel, +Making a famine where abundance lies, +Thyself thy foe, to thy sweet self too cruel. +Thou that art now the world's fresh ornament +And only herald to the gaudy spring, +Within thine own bud buriest thy content +And, tender churl, makest waste in niggarding. +Pity the world, or else this glutton be, +To eat the world's due, by the grave and thee. +When forty winters shall beseige thy brow, +And dig deep trenches in thy beauty's field, +Thy youth's proud livery, so gazed on now, +Will be a tatter'd weed, of small worth held: +Then being ask'd where all thy beauty lies, +Where all the treasure of thy lusty days, +To say, within thine own deep-sunken eyes, +Were an all-eating shame and thriftless praise. +How much more praise deserved thy beauty's use, +If thou couldst answer 'This fair child of mine +Shall sum my count and make my old excuse,' +Proving his beauty by succession thine! +This were to be new made when thou art old, +And see thy blood warm when thou feel'st it cold. +Look in thy glass, and tell the face thou viewest +Now is the time that face should form another; +Whose fresh repair if now thou not renewest, +Thou dost beguile the world, unbless some mother. +For where is she so fair whose unear'd womb +Disdains the tillage of thy husbandry? +Or who is he so fond will be the tomb +Of his self-love, to stop posterity? +Thou art thy mother's glass, and she in thee +Calls back the lovely April of her prime: +So thou through windows of thine age shall see +Despite of wrinkles this thy golden time. +But if thou live, remember'd not to be, +Die single, and thine image dies with thee. +Unthrifty loveliness, why dost thou spend +Upon thyself thy beauty's legacy? +Nature's bequest gives nothing but doth lend, +And being frank she lends to those are free. +Then, beauteous niggard, why dost thou abuse +The bounteous largess given thee to give? +Profitless usurer, why dost thou use +So great a sum of sums, yet canst not live? +For having traffic with thyself alone, +Thou of thyself thy sweet self dost deceive. +Then how, when nature calls thee to be gone, +What acceptable audit canst thou leave? +Thy unused beauty must be tomb'd with thee, +Which, used, lives th' executor to be. +Those hours, that with gentle work did frame +The lovely gaze where every eye doth dwell, +Will play the tyrants to the very same +And that unfair which fairly doth excel: +For never-resting time leads summer on +To hideous winter and confounds him there; +Sap cheque'd with frost and lusty leaves quite gone, +Beauty o'ersnow'd and bareness every where: +Then, were not summer's distillation left, +A liquid prisoner pent in walls of glass, +Beauty's effect with beauty were bereft, +Nor it nor no remembrance what it was: +But flowers distill'd though they with winter meet, +Leese but their show; their substance still lives sweet. +Then let not winter's ragged hand deface +In thee thy summer, ere thou be distill'd: +Make sweet some vial; treasure thou some place +With beauty's treasure, ere it be self-kill'd. +That use is not forbidden usury, +Which happies those that pay the willing loan; +That's for thyself to breed another thee, +Or ten times happier, be it ten for one; +Ten times thyself were happier than thou art, +If ten of thine ten times refigured thee: +Then what could death do, if thou shouldst depart, +Leaving thee living in posterity? +Be not self-will'd, for thou art much too fair +To be death's conquest and make worms thine heir. +Lo! in the orient when the gracious light +Lifts up his burning head, each under eye +Doth homage to his new-appearing sight, +Serving with looks his sacred majesty; +And having climb'd the steep-up heavenly hill, +Resembling strong youth in his middle age, +yet mortal looks adore his beauty still, +Attending on his golden pilgrimage; +But when from highmost pitch, with weary car, +Like feeble age, he reeleth from the day, +The eyes, 'fore duteous, now converted are +From his low tract and look another way: +So thou, thyself out-going in thy noon, +Unlook'd on diest, unless thou get a son. +Music to hear, why hear'st thou music sadly? +Sweets with sweets war not, joy delights in joy. +Why lovest thou that which thou receivest not gladly, +Or else receivest with pleasure thine annoy? +If the true concord of well-tuned sounds, +By unions married, do offend thine ear, +They do but sweetly chide thee, who confounds +In singleness the parts that thou shouldst bear. +Mark how one string, sweet husband to another, +Strikes each in each by mutual ordering, +Resembling sire and child and happy mother +Who all in one, one pleasing note do sing: +Whose speechless song, being many, seeming one, +Sings this to thee: 'thou single wilt prove none.' +Is it for fear to wet a widow's eye +That thou consumest thyself in single life? +Ah! if thou issueless shalt hap to die. +The world will wail thee, like a makeless wife; +The world will be thy widow and still weep +That thou no form of thee hast left behind, +When every private widow well may keep +By children's eyes her husband's shape in mind. +Look, what an unthrift in the world doth spend +Shifts but his place, for still the world enjoys it; +But beauty's waste hath in the world an end, +And kept unused, the user so destroys it. +No love toward others in that bosom sits +That on himself such murderous shame commits. +For shame! deny that thou bear'st love to any, +Who for thyself art so unprovident. +Grant, if thou wilt, thou art beloved of many, +But that thou none lovest is most evident; +For thou art so possess'd with murderous hate +That 'gainst thyself thou stick'st not to conspire. +Seeking that beauteous roof to ruinate +Which to repair should be thy chief desire. +O, change thy thought, that I may change my mind! +Shall hate be fairer lodged than gentle love? +Be, as thy presence is, gracious and kind, +Or to thyself at least kind-hearted prove: +Make thee another self, for love of me, +That beauty still may live in thine or thee. +As fast as thou shalt wane, so fast thou growest +In one of thine, from that which thou departest; +And that fresh blood which youngly thou bestowest +Thou mayst call thine when thou from youth convertest. +Herein lives wisdom, beauty and increase: +Without this, folly, age and cold decay: +If all were minded so, the times should cease +And threescore year would make the world away. +Let those whom Nature hath not made for store, +Harsh featureless and rude, barrenly perish: +Look, whom she best endow'd she gave the more; +Which bounteous gift thou shouldst in bounty cherish: +She carved thee for her seal, and meant thereby +Thou shouldst print more, not let that copy die. +When I do count the clock that tells the time, +And see the brave day sunk in hideous night; +When I behold the violet past prime, +And sable curls all silver'd o'er with white; +When lofty trees I see barren of leaves +Which erst from heat did canopy the herd, +And summer's green all girded up in sheaves +Borne on the bier with white and bristly beard, +Then of thy beauty do I question make, +That thou among the wastes of time must go, +Since sweets and beauties do themselves forsake +And die as fast as they see others grow; +And nothing 'gainst Time's scythe can make defence +Save breed, to brave him when he takes thee hence. +O, that you were yourself! but, love, you are +No longer yours than you yourself here live: +Against this coming end you should prepare, +And your sweet semblance to some other give. +So should that beauty which you hold in lease +Find no determination: then you were +Yourself again after yourself's decease, +When your sweet issue your sweet form should bear. +Who lets so fair a house fall to decay, +Which husbandry in honour might uphold +Against the stormy gusts of winter's day +And barren rage of death's eternal cold? +O, none but unthrifts! Dear my love, you know +You had a father: let your son say so. +Not from the stars do I my judgment pluck; +And yet methinks I have astronomy, +But not to tell of good or evil luck, +Of plagues, of dearths, or seasons' quality; +Nor can I fortune to brief minutes tell, +Pointing to each his thunder, rain and wind, +Or say with princes if it shall go well, +By oft predict that I in heaven find: +But from thine eyes my knowledge I derive, +And, constant stars, in them I read such art +As truth and beauty shall together thrive, +If from thyself to store thou wouldst convert; +Or else of thee this I prognosticate: +Thy end is truth's and beauty's doom and date. +When I consider every thing that grows +Holds in perfection but a little moment, +That this huge stage presenteth nought but shows +Whereon the stars in secret influence comment; +When I perceive that men as plants increase, +Cheered and cheque'd even by the self-same sky, +Vaunt in their youthful sap, at height decrease, +And wear their brave state out of memory; +Then the conceit of this inconstant stay +Sets you most rich in youth before my sight, +Where wasteful Time debateth with Decay, +To change your day of youth to sullied night; +And all in war with Time for love of you, +As he takes from you, I engraft you new. +But wherefore do not you a mightier way +Make war upon this bloody tyrant, Time? +And fortify yourself in your decay +With means more blessed than my barren rhyme? +Now stand you on the top of happy hours, +And many maiden gardens yet unset +With virtuous wish would bear your living flowers, +Much liker than your painted counterfeit: +So should the lines of life that life repair, +Which this, Time's pencil, or my pupil pen, +Neither in inward worth nor outward fair, +Can make you live yourself in eyes of men. +To give away yourself keeps yourself still, +And you must live, drawn by your own sweet skill. +Who will believe my verse in time to come, +If it were fill'd with your most high deserts? +Though yet, heaven knows, it is but as a tomb +Which hides your life and shows not half your parts. +If I could write the beauty of your eyes +And in fresh numbers number all your graces, +The age to come would say 'This poet lies: +Such heavenly touches ne'er touch'd earthly faces.' +So should my papers yellow'd with their age +Be scorn'd like old men of less truth than tongue, +And your true rights be term'd a poet's rage +And stretched metre of an antique song: +But were some child of yours alive that time, +You should live twice; in it and in my rhyme. +Shall I compare thee to a summer's day? +Thou art more lovely and more temperate: +Rough winds do shake the darling buds of May, +And summer's lease hath all too short a date: +Sometime too hot the eye of heaven shines, +And often is his gold complexion dimm'd; +And every fair from fair sometime declines, +By chance or nature's changing course untrimm'd; +But thy eternal summer shall not fade +Nor lose possession of that fair thou owest; +Nor shall Death brag thou wander'st in his shade, +When in eternal lines to time thou growest: +So long as men can breathe or eyes can see, +So long lives this and this gives life to thee. +Devouring Time, blunt thou the lion's paws, +And make the earth devour her own sweet brood; +Pluck the keen teeth from the fierce tiger's jaws, +And burn the long-lived phoenix in her blood; +Make glad and sorry seasons as thou fleets, +And do whate'er thou wilt, swift-footed Time, +To the wide world and all her fading sweets; +But I forbid thee one most heinous crime: +O, carve not with thy hours my love's fair brow, +Nor draw no lines there with thine antique pen; +Him in thy course untainted do allow +For beauty's pattern to succeeding men. +Yet, do thy worst, old Time: despite thy wrong, +My love shall in my verse ever live young. +A woman's face with Nature's own hand painted +Hast thou, the master-mistress of my passion; +A woman's gentle heart, but not acquainted +With shifting change, as is false women's fashion; +An eye more bright than theirs, less false in rolling, +Gilding the object whereupon it gazeth; +A man in hue, all 'hues' in his controlling, +Much steals men's eyes and women's souls amazeth. +And for a woman wert thou first created; +Till Nature, as she wrought thee, fell a-doting, +And by addition me of thee defeated, +By adding one thing to my purpose nothing. +But since she prick'd thee out for women's pleasure, +Mine be thy love and thy love's use their treasure. +So is it not with me as with that Muse +Stirr'd by a painted beauty to his verse, +Who heaven itself for ornament doth use +And every fair with his fair doth rehearse +Making a couplement of proud compare, +With sun and moon, with earth and sea's rich gems, +With April's first-born flowers, and all things rare +That heaven's air in this huge rondure hems. +O' let me, true in love, but truly write, +And then believe me, my love is as fair +As any mother's child, though not so bright +As those gold candles fix'd in heaven's air: +Let them say more than like of hearsay well; +I will not praise that purpose not to sell. +My glass shall not persuade me I am old, +So long as youth and thou are of one date; +But when in thee time's furrows I behold, +Then look I death my days should expiate. +For all that beauty that doth cover thee +Is but the seemly raiment of my heart, +Which in thy breast doth live, as thine in me: +How can I then be elder than thou art? +O, therefore, love, be of thyself so wary +As I, not for myself, but for thee will; +Bearing thy heart, which I will keep so chary +As tender nurse her babe from faring ill. +Presume not on thy heart when mine is slain; +Thou gavest me thine, not to give back again. +As an unperfect actor on the stage +Who with his fear is put besides his part, +Or some fierce thing replete with too much rage, +Whose strength's abundance weakens his own heart. +So I, for fear of trust, forget to say +The perfect ceremony of love's rite, +And in mine own love's strength seem to decay, +O'ercharged with burden of mine own love's might. +O, let my books be then the eloquence +And dumb presagers of my speaking breast, +Who plead for love and look for recompense +More than that tongue that more hath more express'd. +O, learn to read what silent love hath writ: +To hear with eyes belongs to love's fine wit. +Mine eye hath play'd the painter and hath stell'd +Thy beauty's form in table of my heart; +My body is the frame wherein 'tis held, +And perspective it is the painter's art. +For through the painter must you see his skill, +To find where your true image pictured lies; +Which in my bosom's shop is hanging still, +That hath his windows glazed with thine eyes. +Now see what good turns eyes for eyes have done: +Mine eyes have drawn thy shape, and thine for me +Are windows to my breast, where-through the sun +Delights to peep, to gaze therein on thee; +Yet eyes this cunning want to grace their art; +They draw but what they see, know not the heart. +Let those who are in favour with their stars +Of public honour and proud titles boast, +Whilst I, whom fortune of such triumph bars, +Unlook'd for joy in that I honour most. +Great princes' favourites their fair leaves spread +But as the marigold at the sun's eye, +And in themselves their pride lies buried, +For at a frown they in their glory die. +The painful warrior famoused for fight, +After a thousand victories once foil'd, +Is from the book of honour razed quite, +And all the rest forgot for which he toil'd: +Then happy I, that love and am beloved +Where I may not remove nor be removed. +Lord of my love, to whom in vassalage +Thy merit hath my duty strongly knit, +To thee I send this written embassage, +To witness duty, not to show my wit: +Duty so great, which wit so poor as mine +May make seem bare, in wanting words to show it, +But that I hope some good conceit of thine +In thy soul's thought, all naked, will bestow it; +Till whatsoever star that guides my moving +Points on me graciously with fair aspect +And puts apparel on my tatter'd loving, +To show me worthy of thy sweet respect: +Then may I dare to boast how I do love thee; +Till then not show my head where thou mayst prove me. +Weary with toil, I haste me to my bed, +The dear repose for limbs with travel tired; +But then begins a journey in my head, +To work my mind, when body's work's expired: +For then my thoughts, from far where I abide, +Intend a zealous pilgrimage to thee, +And keep my drooping eyelids open wide, +Looking on darkness which the blind do see +Save that my soul's imaginary sight +Presents thy shadow to my sightless view, +Which, like a jewel hung in ghastly night, +Makes black night beauteous and her old face new. +Lo! thus, by day my limbs, by night my mind, +For thee and for myself no quiet find. +How can I then return in happy plight, +That am debarr'd the benefit of rest? +When day's oppression is not eased by night, +But day by night, and night by day, oppress'd? +And each, though enemies to either's reign, +Do in consent shake hands to torture me; +The one by toil, the other to complain +How far I toil, still farther off from thee. +I tell the day, to please them thou art bright +And dost him grace when clouds do blot the heaven: +So flatter I the swart-complexion'd night, +When sparkling stars twire not thou gild'st the even. +But day doth daily draw my sorrows longer +And night doth nightly make grief's strength seem stronger. +When, in disgrace with fortune and men's eyes, +I all alone beweep my outcast state +And trouble deal heaven with my bootless cries +And look upon myself and curse my fate, +Wishing me like to one more rich in hope, +Featured like him, like him with friends possess'd, +Desiring this man's art and that man's scope, +With what I most enjoy contented least; +Yet in these thoughts myself almost despising, +Haply I think on thee, and then my state, +Like to the lark at break of day arising +From sullen earth, sings hymns at heaven's gate; +For thy sweet love remember'd such wealth brings +That then I scorn to change my state with kings. +When to the sessions of sweet silent thought +I summon up remembrance of things past, +I sigh the lack of many a thing I sought, +And with old woes new wail my dear time's waste: +Then can I drown an eye, unused to flow, +For precious friends hid in death's dateless night, +And weep afresh love's long since cancell'd woe, +And moan the expense of many a vanish'd sight: +Then can I grieve at grievances foregone, +And heavily from woe to woe tell o'er +The sad account of fore-bemoaned moan, +Which I new pay as if not paid before. +But if the while I think on thee, dear friend, +All losses are restored and sorrows end. +Thy bosom is endeared with all hearts, +Which I by lacking have supposed dead, +And there reigns love and all love's loving parts, +And all those friends which I thought buried. +How many a holy and obsequious tear +Hath dear religious love stol'n from mine eye +As interest of the dead, which now appear +But things removed that hidden in thee lie! +Thou art the grave where buried love doth live, +Hung with the trophies of my lovers gone, +Who all their parts of me to thee did give; +That due of many now is thine alone: +Their images I loved I view in thee, +And thou, all they, hast all the all of me. +If thou survive my well-contented day, +When that churl Death my bones with dust shall cover, +And shalt by fortune once more re-survey +These poor rude lines of thy deceased lover, +Compare them with the bettering of the time, +And though they be outstripp'd by every pen, +Reserve them for my love, not for their rhyme, +Exceeded by the height of happier men. +O, then vouchsafe me but this loving thought: +'Had my friend's Muse grown with this growing age, +A dearer birth than this his love had brought, +To march in ranks of better equipage: +But since he died and poets better prove, +Theirs for their style I'll read, his for his love.' +Full many a glorious morning have I seen +Flatter the mountain-tops with sovereign eye, +Kissing with golden face the meadows green, +Gilding pale streams with heavenly alchemy; +Anon permit the basest clouds to ride +With ugly rack on his celestial face, +And from the forlorn world his visage hide, +Stealing unseen to west with this disgrace: +Even so my sun one early morn did shine +With all triumphant splendor on my brow; +But out, alack! he was but one hour mine; +The region cloud hath mask'd him from me now. +Yet him for this my love no whit disdaineth; +Suns of the world may stain when heaven's sun staineth. +Why didst thou promise such a beauteous day, +And make me travel forth without my cloak, +To let base clouds o'ertake me in my way, +Hiding thy bravery in their rotten smoke? +'Tis not enough that through the cloud thou break, +To dry the rain on my storm-beaten face, +For no man well of such a salve can speak +That heals the wound and cures not the disgrace: +Nor can thy shame give physic to my grief; +Though thou repent, yet I have still the loss: +The offender's sorrow lends but weak relief +To him that bears the strong offence's cross. +Ah! but those tears are pearl which thy love sheds, +And they are rich and ransom all ill deeds. +No more be grieved at that which thou hast done: +Roses have thorns, and silver fountains mud; +Clouds and eclipses stain both moon and sun, +And loathsome canker lives in sweetest bud. +All men make faults, and even I in this, +Authorizing thy trespass with compare, +Myself corrupting, salving thy amiss, +Excusing thy sins more than thy sins are; +For to thy sensual fault I bring in sense-- +Thy adverse party is thy advocate-- +And 'gainst myself a lawful plea commence: +Such civil war is in my love and hate +That I an accessary needs must be +To that sweet thief which sourly robs from me. +Let me confess that we two must be twain, +Although our undivided loves are one: +So shall those blots that do with me remain +Without thy help by me be borne alone. +In our two loves there is but one respect, +Though in our lives a separable spite, +Which though it alter not love's sole effect, +Yet doth it steal sweet hours from love's delight. +I may not evermore acknowledge thee, +Lest my bewailed guilt should do thee shame, +Nor thou with public kindness honour me, +Unless thou take that honour from thy name: +But do not so; I love thee in such sort +As, thou being mine, mine is thy good report. +As a decrepit father takes delight +To see his active child do deeds of youth, +So I, made lame by fortune's dearest spite, +Take all my comfort of thy worth and truth. +For whether beauty, birth, or wealth, or wit, +Or any of these all, or all, or more, +Entitled in thy parts do crowned sit, +I make my love engrafted to this store: +So then I am not lame, poor, nor despised, +Whilst that this shadow doth such substance give +That I in thy abundance am sufficed +And by a part of all thy glory live. +Look, what is best, that best I wish in thee: +This wish I have; then ten times happy me! \ No newline at end of file diff --git a/online_apiserver_test/benchmarks/structured_schemas/structured_schema_1.json b/online_apiserver_test/benchmarks/structured_schemas/structured_schema_1.json new file mode 100644 index 0000000000000000000000000000000000000000..13bd6b6d16c601e7e589952b49119205a407acd1 --- /dev/null +++ b/online_apiserver_test/benchmarks/structured_schemas/structured_schema_1.json @@ -0,0 +1,19 @@ +{ + "type": "object", + "properties": { + "name": { "type": "string" }, + "email": { "type": "string" }, + "street": { "type": "string" }, + "city": { "type": "string" }, + "state": { "type": "string" }, + "zip": { "type": "string" }, + "phone": { "type": "string" }, + "website": { "type": "string" }, + "company": { "type": "string" }, + "age": { "type": "integer" } + }, + "required": [ + "name", + "email" + ] +} diff --git a/online_apiserver_test/online_quick_check.sh b/online_apiserver_test/online_quick_check.sh new file mode 100644 index 0000000000000000000000000000000000000000..92dc8a7707631c99f1dab6b4068e87343c085109 --- /dev/null +++ b/online_apiserver_test/online_quick_check.sh @@ -0,0 +1,94 @@ +#!/bin/bash + +HOST=127.0.0.1 +PORT=8081 +TIMEOUT=1200 # 监控超时设置 +INTERVAL=60 # 监控时间间隔 +ENDPOINT=/v1/completions + +# 读取json配置文件 +json_data=$(cat auto_quick_check_config.json) + +DCU=$(echo $json_data | jq -r '.DCU') +vllm_version=$(echo $json_data | jq -r '.vllm_version') +pkg_version=$(echo $json_data | jq -r '.pkg_version') +dst_path=$(echo $json_data | jq -r '.dst_path') +items=$(echo $json_data | jq -c '.items[]') + +while read -r item; do + model_name=$(echo "$item" | jq -r '.model_name') + model_path=$(echo "$item" | jq -r '.model_path') + dtype=$(echo "$item" | jq -r '.dtype') + tensor_parallel=$(echo "$item" | jq -r '.tensor_parallel') + + batch_size=$(echo "$item" | jq -r '.batch_size') + seqlen_tuple=$(echo "$item" | jq -r '.seqlen_tuple') + + result_path=${dst_path}/${model_name}/ + if [ ! -f ${result_path} ]; then + mkdir ${result_path} -p + fi + + if [ -e "${result_path}output.csv" ] && [ -s "${result_path}output.csv" ]; then + : + else + echo "model_name,DCU,DCU nums,precision,input_len,output_len,bs,TTFT_mean(ms),TPOT_mean(ms),ITL_mean(ms),GenerateThroughput(tokens/s),TotalThroughput(tokens/s),Duration(s),version" > ${result_path}output.csv + fi + + echo $tensor_parallel | jq -c '.[]' | while read -r tp; do + # 运行服务端启动脚本 + + nohup bash run_apiserver.sh $model_name $model_path $tp $dtype $HOST $PORT $result_path & + + start_time=$(date +%s) + while true; do + if nc -zv $HOST $PORT; then # 检查端口是否打开 + echo $seqlen_tuple | jq -c '.[]' | while read -r sq; do + echo $batch_size | jq -c '.[]' | while read -r bs; do + IFS=' ' read -r input_len output_len <<< ${sq//\"/} + + # 运行评测脚本 + python3 benchmarks/benchmark_serving.py \ + --dataset-name random \ + --tokenizer ${model_path} \ + --trust-remote-code \ + --model ${model_name} \ + --port ${PORT} \ + --endpoint ${ENDPOINT} \ + --random-input-len ${input_len} \ + --random-output-len ${output_len} \ + --ignore_eos \ + --num-prompts ${bs} \ + --max-concurrency ${bs} 2>&1 | tee ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-bs-${bs}.log + # --num-prompts $((4*${bs})) \ + + output_throughput=`grep -a "^Output token throughput (tok/s):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-bs-${bs}.log | awk -F ' ' '{print $5}'` + total_throughput=`grep -a "^Total Token throughput (tok/s):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-bs-${bs}.log | awk -F ' ' '{print $5}'` + TTFT_mean=`grep -a "^Mean TTFT (ms):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-bs-${bs}.log | awk -F ' ' '{print $4}'` + TPOT_mean=`grep -a "^Mean TPOT (ms):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-bs-${bs}.log | awk -F ' ' '{print $4}'` + ITL_mean=`grep -a "^Mean ITL (ms):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-bs-${bs}.log | awk -F ' ' '{print $4}'` + duration=`grep -a "^Benchmark duration (s):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-bs-${bs}.log | awk -F ' ' '{print $4}'` + # total_time=$(python -c "print(${total_time} / 4)") + + echo "$model_name,$DCU,$tp,$dtype,$input_len,$output_len,$bs,$TTFT_mean,$TPOT_mean,$ITL_mean,$output_throughput,$total_throughput,$duration,$pkg_version" >> ${result_path}output.csv + sleep 10 + done + done + break + else + current_time=$(date +%s) + elapsed_time=$((current_time - start_time)) + if [ $elapsed_time -ge $TIMEOUT ]; then + echo "ERR：PORT ${PORT} launch time out, exit!!!。" + exit 1 + fi + echo "PORT ${PORT} has not been launched yet, please wait...." + sleep $INTERVAL + fi + done + + pkill -f vllm + sleep 60 + done + +done <<< "$items" diff --git a/online_apiserver_test/run_apiserver.sh b/online_apiserver_test/run_apiserver.sh new file mode 100644 index 0000000000000000000000000000000000000000..e580e6a2e8d6465c7cec657fa0ff7bdaaf77e3c8 --- /dev/null +++ b/online_apiserver_test/run_apiserver.sh @@ -0,0 +1,69 @@ +#!/bin/bash + +## 产品部提供的vllm 0.9.2默认环境变量，用于测试非DS 671B模型 +export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +export ALLREDUCE_STREAM_WITH_COMPUTE=1 +export NCCL_MIN_NCHANNELS=16 +export NCCL_MAX_NCHANNELS=16 +# export VLLM_PCIE_USE_CUSTOM_ALLREDUCE=1 #K100-AI需要 + +## 根据TOPO进行NUMA绑核 +export VLLM_NUMA_BIND=1 +export VLLM_RANK0_NUMA=0 +export VLLM_RANK1_NUMA=0 +export VLLM_RANK2_NUMA=0 +export VLLM_RANK3_NUMA=0 +export VLLM_RANK4_NUMA=0 +export VLLM_RANK5_NUMA=0 +export VLLM_RANK6_NUMA=0 +export VLLM_RANK7_NUMA=0 + +## 项目上反馈的可能对性能有提升的环境变量，可酌情使用 +# export HSA_FORCE_FINE_GRAIN_PCIE=1 +# export NCCL_P2P_LEVEL=SYS +# export NCCL_LAUNCH_MODE=GROUP +# export VLLM_RPC_TIMEOUT=1800000 +# export VLLM_SPEC_DECODE_EAGER=1 +# export VLLM_ENFORCE_EAGER_BS_THRESHOLD=44 # K100AI设置为44，BW1000去掉该环境变量 + +# export VLLM_MLA_DISABLE=0 +# export VLLM_USE_FLASH_MLA=1 +# export VLLM_ZERO_OVERHEAD=1 # 该参数对某些模型性能有明显提升；但在某些环境下表现异常，根据实际情况斟酌选用；且使用后对Qwen3模型会关闭其thinking功能 + +# export W8A8_SUPPORT_METHODS=3 # 对W8A8的量化模型有提升 +# export ROCBLAS_INT8_ENABLE=0 # 对W8A8的量化模型有提升 + +# export VLLM_USE_FLASH_ATTN_PA=0 # 解决精度乱码问题 + +# 禁止生成core文件 +ulimit -c 0 + +# 模型参数 +model_name=$1 +model_path=$2 +tp=$3 +dtype=$4 +host=$5 +port=$6 +dst_path=$7 +case $dtype in + "float16" | "bfloat16") + extra="--dtype ${dtype}" ;; + "gptq-int8" | "gptq-int4") + export GPTQ_CK_GEMMBS=15000 + extra="--quantization gptq" ;; + "compressed-tensors-w8a8") + extra="--quantization compressed-tensors" ;; + "awq") + export AWQ_CK_GEMMBS=15000 + extra="--quantization awq" ;; +esac + +# vllm 0.9.2 默认使用V1 Engine及Prefix Caching功能开启（推荐），如果有测试需求可以自行切换V0 Engine或关闭Prefix Caching +# 使用V0 Engine：VLLM_USE_V1=0 +# 开启/关闭Prefix Caching：--enable-chunked-prefill / --no-enable-prefix-caching + +date=$(date +"%Y-%m-%d-%H-%M-%S") +python3 -m vllm.entrypoints.openai.api_server --served-model-name $model_name --gpu-memory-utilization 0.95 $extra \ + --model $model_path --tensor-parallel-size $tp --host $host --port $port --max-model-len 32768 \ + 2>&1 | tee ${dst_path}/launcher_${model_name}_tp_${tp}_dtype_${dtype}_${date}.log diff --git a/online_apiserver_test_maxbs/README.md b/online_apiserver_test_maxbs/README.md new file mode 100644 index 0000000000000000000000000000000000000000..f45a82a68b1b9b2cbeda40a04fad0b2322e766fe --- /dev/null +++ b/online_apiserver_test_maxbs/README.md @@ -0,0 +1,60 @@ +# Online API Server Benchmark Scripts + +vLLM 0.9.2 在线评测测试的代码/脚本管理 + +## 更新日志 + +2025/10/16：新增vllm0.9.2在线测试的自动化测试脚本（by 刘煦） + +## 使用方法 + +1. 需要安装jq和netcat +``` + apt-get install jq + apt-get install netcat +``` + +2. 文件说明： + + 1）auto_quick_check_config.json：配置文件 + + 2）online_quick_check_maxbs.sh：自动化启动脚本 + + 3）run_apiserver.sh：运行服务的脚本 + + 4）benchmarks：评测部分的代码，来源为vllm 0.9.2的GitHub + +``` +# 配置文件说明 auto_quick_check_config.json +{ + "DCU": "BW1000", # DCU型号 + "vllm_version": "0.9.2", # vllm版本号 + "pkg_version": "dtk25.04.1", # DTK版本号 + "dst_path": "./result/", # 结果/日志存放目录 + "items":[ + { + "model_name": "Qwen3-32B", # 模型名称 + "model_path": "/data/models/Qwen3-32B/", # 模型路径 + "dtype": "float16", # 数据类型 + "tensor_parallel": [4], # 使用卡数 + "batch_size_start": 4, # 起始并发数 + "batch_size_interval": 4, # 并发数增加间隔 + "seqlen_tuple": ["512 512", "1024 1024"], # 输入输出序列组合 + "ttft_thres": [3000, 3000], # ttft阈值，单位毫秒（"inf"表示无限制） + "tpot_thres": [100, "inf"] # tpot阈值，单位毫秒（"inf"表示无限制） + } + ] +} + +注： +1. 【输入输出序列组合】、【ttft阈值】、【tpot阈值】是一一对应的关系，数量需要确保一致 +2. 【ttft阈值】、【tpot阈值】不可均设置成"inf" +``` + +3. 运行及逻辑： + + 指令：bash online_quick_check_maxbs.sh + + 逻辑：调用run_apiserver.sh启动服务 -> 监听端口是否启动 -> 启动后运行评测指令评测 -> 汇总数据 + + 最大并发数测试逻辑：从batch_size_start开始测试，并且逐步增加batch_size_interval，直到某一个并发数的ttft/tpot超过了设置的阈值停止测试；从output.csv中可以获取最大并发数 diff --git a/online_apiserver_test_maxbs/auto_quick_check_config.json b/online_apiserver_test_maxbs/auto_quick_check_config.json new file mode 100644 index 0000000000000000000000000000000000000000..2f6e3041376f34803f7e2432fc03d4c0aebaaf2f --- /dev/null +++ b/online_apiserver_test_maxbs/auto_quick_check_config.json @@ -0,0 +1,19 @@ +{ + "DCU": "BW1000", + "vllm_version": "0.9.2", + "pkg_version": "dtk25.04.1", + "dst_path": "./result/", + "items":[ + { + "model_name": "Qwen3-32B", + "model_path": "/data/models/Qwen3-32B/", + "dtype": "float16", + "tensor_parallel": [4], + "batch_size_start": 4, + "batch_size_interval": 4, + "seqlen_tuple": ["512 512", "1024 1024"], + "ttft_thres": [3000, 3000], + "tpot_thres": [100, "inf"] + } + ] +} \ No newline at end of file diff --git a/online_apiserver_test_maxbs/benchmarks/README.md b/online_apiserver_test_maxbs/benchmarks/README.md new file mode 100644 index 0000000000000000000000000000000000000000..fb8690d42db98393c118947d7facce057f3cd48f --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/README.md @@ -0,0 +1,605 @@ +# Benchmarking vLLM + +This README guides you through running benchmark tests with the extensive +datasets supported on vLLM. It’s a living document, updated as new features and datasets +become available. + +**Dataset Overview** + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
DatasetOnlineOfflineData Path
ShareGPT✅✅wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
BurstGPT✅✅wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv
Sonnet✅✅Local file: benchmarks/sonnet.txt
Random✅✅synthetic
HuggingFace-VisionArena✅✅lmarena-ai/VisionArena-Chat
HuggingFace-InstructCoder✅✅likaixin/InstructCoder
HuggingFace-AIMO✅✅AI-MO/aimo-validation-aime , AI-MO/NuminaMath-1.5, AI-MO/NuminaMath-CoT
HuggingFace-Other✅✅lmms-lab/LLaVA-OneVision-Data, Aeala/ShareGPT_Vicuna_unfiltered
Custom✅✅Local file: data.jsonl
+ +✅: supported + +🟡: Partial support + +🚧: to be supported + +**Note**: HuggingFace dataset's `dataset-name` should be set to `hf` + +--- +
+🚀 Example - Online Benchmark + +
+ +First start serving your model + +```bash +vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests +``` + +Then run the benchmarking script + +```bash +# download dataset +# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json +python3 vllm/benchmarks/benchmark_serving.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --endpoint /v1/completions \ + --dataset-name sharegpt \ + --dataset-path /ShareGPT_V3_unfiltered_cleaned_split.json \ + --num-prompts 10 +``` + +If successful, you will see the following output + +``` +============ Serving Benchmark Result ============ +Successful requests: 10 +Benchmark duration (s): 5.78 +Total input tokens: 1369 +Total generated tokens: 2212 +Request throughput (req/s): 1.73 +Output token throughput (tok/s): 382.89 +Total Token throughput (tok/s): 619.85 +---------------Time to First Token---------------- +Mean TTFT (ms): 71.54 +Median TTFT (ms): 73.88 +P99 TTFT (ms): 79.49 +-----Time per Output Token (excl. 1st token)------ +Mean TPOT (ms): 7.91 +Median TPOT (ms): 7.96 +P99 TPOT (ms): 8.03 +---------------Inter-token Latency---------------- +Mean ITL (ms): 7.74 +Median ITL (ms): 7.70 +P99 ITL (ms): 8.39 +================================================== +``` + +**Custom Dataset** + +If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl + +``` +{"prompt": "What is the capital of India?"} +{"prompt": "What is the capital of Iran?"} +{"prompt": "What is the capital of China?"} +``` + +```bash +# start server +VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct --disable-log-requests +``` + +```bash +# run benchmarking script +python3 benchmarks/benchmark_serving.py --port 9001 --save-result --save-detailed \ + --backend vllm \ + --model meta-llama/Llama-3.1-8B-Instruct \ + --endpoint /v1/completions \ + --dataset-name custom \ + --dataset-path \ + --custom-skip-chat-template \ + --num-prompts 80 \ + --max-concurrency 1 \ + --temperature=0.3 \ + --top-p=0.75 \ + --result-dir "./log/" +``` + +You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`. + +**VisionArena Benchmark for Vision Language Models** + +```bash +# need a model with vision capability here +vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests +``` + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend openai-chat \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --endpoint /v1/chat/completions \ + --dataset-name hf \ + --dataset-path lmarena-ai/VisionArena-Chat \ + --hf-split train \ + --num-prompts 1000 +``` + +**InstructCoder Benchmark with Speculative Decoding** + +``` bash +VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \ + --speculative-config $'{"method": "ngram", + "num_speculative_tokens": 5, "prompt_lookup_max": 5, + "prompt_lookup_min": 2}' +``` + +``` bash +python3 benchmarks/benchmark_serving.py \ + --model meta-llama/Meta-Llama-3-8B-Instruct \ + --dataset-name hf \ + --dataset-path likaixin/InstructCoder \ + --num-prompts 2048 +``` + +**Other HuggingFaceDataset Examples** + +```bash +vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests +``` + +**`lmms-lab/LLaVA-OneVision-Data`** + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend openai-chat \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --endpoint /v1/chat/completions \ + --dataset-name hf \ + --dataset-path lmms-lab/LLaVA-OneVision-Data \ + --hf-split train \ + --hf-subset "chart2text(cauldron)" \ + --num-prompts 10 +``` + +**`Aeala/ShareGPT_Vicuna_unfiltered`** + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend openai-chat \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --endpoint /v1/chat/completions \ + --dataset-name hf \ + --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \ + --hf-split train \ + --num-prompts 10 +``` + +**`AI-MO/aimo-validation-aime`** + +``` bash +python3 vllm/benchmarks/benchmark_serving.py \ + --model Qwen/QwQ-32B \ + --dataset-name hf \ + --dataset-path AI-MO/aimo-validation-aime \ + --num-prompts 10 \ + --seed 42 +``` + +**`philschmid/mt-bench`** + +``` bash +python3 vllm/benchmarks/benchmark_serving.py \ + --model Qwen/QwQ-32B \ + --dataset-name hf \ + --dataset-path philschmid/mt-bench \ + --num-prompts 80 +``` + +**Running With Sampling Parameters** + +When using OpenAI-compatible backends such as `vllm`, optional sampling +parameters can be specified. Example client command: + +```bash +python3 vllm/benchmarks/benchmark_serving.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --endpoint /v1/completions \ + --dataset-name sharegpt \ + --dataset-path /ShareGPT_V3_unfiltered_cleaned_split.json \ + --top-k 10 \ + --top-p 0.9 \ + --temperature 0.5 \ + --num-prompts 10 +``` + +**Running With Ramp-Up Request Rate** + +The benchmark tool also supports ramping up the request rate over the +duration of the benchmark run. This can be useful for stress testing the +server or finding the maximum throughput that it can handle, given some latency budget. + +Two ramp-up strategies are supported: +- `linear`: Increases the request rate linearly from a start value to an end value. +- `exponential`: Increases the request rate exponentially. + +The following arguments can be used to control the ramp-up: +- `--ramp-up-strategy`: The ramp-up strategy to use (`linear` or `exponential`). +- `--ramp-up-start-rps`: The request rate at the beginning of the benchmark. +- `--ramp-up-end-rps`: The request rate at the end of the benchmark. + +
+ +
+📈 Example - Offline Throughput Benchmark + +
+ +```bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset-name sonnet \ + --dataset-path vllm/benchmarks/sonnet.txt \ + --num-prompts 10 +``` + +If successful, you will see the following output + +``` +Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s +Total num prompt tokens: 5014 +Total num output tokens: 1500 +``` + +**VisionArena Benchmark for Vision Language Models** + +``` bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --backend vllm-chat \ + --dataset-name hf \ + --dataset-path lmarena-ai/VisionArena-Chat \ + --num-prompts 1000 \ + --hf-split train +``` + +The `num prompt tokens` now includes image token counts + +``` +Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s +Total num prompt tokens: 14527 +Total num output tokens: 1280 +``` + +**InstructCoder Benchmark with Speculative Decoding** + +``` bash +VLLM_WORKER_MULTIPROC_METHOD=spawn \ +VLLM_USE_V1=1 \ +python3 vllm/benchmarks/benchmark_throughput.py \ + --dataset-name=hf \ + --dataset-path=likaixin/InstructCoder \ + --model=meta-llama/Meta-Llama-3-8B-Instruct \ + --input-len=1000 \ + --output-len=100 \ + --num-prompts=2048 \ + --async-engine \ + --speculative-config $'{"method": "ngram", + "num_speculative_tokens": 5, "prompt_lookup_max": 5, + "prompt_lookup_min": 2}' +``` + +``` +Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s +Total num prompt tokens: 261136 +Total num output tokens: 204800 +``` + +**Other HuggingFaceDataset Examples** + +**`lmms-lab/LLaVA-OneVision-Data`** + +```bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --backend vllm-chat \ + --dataset-name hf \ + --dataset-path lmms-lab/LLaVA-OneVision-Data \ + --hf-split train \ + --hf-subset "chart2text(cauldron)" \ + --num-prompts 10 +``` + +**`Aeala/ShareGPT_Vicuna_unfiltered`** + +```bash +python3 vllm/benchmarks/benchmark_throughput.py \ + --model Qwen/Qwen2-VL-7B-Instruct \ + --backend vllm-chat \ + --dataset-name hf \ + --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \ + --hf-split train \ + --num-prompts 10 +``` + +**`AI-MO/aimo-validation-aime`** + +```bash +python3 benchmarks/benchmark_throughput.py \ + --model Qwen/QwQ-32B \ + --backend vllm \ + --dataset-name hf \ + --dataset-path AI-MO/aimo-validation-aime \ + --hf-split train \ + --num-prompts 10 +``` + +**Benchmark with LoRA Adapters** + +``` bash +# download dataset +# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json +python3 vllm/benchmarks/benchmark_throughput.py \ + --model meta-llama/Llama-2-7b-hf \ + --backend vllm \ + --dataset_path /ShareGPT_V3_unfiltered_cleaned_split.json \ + --dataset_name sharegpt \ + --num-prompts 10 \ + --max-loras 2 \ + --max-lora-rank 8 \ + --enable-lora \ + --lora-path yard1/llama-2-7b-sql-lora-test + ``` + +
+ +
+🛠️ Example - Structured Output Benchmark + +
+ +Benchmark the performance of structured output generation (JSON, grammar, regex). + +**Server Setup** + +```bash +vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests +``` + +**JSON Schema Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset json \ + --structured-output-ratio 1.0 \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**Grammar-based Generation Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset grammar \ + --structure-type grammar \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**Regex-based Generation Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset regex \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**Choice-based Generation Benchmark** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset choice \ + --request-rate 10 \ + --num-prompts 1000 +``` + +**XGrammar Benchmark Dataset** + +```bash +python3 benchmarks/benchmark_serving_structured_output.py \ + --backend vllm \ + --model NousResearch/Hermes-3-Llama-3.1-8B \ + --dataset xgrammar_bench \ + --request-rate 10 \ + --num-prompts 1000 +``` + +
+ +
+📚 Example - Long Document QA Benchmark + +
+ +Benchmark the performance of long document question-answering with prefix caching. + +**Basic Long Document QA Test** + +```bash +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 16 \ + --document-length 2000 \ + --output-len 50 \ + --repeat-count 5 +``` + +**Different Repeat Modes** + +```bash +# Random mode (default) - shuffle prompts randomly +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --document-length 3000 \ + --repeat-count 3 \ + --repeat-mode random + +# Tile mode - repeat entire prompt list in sequence +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --document-length 3000 \ + --repeat-count 3 \ + --repeat-mode tile + +# Interleave mode - repeat each prompt consecutively +python3 benchmarks/benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --document-length 3000 \ + --repeat-count 3 \ + --repeat-mode interleave +``` + +
+ +
+🗂️ Example - Prefix Caching Benchmark + +
+ +Benchmark the efficiency of automatic prefix caching. + +**Fixed Prompt with Prefix Caching** + +```bash +python3 benchmarks/benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-prompts 1 \ + --repeat-count 100 \ + --input-length-range 128:256 +``` + +**ShareGPT Dataset with Prefix Caching** + +```bash +# download dataset +# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json + +python3 benchmarks/benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --dataset-path /path/ShareGPT_V3_unfiltered_cleaned_split.json \ + --enable-prefix-caching \ + --num-prompts 20 \ + --repeat-count 5 \ + --input-length-range 128:256 +``` + +
+ +
+⚡ Example - Request Prioritization Benchmark + +
+ +Benchmark the performance of request prioritization in vLLM. + +**Basic Prioritization Test** + +```bash +python3 benchmarks/benchmark_prioritization.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --input-len 128 \ + --output-len 64 \ + --num-prompts 100 \ + --scheduling-policy priority +``` + +**Multiple Sequences per Prompt** + +```bash +python3 benchmarks/benchmark_prioritization.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --input-len 128 \ + --output-len 64 \ + --num-prompts 100 \ + --scheduling-policy priority \ + --n 2 +``` + +
diff --git a/online_apiserver_test_maxbs/benchmarks/auto_tune.sh b/online_apiserver_test_maxbs/benchmarks/auto_tune.sh new file mode 100644 index 0000000000000000000000000000000000000000..b257b57ce06f5038338bc8a8f114fcb13c0e203b --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/auto_tune.sh @@ -0,0 +1,276 @@ +#!/bin/bash + +# This script aims to tune the best server parameter combinations to maximize throughput for given requirement. +# The current server parameter combination is max_num_seqs and max_num_batched_tokens +# It also supports additional requirement: e2e latency and prefix cache. + +# Pre-requisite: +# 1. Checkout to your branch, install/ update the correct running env. For TPU, activate conda env and install the corresponding torch, xla version. +# 2. If the model is customized, replace the MODEL's config with the customized config. +# 3. Set variables (ALL REQUIRED) +# BASE: your directory for vllm repo +# MODEL: the model served by vllm +# SYSTEM: the hardware, choice TPU or GPU, for other systems, "get best profile" might not support. +# TP: ways of tensor parallelism +# DOWNLOAD_DIR: directory to download and load model weights. +# INPUT_LEN: request input len +# OUTPUT_LEN: request output len +# MIN_CACHE_HIT_PCT: prefix cache rate +# MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000 +# NUM_SEQS_LIST: a list of `max-num-seqs` you want to loop with. +# NUM_BATCHED_TOKENS_LIST: a list of `max-num-batched-tokens` you want to loop with. +# Note that the default NUM_SEQS_LIST and NUM_BATCHED_TOKENS_LIST are set for medium size input/output len, for extra short context (such as 20:20), you might need to include larger numbers in NUM_SEQS_LIST. +# 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens. +# 5. The final result will be saved in RESULT file. + + +# Example use cases +# 1. Given input_len=1800, output_len=20, what's the best max_num_seqs and max_num_batched_tokens to get highest throughput? +# Use INPUT_LEN=1800, OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=100000000000 +# 2. If we have latency requirement to be lower than 500ms, what's the best server parameter? +# Use INPUT_LEN=1800, OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=500 +# 3. If we want to reach 60% prefix cache, what's the best server parameter? +# Use INPUT_LEN=1800, OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=60, MAX_LATENCY_ALLOWED_MS=500 + +TAG=$(date +"%Y_%m_%d_%H_%M") +BASE="" +MODEL="meta-llama/Llama-3.1-8B-Instruct" +SYSTEM="TPU" +TP=1 +DOWNLOAD_DIR="" +INPUT_LEN=4000 +OUTPUT_LEN=16 +MIN_CACHE_HIT_PCT=0 +MAX_LATENCY_ALLOWED_MS=100000000000 +NUM_SEQS_LIST="128 256" +NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096" + +LOG_FOLDER="$BASE/auto-benchmark/$TAG" +RESULT="$LOG_FOLDER/result.txt" +PROFILE_PATH="$LOG_FOLDER/profile" + +echo "result file: $RESULT" +echo "model: $MODEL" + +rm -rf $LOG_FOLDER +rm -rf $PROFILE_PATH +mkdir -p $LOG_FOLDER +mkdir -p $PROFILE_PATH + +cd "$BASE/vllm" + +pip install -q datasets + +current_hash=$(git rev-parse HEAD) +echo "hash:$current_hash" >> "$RESULT" +echo "current_hash: $current_hash" + +best_throughput=0 +best_max_num_seqs=0 +best_num_batched_tokens=0 +best_goodput=0 + +start_server() { + local gpu_memory_utilization=$1 + local max_num_seqs=$2 + local max_num_batched_tokens=$3 + local vllm_log=$4 + local profile_dir=$5 + + pkill -f vllm + + VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \ + --disable-log-requests \ + --port 8004 \ + --gpu-memory-utilization $gpu_memory_utilization \ + --max-num-seqs $max_num_seqs \ + --max-num-batched-tokens $max_num_batched_tokens \ + --tensor-parallel-size $TP \ + --enable-prefix-caching \ + --load-format dummy \ + --download-dir "$DOWNLOAD_DIR" \ + --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 & + + # wait for 10 minutes... + server_started=0 + for i in {1..60}; do + RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout) + STATUS_CODE=$(echo "$RESPONSE" | tail -n 1) + if [[ "$STATUS_CODE" -eq 200 ]]; then + server_started=1 + break + else + sleep 10 + fi + done + if (( ! server_started )); then + echo "server did not start within 10 minutes. Please check server log at $vllm_log". + return 1 + else + return 0 + fi +} + +update_best_profile() { + local profile_dir=$1 + local profile_index=$2 + sorted_paths=($(find "$profile_dir" -maxdepth 1 -not -path "$profile_dir" | sort)) + selected_profile_file= + if [[ "$SYSTEM" == "TPU" ]]; then + selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb" + fi + if [[ "$SYSTEM" == "GPU" ]]; then + selected_profile_file="${sorted_paths[$profile_index]}" + fi + rm -f $PROFILE_PATH/* + cp $selected_profile_file $PROFILE_PATH +} + +run_benchmark() { + local max_num_seqs=$1 + local max_num_batched_tokens=$2 + local gpu_memory_utilization=$3 + echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens" + local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt" + local profile_dir="$LOG_FOLDER/profile_${max_num_seqs}_${max_num_batched_tokens}" + echo "vllm_log: $vllm_log" + echo + rm -f $vllm_log + mkdir -p $profile_dir + pkill -f vllm + local profile_index=0 + + echo "starting server..." + start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log $profile_dir + result=$? + if [[ "$result" -eq 1 ]]; then + echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens" + else + echo "server started." + fi + echo + + echo "run benchmark test..." + meet_latency_requirement=0 + # get a basic qps by using request-rate inf + bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt" + prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 )) + python benchmarks/benchmark_serving.py \ + --backend vllm \ + --model $MODEL \ + --dataset-name random \ + --random-input-len $INPUT_LEN \ + --random-output-len $OUTPUT_LEN \ + --ignore-eos \ + --disable-tqdm \ + --request-rate inf \ + --percentile-metrics ttft,tpot,itl,e2el \ + --goodput e2el:$MAX_LATENCY_ALLOWED_MS \ + --num-prompts 1000 \ + --random-prefix-len $prefix_len \ + --port 8004 \ + --profile &> "$bm_log" + throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}') + goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + + if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then + meet_latency_requirement=1 + request_rate=inf + fi + + if (( ! meet_latency_requirement )); then + # start from request-rate as int(throughput) + 1 + request_rate=$((${throughput%.*} + 1)) + while ((request_rate > 0)); do + profile_index=$((profile_index+1)) + # clear prefix cache + curl -X POST http://0.0.0.0:8004/reset_prefix_cache + sleep 5 + bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt" + python benchmarks/benchmark_serving.py \ + --backend vllm \ + --model $MODEL \ + --dataset-name random \ + --random-input-len $INPUT_LEN \ + --random-output-len $OUTPUT_LEN \ + --ignore-eos \ + --disable-tqdm \ + --request-rate $request_rate \ + --percentile-metrics ttft,tpot,itl,e2el \ + --goodput e2el:$MAX_LATENCY_ALLOWED_MS \ + --num-prompts 100 \ + --random-prefix-len $prefix_len \ + --port 8004 &> "$bm_log" + throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}') + goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g') + if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then + meet_latency_requirement=1 + break + fi + request_rate=$((request_rate-1)) + done + fi + # write the results and update the best result. + if ((meet_latency_requirement)); then + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" >> "$RESULT" + if (( $(echo "$throughput > $best_throughput" | bc -l) )); then + best_throughput=$throughput + best_max_num_seqs=$max_num_seqs + best_num_batched_tokens=$max_num_batched_tokens + best_goodput=$goodput + if [[ "$SYSTEM" == "TPU" ]]; then + update_best_profile "$profile_dir/plugins/profile" $profile_index + fi + if [[ "$SYSTEM" == "GPU" ]]; then + update_best_profile "$profile_dir" $profile_index + fi + fi + else + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" + echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" >> "$RESULT" + fi + + echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput" + + pkill vllm + sleep 10 + printf '=%.0s' $(seq 1 20) + return 0 +} + +read -r -a num_seqs_list <<< "$NUM_SEQS_LIST" +read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST" + +# first find out the max gpu-memory-utilization without HBM OOM. +gpu_memory_utilization=0.98 +find_gpu_memory_utilization=0 +while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do + start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log" + result=$? + if [[ "$result" -eq 0 ]]; then + find_gpu_memory_utilization=1 + break + else + gpu_memory_utilization=$(echo "$gpu_memory_utilization - 0.01" | bc) + fi +done + +if [[ "$find_gpu_memory_utilization" -eq 1 ]]; then + echo "Using gpu_memory_utilization=$gpu_memory_utilization to serve model." +else + echo "Cannot find a proper gpu_memory_utilization over 0.9 to serve the model, please check logs in $LOG_FOLDER." + exit 1 +fi + +for num_seqs in "${num_seqs_list[@]}"; do + for num_batched_tokens in "${num_batched_tokens_list[@]}"; do + run_benchmark $num_seqs $num_batched_tokens $gpu_memory_utilization + done +done +echo "finish permutations" +echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" +echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT" + diff --git a/online_apiserver_test_maxbs/benchmarks/backend_request_func.py b/online_apiserver_test_maxbs/benchmarks/backend_request_func.py new file mode 100644 index 0000000000000000000000000000000000000000..c7229dbb8e90d1167eab8fda1a5fb353e31e81a4 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/backend_request_func.py @@ -0,0 +1,628 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import io +import json +import os +import sys +import time +import traceback +from dataclasses import dataclass, field +from typing import Optional, Union + +import aiohttp +import huggingface_hub.constants +from tqdm.asyncio import tqdm +from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast + +# NOTE(simon): do not import vLLM here so the benchmark script +# can run without vLLM installed. + +AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60) + + +@dataclass +class RequestFuncInput: + prompt: str + api_url: str + prompt_len: int + output_len: int + model: str + model_name: Optional[str] = None + logprobs: Optional[int] = None + extra_body: Optional[dict] = None + multi_modal_content: Optional[dict] = None + ignore_eos: bool = False + language: Optional[str] = None + + +@dataclass +class RequestFuncOutput: + generated_text: str = "" + success: bool = False + latency: float = 0.0 + output_tokens: int = 0 + ttft: float = 0.0 # Time to first token + itl: list[float] = field(default_factory=list) # list of inter-token latencies + tpot: float = 0.0 # avg next-token latencies + prompt_len: int = 0 + error: str = "" + + +async def async_request_tgi( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith("generate_stream") + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + params = { + "max_new_tokens": request_func_input.output_len, + "do_sample": True, + "temperature": 0.01, # TGI does not accept 0.0 temperature. + "top_p": 0.99, # TGI does not accept 1.0 top_p. + "truncate": request_func_input.prompt_len, + "ignore_eos_token": request_func_input.ignore_eos, + } + payload = { + "inputs": request_func_input.prompt, + "parameters": params, + } + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + if request_func_input.ignore_eos: + output.output_tokens = request_func_input.output_len + else: + output.output_tokens = None + + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post(url=api_url, json=payload) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + chunk_bytes = chunk_bytes.decode("utf-8") + + # NOTE: Sometimes TGI returns a ping response without + # any data, we should skip it. + if chunk_bytes.startswith(":"): + continue + chunk = chunk_bytes.removeprefix("data:") + + data = json.loads(chunk) + timestamp = time.perf_counter() + # First token + if ttft == 0.0: + ttft = time.perf_counter() - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + most_recent_timestamp = timestamp + + output.latency = most_recent_timestamp - st + output.success = True + output.generated_text = data["generated_text"] + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_trt_llm( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith("generate_stream") + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + payload = { + "accumulate_tokens": True, + "text_input": request_func_input.prompt, + "temperature": 0.0, + "top_p": 1.0, + "max_tokens": request_func_input.output_len, + "stream": True, + } + if request_func_input.ignore_eos: + payload["min_length"] = request_func_input.output_len + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post(url=api_url, json=payload) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + + chunk = chunk_bytes.decode("utf-8").removeprefix("data:") + + data = json.loads(chunk) + output.generated_text += data["text_output"] + timestamp = time.perf_counter() + # First token + if ttft == 0.0: + ttft = timestamp - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + most_recent_timestamp = timestamp + + output.latency = most_recent_timestamp - st + output.success = True + + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_deepspeed_mii( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith(("completions", "profile")), ( + "OpenAI Completions API URL must end with 'completions' or 'profile'." + ) + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + payload = { + "model": request_func_input.model, + "prompt": request_func_input.prompt, + "max_tokens": request_func_input.output_len, + "temperature": 0.01, # deepspeed-mii does not accept 0.0 temp. + "top_p": 1.0, + } + headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"} + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + # NOTE: DeepSpeed-MII doesn't support streaming as of Jan 28 2024, + # will use 0 as placeholder. + # See https://github.com/microsoft/DeepSpeed-MII/pull/311 + output.ttft = 0 + + st = time.perf_counter() + try: + async with session.post( + url=api_url, json=payload, headers=headers + ) as response: + if response.status == 200: + parsed_resp = await response.json() + output.latency = time.perf_counter() - st + if "choices" in parsed_resp: + output.generated_text = parsed_resp["choices"][0]["text"] + elif "text" in parsed_resp: + output.generated_text = parsed_resp["text"][0] + else: + output.error = ( + "Unexpected response format: " + "neither 'choices' nor 'text' found" + ) + output.success = False + output.success = True + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_openai_completions( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith(("completions", "profile")), ( + "OpenAI Completions API URL must end with 'completions' or 'profile'." + ) + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + payload = { + "model": request_func_input.model_name + if request_func_input.model_name + else request_func_input.model, + "prompt": request_func_input.prompt, + "temperature": 0.0, + "repetition_penalty": 1.0, + "max_tokens": request_func_input.output_len, + "logprobs": request_func_input.logprobs, + "stream": True, + "stream_options": { + "include_usage": True, + }, + } + if request_func_input.ignore_eos: + payload["ignore_eos"] = request_func_input.ignore_eos + if request_func_input.extra_body: + payload.update(request_func_input.extra_body) + headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"} + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + generated_text = "" + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post( + url=api_url, json=payload, headers=headers + ) as response: + if response.status == 200: + first_chunk_received = False + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + + chunk = chunk_bytes.decode("utf-8").removeprefix("data: ") + if chunk != "[DONE]": + data = json.loads(chunk) + + # NOTE: Some completion API might have a last + # usage summary response without a token so we + # want to check a token was generated + if choices := data.get("choices"): + # Note that text could be empty here + # e.g. for special tokens + text = choices[0].get("text") + timestamp = time.perf_counter() + # First token + if not first_chunk_received: + first_chunk_received = True + ttft = time.perf_counter() - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + most_recent_timestamp = timestamp + generated_text += text or "" + if usage := data.get("usage"): + output.output_tokens = usage.get("completion_tokens") + if first_chunk_received: + output.success = True + else: + output.success = False + output.error = ( + "Never received a valid chunk to calculate TTFT." + "This response will be marked as failed!" + ) + output.generated_text = generated_text + output.latency = most_recent_timestamp - st + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_openai_chat_completions( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + api_url = request_func_input.api_url + assert api_url.endswith(("chat/completions", "profile")), ( + "OpenAI Chat Completions API URL must end with 'chat/completions'." + ) + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + content = [{"type": "text", "text": request_func_input.prompt}] + if request_func_input.multi_modal_content: + content.append(request_func_input.multi_modal_content) + payload = { + "model": request_func_input.model_name + if request_func_input.model_name + else request_func_input.model, + "messages": [ + {"role": "user", "content": content}, + ], + "temperature": 0.0, + "max_completion_tokens": request_func_input.output_len, + "stream": True, + "stream_options": { + "include_usage": True, + }, + } + if request_func_input.ignore_eos: + payload["ignore_eos"] = request_func_input.ignore_eos + if request_func_input.extra_body: + payload.update(request_func_input.extra_body) + headers = { + "Content-Type": "application/json", + "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}", + } + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + generated_text = "" + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post( + url=api_url, json=payload, headers=headers + ) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + chunk_bytes = chunk_bytes.decode("utf-8") + # NOTE: SSE comments (often used as pings) start with a colon. + # These are not JSON data payload and should be skipped. + if chunk_bytes.startswith(":"): + continue + + chunk = chunk_bytes.removeprefix("data: ") + + if chunk != "[DONE]": + timestamp = time.perf_counter() + data = json.loads(chunk) + + if choices := data.get("choices"): + content = choices[0]["delta"].get("content") + # First token + if ttft == 0.0: + ttft = timestamp - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append(timestamp - most_recent_timestamp) + + generated_text += content or "" + elif usage := data.get("usage"): + output.output_tokens = usage.get("completion_tokens") + + most_recent_timestamp = timestamp + + output.generated_text = generated_text + output.success = True + output.latency = most_recent_timestamp - st + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +async def async_request_openai_audio( + request_func_input: RequestFuncInput, + pbar: Optional[tqdm] = None, +) -> RequestFuncOutput: + # Lazy import without PlaceholderModule to avoid vllm dep. + import soundfile + + api_url = request_func_input.api_url + assert api_url.endswith(("transcriptions", "translations")), ( + "OpenAI Chat Completions API URL must end with 'transcriptions' " + ) + "or `translations`." + + async with aiohttp.ClientSession( + trust_env=True, timeout=AIOHTTP_TIMEOUT + ) as session: + content = [{"type": "text", "text": request_func_input.prompt}] + payload = { + "model": request_func_input.model_name + if request_func_input.model_name + else request_func_input.model, + "temperature": 0.0, + "max_completion_tokens": request_func_input.output_len, + "stream": True, + "language": "en", + # Flattened due to multipart/form-data + "stream_include_usage": True, + "stream_continuous_usage_stats": True, + } + if request_func_input.extra_body: + payload.update(request_func_input.extra_body) + headers = { + "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}", + } + + # Send audio file + def to_bytes(y, sr): + buffer = io.BytesIO() + soundfile.write(buffer, y, sr, format="WAV") + buffer.seek(0) + return buffer + + with to_bytes(*request_func_input.multi_modal_content["audio"]) as f: + form = aiohttp.FormData() + form.add_field("file", f, content_type="audio/wav") + for key, value in payload.items(): + form.add_field(key, str(value)) + + output = RequestFuncOutput() + output.prompt_len = request_func_input.prompt_len + + generated_text = "" + ttft = 0.0 + st = time.perf_counter() + most_recent_timestamp = st + try: + async with session.post( + url=api_url, data=form, headers=headers + ) as response: + if response.status == 200: + async for chunk_bytes in response.content: + chunk_bytes = chunk_bytes.strip() + if not chunk_bytes: + continue + + chunk = chunk_bytes.decode("utf-8").removeprefix("data: ") + if chunk != "[DONE]": + timestamp = time.perf_counter() + data = json.loads(chunk) + + if choices := data.get("choices"): + content = choices[0]["delta"].get("content") + # First token + if ttft == 0.0: + ttft = timestamp - st + output.ttft = ttft + + # Decoding phase + else: + output.itl.append( + timestamp - most_recent_timestamp + ) + + generated_text += content or "" + elif usage := data.get("usage"): + output.output_tokens = usage.get( + "completion_tokens" + ) + + most_recent_timestamp = timestamp + + output.generated_text = generated_text + output.success = True + output.latency = most_recent_timestamp - st + else: + output.error = response.reason or "" + output.success = False + except Exception: + output.success = False + exc_info = sys.exc_info() + output.error = "".join(traceback.format_exception(*exc_info)) + + if pbar: + pbar.update(1) + return output + + +def get_model(pretrained_model_name_or_path: str) -> str: + if os.getenv("VLLM_USE_MODELSCOPE", "False").lower() == "true": + from modelscope import snapshot_download + + from vllm.model_executor.model_loader.weight_utils import get_lock + + # Use file lock to prevent multiple processes from + # downloading the same model weights at the same time. + with get_lock(pretrained_model_name_or_path): + model_path = snapshot_download( + model_id=pretrained_model_name_or_path, + local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, + ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"], + ) + + return model_path + return pretrained_model_name_or_path + + +def get_tokenizer( + pretrained_model_name_or_path: str, + tokenizer_mode: str = "auto", + trust_remote_code: bool = False, + **kwargs, +) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]: + if pretrained_model_name_or_path is not None and not os.path.exists( + pretrained_model_name_or_path + ): + pretrained_model_name_or_path = get_model(pretrained_model_name_or_path) + if tokenizer_mode == "slow": + if kwargs.get("use_fast", False): + raise ValueError("Cannot use the fast tokenizer in slow tokenizer mode.") + kwargs["use_fast"] = False + if tokenizer_mode == "mistral": + try: + from vllm.transformers_utils.tokenizer import MistralTokenizer + except ImportError as e: + raise ImportError( + "MistralTokenizer requires vllm package.\n" + "Please install it with `pip install vllm` " + "to use mistral tokenizer mode." + ) from e + return MistralTokenizer.from_pretrained(str(pretrained_model_name_or_path)) + else: + return AutoTokenizer.from_pretrained( + pretrained_model_name_or_path, + trust_remote_code=trust_remote_code, + **kwargs, + ) + + +ASYNC_REQUEST_FUNCS = { + "tgi": async_request_tgi, + "vllm": async_request_openai_completions, + "lmdeploy": async_request_openai_completions, + "deepspeed-mii": async_request_deepspeed_mii, + "openai": async_request_openai_completions, + "openai-chat": async_request_openai_chat_completions, + "openai-audio": async_request_openai_audio, + "tensorrt-llm": async_request_trt_llm, + "scalellm": async_request_openai_completions, + "sglang": async_request_openai_completions, + "llama.cpp": async_request_openai_completions, +} + +OPENAI_COMPATIBLE_BACKENDS = [ + k + for k, v in ASYNC_REQUEST_FUNCS.items() + if v in (async_request_openai_completions, async_request_openai_chat_completions) +] diff --git a/online_apiserver_test_maxbs/benchmarks/benchmark_dataset.py b/online_apiserver_test_maxbs/benchmarks/benchmark_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..55c0cf851264f37d271f7a8cd1fc5ad17451f386 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/benchmark_dataset.py @@ -0,0 +1,1168 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +This module defines a framework for sampling benchmark requests from various +datasets. Each dataset subclass of BenchmarkDataset must implement sample +generation. Supported dataset types include: + - ShareGPT + - Random (synthetic) + - Sonnet + - BurstGPT + - HuggingFace + - VisionArena +""" + +import base64 +import io +import json +import logging +import random +from abc import ABC, abstractmethod +from collections.abc import Mapping +from dataclasses import dataclass +from functools import cache +from io import BytesIO +from typing import Any, Callable, Optional, Union + +import numpy as np +import pandas as pd +from datasets import load_dataset +from PIL import Image +from transformers import PreTrainedTokenizerBase + +from vllm.lora.request import LoRARequest +from vllm.lora.utils import get_adapter_absolute_path +from vllm.multimodal import MultiModalDataDict +from vllm.multimodal.image import convert_image_mode +from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer + +logger = logging.getLogger(__name__) + +# ----------------------------------------------------------------------------- +# Data Classes +# ----------------------------------------------------------------------------- + + +@dataclass +class SampleRequest: + """ + Represents a single inference request for benchmarking. + """ + + prompt: Union[str, Any] + prompt_len: int + expected_output_len: int + multi_modal_data: Optional[Union[MultiModalDataDict, dict]] = None + lora_request: Optional[LoRARequest] = None + + +# ----------------------------------------------------------------------------- +# Benchmark Dataset Base Class +# ----------------------------------------------------------------------------- + + +class BenchmarkDataset(ABC): + DEFAULT_SEED = 0 + IS_MULTIMODAL = False + + def __init__( + self, + dataset_path: Optional[str] = None, + random_seed: int = DEFAULT_SEED, + ) -> None: + """ + Initialize the BenchmarkDataset with an optional dataset path and random + seed. Args: + dataset_path (Optional[str]): Path to the dataset. If None, it + indicates that a default or random dataset might be used. + random_seed (int): Seed value for reproducible shuffling or + sampling. Defaults to DEFAULT_SEED. + """ + self.dataset_path = dataset_path + # Set the random seed, ensuring that a None value is replaced with the + # default seed. + self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED + self.data = None + + def apply_multimodal_chat_transformation( + self, prompt: str, mm_content: Optional[MultiModalDataDict] = None + ) -> list[dict]: + """ + Transform a prompt and optional multimodal content into a chat format. + This method is used for chat models that expect a specific conversation + format. + """ + content = [{"text": prompt, "type": "text"}] + if mm_content is not None: + content.append(mm_content) + return [{"role": "user", "content": content}] + + def load_data(self) -> None: + """ + Load data from the dataset path into self.data. + + This method must be overridden by subclasses since the method to load + data will vary depending on the dataset format and source. + + Raises: + NotImplementedError: If a subclass does not implement this method. + """ + # TODO (jenniferzhao): add support for downloading data + raise NotImplementedError("load_data must be implemented in subclasses.") + + def get_random_lora_request( + self, + tokenizer: PreTrainedTokenizerBase, + max_loras: Optional[int] = None, + lora_path: Optional[str] = None, + ) -> tuple[Optional[LoRARequest], AnyTokenizer]: + """ + Optionally select a random LoRA request and return its associated + tokenizer. + + This method is used when LoRA parameters are provided. It randomly + selects a LoRA based on max_loras and retrieves a cached tokenizer for + that LoRA if available. Otherwise, it returns the base tokenizer. + + Args: + tokenizer (PreTrainedTokenizerBase): The base tokenizer to use if no + LoRA is selected. max_loras (Optional[int]): The maximum number of + LoRAs available. If None, LoRA is not used. lora_path + (Optional[str]): Path to the LoRA parameters on disk. If None, LoRA + is not used. + + Returns: + tuple[Optional[LoRARequest], AnyTokenizer]: A tuple where the first + element is a LoRARequest (or None if not applicable) and the second + element is the tokenizer associated with the LoRA request (or the + base tokenizer). + """ + if max_loras is None or lora_path is None: + return None, tokenizer + + # Generate a random LoRA ID in the range [1, max_loras]. + lora_id = random.randint(1, max_loras) + lora_request = LoRARequest( + lora_name=str(lora_id), + lora_int_id=lora_id, + lora_path=lora_path_on_disk(lora_path), + ) + if lora_id not in lora_tokenizer_cache: + lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request) + # Return lora_request and the cached tokenizer if available; otherwise, + # return the base tokenizer + return lora_request, lora_tokenizer_cache[lora_id] or tokenizer + + @abstractmethod + def sample( + self, tokenizer: PreTrainedTokenizerBase, num_requests: int + ) -> list[SampleRequest]: + """ + Abstract method to generate sample requests from the dataset. + + Subclasses must override this method to implement dataset-specific logic + for generating a list of SampleRequest objects. + + Args: + tokenizer (PreTrainedTokenizerBase): The tokenizer to be used + for processing the dataset's text. + num_requests (int): The number of sample requests to generate. + + Returns: + list[SampleRequest]: A list of sample requests generated from the + dataset. + """ + raise NotImplementedError("sample must be implemented in subclasses.") + + def maybe_oversample_requests( + self, requests: list[SampleRequest], num_requests: int + ) -> None: + """ + Oversamples the list of requests if its size is less than the desired + number. + + Args: + requests (List[SampleRequest]): The current list of sampled + requests. num_requests (int): The target number of requests. + """ + if len(requests) < num_requests: + random.seed(self.random_seed) + additional = random.choices(requests, k=num_requests - len(requests)) + requests.extend(additional) + logger.info("Oversampled requests to reach %d total samples.", num_requests) + + +# ----------------------------------------------------------------------------- +# Utility Functions and Global Caches +# ----------------------------------------------------------------------------- + + +def is_valid_sequence( + prompt_len: int, + output_len: int, + min_len: int = 4, + max_prompt_len: int = 1024, + max_total_len: int = 2048, + skip_min_output_len_check: bool = False, +) -> bool: + """ + Validate a sequence based on prompt and output lengths. + + Default pruning criteria are copied from the original `sample_hf_requests` + and `sample_sharegpt_requests` functions in benchmark_serving.py, as well as + from `sample_requests` in benchmark_throughput.py. + """ + # Check for invalid conditions + prompt_too_short = prompt_len < min_len + output_too_short = (not skip_min_output_len_check) and (output_len < min_len) + prompt_too_long = prompt_len > max_prompt_len + combined_too_long = (prompt_len + output_len) > max_total_len + + # Return True if none of the invalid conditions are met + return not ( + prompt_too_short or output_too_short or prompt_too_long or combined_too_long + ) + + +@cache +def lora_path_on_disk(lora_path: str) -> str: + return get_adapter_absolute_path(lora_path) + + +# Global cache for LoRA tokenizers. +lora_tokenizer_cache: dict[int, AnyTokenizer] = {} + + +def process_image(image: Any) -> Mapping[str, Any]: + """ + Process a single image input and return a multimedia content dictionary. + + Supports three input types: + + 1. Dictionary with raw image bytes: - Expects a dict with a 'bytes' key + containing raw image data. - Loads the bytes as a PIL.Image.Image. + + 2. PIL.Image.Image input: - Converts the image to RGB. - Saves the image as + a JPEG in memory. - Encodes the JPEG data as a base64 string. - Returns + a dictionary with the image as a base64 data URL. + + 3. String input: - Treats the string as a URL or local file path. - + Prepends "file://" if the string doesn't start with "http://" or + "file://". - Returns a dictionary with the image URL. + + Raises: + ValueError: If the input is not a supported type. + """ + if isinstance(image, dict) and "bytes" in image: + image = Image.open(BytesIO(image["bytes"])) + if isinstance(image, Image.Image): + image = convert_image_mode(image, "RGB") + with io.BytesIO() as image_data: + image.save(image_data, format="JPEG") + image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8") + return { + "type": "image_url", + "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"}, + } + + if isinstance(image, str): + image_url = ( + image if image.startswith(("http://", "file://")) else f"file://{image}" + ) + return {"type": "image_url", "image_url": {"url": image_url}} + + raise ValueError( + f"Invalid image input {image}. Must be a PIL.Image.Image" + " or str or dictionary with raw image bytes." + ) + + +# ----------------------------------------------------------------------------- +# Random Dataset Implementation (Synthetic Data) +# ----------------------------------------------------------------------------- + + +class RandomDataset(BenchmarkDataset): + # Default values copied from benchmark_serving.py for the random dataset. + DEFAULT_PREFIX_LEN = 0 + DEFAULT_RANGE_RATIO = 0.0 + DEFAULT_INPUT_LEN = 1024 + DEFAULT_OUTPUT_LEN = 128 + + def __init__( + self, + **kwargs, + ) -> None: + super().__init__(**kwargs) + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + prefix_len: int = DEFAULT_PREFIX_LEN, + range_ratio: float = DEFAULT_RANGE_RATIO, + input_len: int = DEFAULT_INPUT_LEN, + output_len: int = DEFAULT_OUTPUT_LEN, + **kwargs, + ) -> list[SampleRequest]: + # Enforce range_ratio < 1 + assert range_ratio < 1.0, ( + "random_range_ratio must be < 1.0 to ensure a valid sampling range" + ) + + vocab_size = tokenizer.vocab_size + num_special_tokens = tokenizer.num_special_tokens_to_add() + real_input_len = input_len - num_special_tokens + + prefix_token_ids = ( + np.random.randint(0, vocab_size, size=prefix_len).tolist() + if prefix_len > 0 + else [] + ) + + # New sampling logic: [X * (1 - b), X * (1 + b)] + input_low = int(real_input_len * (1 - range_ratio)) + input_high = int(real_input_len * (1 + range_ratio)) + output_low = int(output_len * (1 - range_ratio)) + output_high = int(output_len * (1 + range_ratio)) + + # Add logging for debugging + logger.info("Sampling input_len from [%s, %s]", input_low, input_high) + logger.info("Sampling output_len from [%s, %s]", output_low, output_high) + + input_lens = np.random.randint(input_low, input_high + 1, size=num_requests) + output_lens = np.random.randint(output_low, output_high + 1, size=num_requests) + offsets = np.random.randint(0, vocab_size, size=num_requests) + + requests = [] + for i in range(num_requests): + inner_seq = ( + (offsets[i] + i + np.arange(input_lens[i])) % vocab_size + ).tolist() + token_sequence = prefix_token_ids + inner_seq + prompt = tokenizer.decode(token_sequence) + # After decoding the prompt we have to encode and decode it again. + # This is done because in some cases N consecutive tokens + # give a string tokenized into != N number of tokens. + # For example for GPT2Tokenizer: + # [6880, 6881] -> ['Ġcalls', 'here'] -> + # [1650, 939, 486] -> ['Ġcall', 'sh', 'ere'] + # To avoid uncontrolled change of the prompt length, + # the encoded sequence is truncated before being decode again. + total_input_len = prefix_len + int(input_lens[i]) + re_encoded_sequence = tokenizer.encode(prompt, add_special_tokens=False)[ + :total_input_len + ] + prompt = tokenizer.decode(re_encoded_sequence) + total_input_len = len(re_encoded_sequence) + requests.append( + SampleRequest( + prompt=prompt, + prompt_len=total_input_len, + expected_output_len=int(output_lens[i]), + ) + ) + return requests + + +# ----------------------------------------------------------------------------- +# ShareGPT Dataset Implementation +# ----------------------------------------------------------------------------- + + +class ShareGPTDataset(BenchmarkDataset): + """ + Implements the ShareGPT dataset. Loads data from a JSON file and generates + sample requests based on conversation turns. + """ + + def __init__(self, **kwargs) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data(self) -> None: + if self.dataset_path is None: + raise ValueError("dataset_path must be provided for loading data.") + + with open(self.dataset_path, encoding="utf-8") as f: + self.data = json.load(f) + # Filter entries with at least two conversation turns. + self.data = [ + entry + for entry in self.data + if "conversations" in entry and len(entry["conversations"]) >= 2 + ] + random.seed(self.random_seed) + random.shuffle(self.data) + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + lora_path: Optional[str] = None, + max_loras: Optional[int] = None, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + samples: list = [] + for entry in self.data: + if len(samples) >= num_requests: + break + prompt, completion = ( + entry["conversations"][0]["value"], + entry["conversations"][1]["value"], + ) + + lora_request, tokenizer = self.get_random_lora_request( + tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path + ) + prompt_ids = tokenizer(prompt).input_ids + completion_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_ids) + new_output_len = len(completion_ids) if output_len is None else output_len + if not is_valid_sequence( + prompt_len, + new_output_len, + skip_min_output_len_check=output_len is not None, + ): + continue + if enable_multimodal_chat: + prompt = self.apply_multimodal_chat_transformation(prompt, None) + samples.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=new_output_len, + lora_request=lora_request, + ) + ) + self.maybe_oversample_requests(samples, num_requests) + return samples + + +# ----------------------------------------------------------------------------- +# Custom Dataset Implementation +# ----------------------------------------------------------------------------- + + +class CustomDataset(BenchmarkDataset): + """ + Implements the Custom dataset. Loads data from a JSONL file and generates + sample requests based on conversation turns. E.g., + ``` + {"prompt": "What is the capital of India?"} + {"prompt": "What is the capital of Iran?"} + {"prompt": "What is the capital of China?"} + ``` + """ + + def __init__(self, **kwargs) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data(self) -> None: + if self.dataset_path is None: + raise ValueError("dataset_path must be provided for loading data.") + + # self.data will be a list of dictionaries + # e.g., [{"prompt": "What is the capital of India?"}, ...] + # This will be the standardized format which load_data() + # has to convert into depending on the filetype of dataset_path. + # sample() will assume this standardized format of self.data + self.data = [] + + # Load the JSONL file + if self.dataset_path.endswith(".jsonl"): + jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True) + + # check if the JSONL file has a 'prompt' column + if "prompt" not in jsonl_data.columns: + raise ValueError("JSONL file must contain a 'prompt' column.") + + # Convert each row to a dictionary and append to self.data + # This will convert the DataFrame to a list of dictionaries + # where each dictionary corresponds to a row in the DataFrame. + # This is the standardized format we want for self.data + for _, row in jsonl_data.iterrows(): + self.data.append(row.to_dict()) + else: + raise NotImplementedError( + "Only JSONL format is supported for CustomDataset." + ) + + random.seed(self.random_seed) + random.shuffle(self.data) + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + lora_path: Optional[str] = None, + max_loras: Optional[int] = None, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + skip_chat_template: bool = False, + **kwargs, + ) -> list: + sampled_requests = [] + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt = item["prompt"] + + # apply template + if not skip_chat_template: + prompt = tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=True, + tokenize=False, + ) + + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Sonnet Dataset Implementation +# ----------------------------------------------------------------------------- + + +class SonnetDataset(BenchmarkDataset): + """ + Simplified implementation of the Sonnet dataset. Loads poem lines from a + text file and generates sample requests. Default values here copied from + `benchmark_serving.py` for the sonnet dataset. + """ + + DEFAULT_PREFIX_LEN = 200 + DEFAULT_INPUT_LEN = 550 + DEFAULT_OUTPUT_LEN = 150 + + def __init__( + self, + **kwargs, + ) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data(self) -> None: + if not self.dataset_path: + raise ValueError("dataset_path must be provided.") + with open(self.dataset_path, encoding="utf-8") as f: + self.data = f.readlines() + + def sample( + self, + tokenizer, + num_requests: int, + prefix_len: int = DEFAULT_PREFIX_LEN, + input_len: int = DEFAULT_INPUT_LEN, + output_len: int = DEFAULT_OUTPUT_LEN, + return_prompt_formatted: bool = False, + **kwargs, + ) -> list: + # Calculate average token length for a poem line. + tokenized_lines = [tokenizer(line).input_ids for line in self.data] + avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines) + + # Build the base prompt. + base_prompt = "Pick as many lines as you can from these poem lines:\n" + base_msg = [{"role": "user", "content": base_prompt}] + base_fmt = tokenizer.apply_chat_template( + base_msg, add_generation_prompt=True, tokenize=False + ) + base_offset = len(tokenizer(base_fmt).input_ids) + if input_len <= base_offset: + raise ValueError( + f"'input_len' must be higher than the base prompt length " + f"({base_offset})." + ) + + # Determine how many poem lines to use. + num_input_lines = round((input_len - base_offset) / avg_len) + num_prefix_lines = max(round((prefix_len - base_offset) / avg_len), 0) + prefix_lines = self.data[:num_prefix_lines] + + samples = [] + while len(samples) < num_requests: + extra_lines = random.choices( + self.data, k=num_input_lines - num_prefix_lines + ) + prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}" + msg = [{"role": "user", "content": prompt}] + prompt_formatted = tokenizer.apply_chat_template( + msg, add_generation_prompt=True, tokenize=False + ) + prompt_len = len(tokenizer(prompt_formatted).input_ids) + if prompt_len <= input_len: + samples.append( + SampleRequest( + prompt=prompt_formatted if return_prompt_formatted else prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + return samples + + +# ----------------------------------------------------------------------------- +# BurstGPT Dataset Implementation +# ----------------------------------------------------------------------------- + + +class BurstGPTDataset(BenchmarkDataset): + """ + Implements the BurstGPT dataset. Loads data from a CSV file and generates + sample requests based on synthetic prompt generation. Only rows with Model + "GPT-4" and positive response tokens are used. + """ + + def __init__(self, **kwargs) -> None: + super().__init__(**kwargs) + self.load_data() + + def load_data( + self, + ): + if self.dataset_path is None: + raise ValueError("dataset_path must be provided for loading data.") + + df = pd.read_csv(self.dataset_path) + # Filter to keep only GPT-4 rows. + gpt4_df = df[df["Model"] == "GPT-4"] + # Remove failed requests (where Response tokens is 0 or less). + gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0] + # Sample the desired number of rows. + self.data = gpt4_df + + def _sample_loaded_data(self, num_requests: int) -> list: + if num_requests <= len(self.data): + data = self.data.sample(n=num_requests, random_state=self.random_seed) + else: + data = self.data.sample( + n=num_requests, + random_state=self.random_seed, + replace=True, + ) + # Convert the dataframe to a list of lists. + return data.values.tolist() + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + max_loras: Optional[int] = None, + lora_path: Optional[str] = None, + **kwargs, + ) -> list[SampleRequest]: + samples = [] + data = self._sample_loaded_data(num_requests=num_requests) + for i in range(num_requests): + input_len = int(data[i][2]) + output_len = int(data[i][3]) + lora_req, tokenizer = self.get_random_lora_request( + tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path + ) + vocab_size = tokenizer.vocab_size + # Generate a synthetic prompt: a list of token IDs computed as (i + + # j) modulo vocab_size. + token_ids = [(i + j) % vocab_size for j in range(input_len)] + prompt = tokenizer.decode(token_ids) + samples.append( + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=output_len, + lora_request=lora_req, + ) + ) + return samples + + +# ----------------------------------------------------------------------------- +# HuggingFace Dataset Base Implementation +# ----------------------------------------------------------------------------- +class HuggingFaceDataset(BenchmarkDataset): + """Base class for datasets hosted on HuggingFace.""" + + SUPPORTED_DATASET_PATHS: Union[set[str], dict[str, Callable]] = set() + + def __init__( + self, + dataset_path: str, + dataset_split: str, + dataset_subset: Optional[str] = None, + **kwargs, + ) -> None: + super().__init__(dataset_path=dataset_path, **kwargs) + + self.dataset_split = dataset_split + self.dataset_subset = dataset_subset + self.load_data() + + def load_data(self) -> None: + """Load data from HuggingFace datasets.""" + self.data = load_dataset( + self.dataset_path, + name=self.dataset_subset, + split=self.dataset_split, + streaming=True, + ) + self.data = self.data.shuffle(seed=self.random_seed) + + +# ----------------------------------------------------------------------------- +# Conversation Dataset Implementation +# ----------------------------------------------------------------------------- + + +class ConversationDataset(HuggingFaceDataset): + """Dataset for conversation data with multimodal support.""" + + SUPPORTED_DATASET_PATHS = { + "lmms-lab/LLaVA-OneVision-Data", + "Aeala/ShareGPT_Vicuna_unfiltered", + } + IS_MULTIMODAL = True + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + # Filter examples with at least 2 conversations + filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2) + sampled_requests = [] + dynamic_output = output_len is None + + for item in filtered_data: + if len(sampled_requests) >= num_requests: + break + conv = item["conversations"] + prompt, completion = conv[0]["value"], conv[1]["value"] + + prompt_ids = tokenizer(prompt).input_ids + completion_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_ids) + completion_len = len(completion_ids) + output_len = completion_len if dynamic_output else output_len + assert isinstance(output_len, int) and output_len > 0 + if dynamic_output and not is_valid_sequence(prompt_len, completion_len): + continue + mm_content = process_image(item["image"]) if "image" in item else None + if enable_multimodal_chat: + # Note: when chat is enabled the request prompt_len is no longer + # accurate and we will be using request output to count the + # actual prompt len and output len + prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=mm_content, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Vision Arena Dataset Implementation +# ----------------------------------------------------------------------------- + + +class VisionArenaDataset(HuggingFaceDataset): + """ + Vision Arena Dataset. + """ + + DEFAULT_OUTPUT_LEN = 128 + SUPPORTED_DATASET_PATHS = { + "lmarena-ai/VisionArena-Chat": lambda x: x["conversation"][0][0]["content"], + "lmarena-ai/vision-arena-bench-v0.1": lambda x: x["turns"][0][0]["content"], + } + IS_MULTIMODAL = True + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + sampled_requests = [] + for item in self.data: + if len(sampled_requests) >= num_requests: + break + parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.dataset_path) + if parser_fn is None: + raise ValueError(f"Unsupported dataset path: {self.dataset_path}") + prompt = parser_fn(item) + mm_content = process_image(item["images"][0]) + prompt_len = len(tokenizer(prompt).input_ids) + if enable_multimodal_chat: + # Note: when chat is enabled the request prompt_len is no longer + # accurate and we will be using request output to count the + # actual prompt len + prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=mm_content, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Instruct Coder Dataset Implementation +# ----------------------------------------------------------------------------- + + +class InstructCoderDataset(HuggingFaceDataset): + """ + InstructCoder Dataset. + https://huggingface.co/datasets/likaixin/InstructCoder + + InstructCoder is the dataset designed for general code editing. It consists + of 114,239 instruction-input-output triplets, and covers multiple distinct + code editing scenario. + """ + + DEFAULT_OUTPUT_LEN = 200 # this is the average default output length + SUPPORTED_DATASET_PATHS = { + "likaixin/InstructCoder", + } + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + sampled_requests = [] + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt = f"{item['input']}\n\n{item['instruction']} Just output \ + the code, do not include any explanation." + + # apply template + prompt = tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=True, + tokenize=False, + ) + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# MT-Bench Dataset Implementation +# ----------------------------------------------------------------------------- + + +class MTBenchDataset(HuggingFaceDataset): + """ + MT-Bench Dataset. + https://huggingface.co/datasets/philschmid/mt-bench + + We create a single turn dataset for MT-Bench. + This is similar to Spec decoding benchmark setup in vLLM + https://github.com/vllm-project/vllm/blob/9d98ab5ec/examples/offline_inference/eagle.py#L14-L18 + """ # noqa: E501 + + DEFAULT_OUTPUT_LEN = 256 # avg len used in SD bench in vLLM + SUPPORTED_DATASET_PATHS = { + "philschmid/mt-bench", + } + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + enable_multimodal_chat: bool = False, + **kwargs, + ) -> list: + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + sampled_requests = [] + + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt = item["turns"][0] + + # apply template + prompt = tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=True, + tokenize=False, + ) + + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# AIMO Dataset Implementation +# ----------------------------------------------------------------------------- + + +class AIMODataset(HuggingFaceDataset): + """ + Dataset class for processing a AIMO dataset with reasoning questions. + """ + + SUPPORTED_DATASET_PATHS = { + "AI-MO/aimo-validation-aime", + "AI-MO/NuminaMath-1.5", + "AI-MO/NuminaMath-CoT", + } + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + **kwargs, + ) -> list: + sampled_requests = [] + dynamic_output = output_len is None + + for item in self.data: + if len(sampled_requests) >= num_requests: + break + prompt, completion = item["problem"], item["solution"] + + prompt_ids = tokenizer(prompt).input_ids + completion_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_ids) + completion_len = len(completion_ids) + output_len = completion_len if dynamic_output else output_len + assert isinstance(output_len, int) and output_len > 0 + if dynamic_output and not is_valid_sequence( + prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000 + ): + continue + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=None, + ) + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests + + +# ----------------------------------------------------------------------------- +# Next Edit Prediction Dataset Implementation +# ----------------------------------------------------------------------------- + + +zeta_prompt = """### Instruction: +You are a code completion assistant and your task is to analyze user edits and then rewrite an excerpt that the user provides, suggesting the appropriate edits within the excerpt, taking into account the cursor location. + +### User Edits: + +{} + +### User Excerpt: + +{} + +### Response: + +""" # noqa: E501 + + +def _format_zeta_prompt( + sample: dict, original_start_marker: str = "<|editable_region_start|>" +) -> dict: + """Format the zeta prompt for the Next Edit Prediction (NEP) dataset. + + This function formats examples from the NEP dataset + into prompts and expected outputs. It could be + further extended to support more NEP datasets. + + Args: + sample: The dataset sample containing events, + inputs, and outputs. + original_start_marker: The marker indicating the + start of the editable region. Defaults to + "<|editable_region_start|>". + + Returns: + A dictionary with the formatted prompts and expected outputs. + """ + events = sample["events"] + input = sample["input"] + output = sample["output"] + prompt = zeta_prompt.format(events, input) + + # following the original implementation, extract the focused region + # from the raw output + output_start_index = output.find(original_start_marker) + output_focused_region = output[output_start_index:] + expected_output = output_focused_region + + return {"prompt": prompt, "expected_output": expected_output} + + +class NextEditPredictionDataset(HuggingFaceDataset): + """ + Dataset class for processing a Next Edit Prediction dataset. + """ + + SUPPORTED_DATASET_PATHS = { + "zed-industries/zeta", + } + MAPPING_PROMPT_FUNCS = { + "zed-industries/zeta": _format_zeta_prompt, + } + + def sample(self, tokenizer: PreTrainedTokenizerBase, num_requests: int, **kwargs): + formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.dataset_path) + if formatting_prompt_func is None: + raise ValueError(f"Unsupported dataset path: {self.dataset_path}") + samples = [] + for sample in self.data: + sample = formatting_prompt_func(sample) + samples.append( + SampleRequest( + prompt=sample["prompt"], + prompt_len=len(tokenizer(sample["prompt"]).input_ids), + expected_output_len=len( + tokenizer(sample["expected_output"]).input_ids + ), + ) + ) + if len(samples) >= num_requests: + break + self.maybe_oversample_requests(samples, num_requests) + return samples + + +# ----------------------------------------------------------------------------- +# ASR Dataset Implementation +# ----------------------------------------------------------------------------- + + +class ASRDataset(HuggingFaceDataset): + """ + Dataset class for processing a ASR dataset for transcription. + Tested on the following set: + + +----------------+----------------------------------------+--------------------------+-----------------------------+ + | Dataset | Domain | Speaking Style | hf-subset | + +----------------+----------------------------------------+--------------------------+-----------------------------+ + | TED-LIUM | TED talks | Oratory | release1, release2, release3| + | | | | release3-speaker-adaptation | + | VoxPopuli | European Parliament | Oratory | en, de, it, fr, ... | + | LibriSpeech | Audiobook | Narrated | "LIUM/tedlium" | + | GigaSpeech | Audiobook, podcast, YouTube | Narrated, spontaneous | xs, s, m, l, xl, dev, test | + | SPGISpeech | Financial meetings | Oratory, spontaneous | S, M, L, dev, test | + | AMI | Meetings | Spontaneous | ihm, sdm | + +----------------+----------------------------------------+--------------------------+-----------------------------+ + + """ # noqa: E501 + + SUPPORTED_DATASET_PATHS = { + "openslr/librispeech_asr", + "facebook/voxpopuli", + "LIUM/tedlium", + "edinburghcstr/ami", + "speechcolab/gigaspeech", + "kensho/spgispeech", + } + + DEFAULT_OUTPUT_LEN = 128 + IS_MULTIMODAL = True + + # TODO Whisper-specific. Abstract interface when more models are supported. + TRANSCRIPTION_PREAMBLE = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>" + skip_long_audios: bool = True + + def sample( + self, + tokenizer: PreTrainedTokenizerBase, + num_requests: int, + output_len: Optional[int] = None, + **kwargs, + ) -> list: + import librosa + + output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN + prompt = ASRDataset.TRANSCRIPTION_PREAMBLE + prompt_len = len(tokenizer(prompt).input_ids) + sampled_requests = [] + skipped = 0 + for item in self.data: + if len(sampled_requests) >= num_requests: + break + audio = item["audio"] + y, sr = audio["array"], audio["sampling_rate"] + duration_s = librosa.get_duration(y=y, sr=sr) + # Whisper max supported duration + if self.skip_long_audios and duration_s > 30: + skipped += 1 + continue + + mm_content = {"audio": (y, sr)} + sampled_requests.append( + SampleRequest( + prompt=prompt, + prompt_len=prompt_len, + expected_output_len=output_len, + multi_modal_data=mm_content, + ) + ) + if skipped: + logger.warning( + "%d samples discarded from dataset due to" + " their length being greater than" + " what Whisper supports.", + skipped, + ) + self.maybe_oversample_requests(sampled_requests, num_requests) + return sampled_requests diff --git a/online_apiserver_test_maxbs/benchmarks/benchmark_latency.py b/online_apiserver_test_maxbs/benchmarks/benchmark_latency.py new file mode 100644 index 0000000000000000000000000000000000000000..4d2ea126b24a514bde2e7e3808d2120ede353cc7 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/benchmark_latency.py @@ -0,0 +1,186 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Benchmark the latency of processing a single batch of requests.""" + +import argparse +import dataclasses +import json +import os +import time +from typing import Any, Optional + +import numpy as np +from tqdm import tqdm + +import vllm.envs as envs +from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json +from vllm import LLM, SamplingParams +from vllm.engine.arg_utils import EngineArgs +from vllm.inputs import PromptType +from vllm.sampling_params import BeamSearchParams +from vllm.utils import FlexibleArgumentParser + + +def save_to_pytorch_benchmark_format( + args: argparse.Namespace, results: dict[str, Any] +) -> None: + pt_records = convert_to_pytorch_benchmark_format( + args=args, + metrics={"latency": results["latencies"]}, + extra_info={k: results[k] for k in ["avg_latency", "percentiles"]}, + ) + if pt_records: + pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json" + write_to_json(pt_file, pt_records) + + +def main(args: argparse.Namespace): + print(args) + + engine_args = EngineArgs.from_cli_args(args) + + # NOTE(woosuk): If the request cannot be processed in a single batch, + # the engine will automatically process the request in multiple batches. + llm = LLM(**dataclasses.asdict(engine_args)) + assert llm.llm_engine.model_config.max_model_len >= ( + args.input_len + args.output_len + ), ( + "Please ensure that max_model_len is greater than" + " the sum of input_len and output_len." + ) + + sampling_params = SamplingParams( + n=args.n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=args.output_len, + detokenize=not args.disable_detokenize, + ) + print(sampling_params) + dummy_prompt_token_ids = np.random.randint( + 10000, size=(args.batch_size, args.input_len) + ) + dummy_prompts: list[PromptType] = [ + {"prompt_token_ids": batch} for batch in dummy_prompt_token_ids.tolist() + ] + + def llm_generate(): + if not args.use_beam_search: + llm.generate(dummy_prompts, sampling_params=sampling_params, use_tqdm=False) + else: + llm.beam_search( + dummy_prompts, + BeamSearchParams( + beam_width=args.n, + max_tokens=args.output_len, + ignore_eos=True, + ), + ) + + def run_to_completion(profile_dir: Optional[str] = None): + if profile_dir: + llm.start_profile() + llm_generate() + llm.stop_profile() + else: + start_time = time.perf_counter() + llm_generate() + end_time = time.perf_counter() + latency = end_time - start_time + return latency + + print("Warming up...") + for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"): + run_to_completion(profile_dir=None) + + if args.profile: + profile_dir = envs.VLLM_TORCH_PROFILER_DIR + print(f"Profiling (results will be saved to '{profile_dir}')...") + run_to_completion(profile_dir=profile_dir) + return + + # Benchmark. + latencies = [] + for _ in tqdm(range(args.num_iters), desc="Profiling iterations"): + latencies.append(run_to_completion(profile_dir=None)) + latencies = np.array(latencies) + percentages = [10, 25, 50, 75, 90, 99] + percentiles = np.percentile(latencies, percentages) + print(f"Avg latency: {np.mean(latencies)} seconds") + for percentage, percentile in zip(percentages, percentiles): + print(f"{percentage}% percentile latency: {percentile} seconds") + + # Output JSON results if specified + if args.output_json: + results = { + "avg_latency": np.mean(latencies), + "latencies": latencies.tolist(), + "percentiles": dict(zip(percentages, percentiles.tolist())), + } + with open(args.output_json, "w") as f: + json.dump(results, f, indent=4) + save_to_pytorch_benchmark_format(args, results) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the latency of processing a single batch of " + "requests till completion." + ) + parser.add_argument("--input-len", type=int, default=32) + parser.add_argument("--output-len", type=int, default=128) + parser.add_argument("--batch-size", type=int, default=8) + parser.add_argument( + "--n", + type=int, + default=1, + help="Number of generated sequences per prompt.", + ) + parser.add_argument("--use-beam-search", action="store_true") + parser.add_argument( + "--num-iters-warmup", + type=int, + default=10, + help="Number of iterations to run for warmup.", + ) + parser.add_argument( + "--num-iters", type=int, default=30, help="Number of iterations to run." + ) + parser.add_argument( + "--profile", + action="store_true", + help="profile the generation process of a single batch", + ) + parser.add_argument( + "--output-json", + type=str, + default=None, + help="Path to save the latency results in JSON format.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize responses (i.e. do not include " + "detokenization time in the latency measurement)" + ), + ) + + parser = EngineArgs.add_cli_args(parser) + # V1 enables prefix caching by default which skews the latency + # numbers. We need to disable prefix caching by default. + parser.set_defaults(enable_prefix_caching=False) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + if args.profile and not envs.VLLM_TORCH_PROFILER_DIR: + raise OSError( + "The environment variable 'VLLM_TORCH_PROFILER_DIR' is not set. " + "Please set it to a valid path to use torch profiler." + ) + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/benchmark_long_document_qa_throughput.py b/online_apiserver_test_maxbs/benchmarks/benchmark_long_document_qa_throughput.py new file mode 100644 index 0000000000000000000000000000000000000000..6e0f3b51c9d284d22dd1acdc1b794b9a5c0f31b1 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/benchmark_long_document_qa_throughput.py @@ -0,0 +1,202 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +Offline benchmark to test the long document QA throughput. + +Example usage: + # This workload samples 8 different prompts with a default input + # length of 20000 tokens, then replicates each prompt 2 times + # in random order. + python benchmark_long_document_qa_throughput.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-documents 8 \ + --repeat-count 2 + +Commandline arguments: + --num-documents: The number of documents to sample prompts from. + + --document-length: The length of each document in tokens. + (Optional, default: 20000) + + --output-len: The number of tokens to generate for each prompt. + (Optional, default: 10) + + --repeat-count: The number of times to repeat each prompt. + (Optional, default: 2) + + --repeat-mode: The mode to repeat prompts. The supported modes are: + - 'random': shuffle the prompts randomly. (Default) + - 'tile': the entire prompt list is repeated in sequence. (Potentially + lowest cache hit) + - 'interleave': each prompt is repeated consecutively before + moving to the next element. (Highest cache hit) + + --shuffle-seed: Random seed when the repeat mode is "random". + (Optional, default: 0) + +In the meantime, it also supports all the vLLM engine args to initialize the +LLM engine. You can refer to the `vllm.engine.arg_utils.EngineArgs` for more +details. +""" + +import dataclasses +import random +import time + +from vllm import LLM, SamplingParams +from vllm.engine.arg_utils import EngineArgs +from vllm.utils import FlexibleArgumentParser + + +def test_long_document_qa(llm=None, sampling_params=None, prompts=None): + """ + Test long document QA with the given prompts and sampling parameters. + Print the time spent in processing all the prompts. + + Args: + llm: The language model used for generating responses. + sampling_params: Sampling parameter used to generate the response. + prompts: A list of prompt strings to be processed by the LLM. + """ + start_time = time.time() + llm.generate(prompts, sampling_params=sampling_params) + end_time = time.time() + print(f"Time to execute all requests: {end_time - start_time:.4f} secs") + + +def repeat_prompts(prompts, repeat_count, mode: str): + """ + Repeat each prompt in the list for a specified number of times. + The order of prompts in the output list depends on the mode. + + Args: + prompts: A list of prompts to be repeated. + repeat_count: The number of times each prompt is repeated. + mode: The mode of repetition. Supported modes are: + - 'random': Shuffle the prompts randomly after repetition. + - 'tile': Repeat the entire prompt list in sequence. + Example: [1, 2, 3] -> [1, 2, 3, 1, 2, 3]. + - 'interleave': Repeat each prompt consecutively before moving to + the next. Example: [1, 2, 3] -> [1, 1, 2, 2, 3, 3]. + + Returns: + A list of repeated prompts in the specified order. + + Raises: + ValueError: If an invalid mode is provided. + """ + print("Repeat mode: ", mode) + if mode == "random": + repeated_prompts = prompts * repeat_count + random.shuffle(repeated_prompts) + return repeated_prompts + elif mode == "tile": + return prompts * repeat_count + elif mode == "interleave": + repeated_prompts = [] + for prompt in prompts: + repeated_prompts.extend([prompt] * repeat_count) + return repeated_prompts + else: + raise ValueError( + f"Invalid mode: {mode}, only support 'random', 'tile', 'interleave'" + ) + + +def main(args): + random.seed(args.shuffle_seed) + + # Prepare the prompts: + # we append the document id at the beginning to avoid any of the document + # being the prefix of other documents + prompts = [ + str(i) + " ".join(["hi"] * args.document_length) + for i in range(args.num_documents) + ] + + prompts = repeat_prompts(prompts, args.repeat_count, mode=args.repeat_mode) + + warmup_prompts = [ + "This is warm up request " + str(i) + " ".join(["hi"] * args.document_length) + for i in range(args.num_documents) + ] + + # Create the LLM engine + engine_args = EngineArgs.from_cli_args(args) + llm = LLM(**dataclasses.asdict(engine_args)) + sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len) + + print("------warm up------") + test_long_document_qa( + llm=llm, + prompts=warmup_prompts, + sampling_params=sampling_params, + ) + + print("------start generating------") + test_long_document_qa( + llm=llm, + prompts=prompts, + sampling_params=sampling_params, + ) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the performance with or " + "without automatic prefix caching." + ) + + parser.add_argument( + "--document-length", + type=int, + # Roughly the number of tokens for a system paper, + # excluding images + default=20000, + help="Range of input lengths for sampling prompts, " + 'specified as "min:max" (e.g., "128:256").', + ) + + parser.add_argument( + "--num-documents", + type=int, + default=8, + help="Range of input lengths for sampling prompts, " + 'specified as "min:max" (e.g., "128:256").', + ) + + parser.add_argument("--output-len", type=int, default=10) + + parser.add_argument( + "--repeat-count", + type=int, + default=2, + help="Number of times to repeat each prompt", + ) + + parser.add_argument( + "--repeat-mode", + type=str, + default="random", + help="The mode to repeat prompts. The supported " + 'modes are "random", "tile", and "interleave". ' + "See repeat_prompts() in the source code for details.", + ) + + parser.add_argument( + "--shuffle-seed", + type=int, + default=0, + help='Random seed when the repeat mode is "random"', + ) + + parser = EngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/benchmark_prefix_caching.py b/online_apiserver_test_maxbs/benchmarks/benchmark_prefix_caching.py new file mode 100644 index 0000000000000000000000000000000000000000..b5e2613de1cd4a0a2557cf09baa458aa15a6049c --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/benchmark_prefix_caching.py @@ -0,0 +1,278 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +Benchmark the efficiency of prefix caching. + +This script allows you to benchmark the performance of +a model with and without prefix caching using either fixed prompts +or prompts sampled from the ShareGPT dataset. + +Fixed example usage: + python benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --enable-prefix-caching \ + --num-prompts 1 \ + --repeat-count 100 \ + --input-length-range 128:256 + +ShareGPT example usage: + # This command samples 20 prompts with input lengths + # between 128 and 256 tokens from the ShareGPT dataset, + # then replicates each prompt 5 times. + python benchmark_prefix_caching.py \ + --model meta-llama/Llama-2-7b-chat-hf \ + --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json \ + --enable-prefix-caching \ + --num-prompts 20 \ + --repeat-count 5 \ + --input-length-range 128:256 +""" + +import dataclasses +import json +import random +import time +from typing import Optional + +from transformers import PreTrainedTokenizerBase + +from vllm import LLM, SamplingParams +from vllm.engine.arg_utils import EngineArgs +from vllm.utils import FlexibleArgumentParser + +try: + from vllm.transformers_utils.tokenizer import get_tokenizer +except ImportError: + from backend_request_func import get_tokenizer + +PROMPT = "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as fellows. You need to answer my question about the table.\n# Table\n|Opening|Opening|Sl. No.|Film|Cast|Director|Music Director|Notes|\n|----|----|----|----|----|----|----|----|\n|J A N|9|1|Agni Pushpam|Jayabharathi, Kamalahasan|Jeassy|M. K. Arjunan||\n|J A N|16|2|Priyamvada|Mohan Sharma, Lakshmi, KPAC Lalitha|K. S. Sethumadhavan|V. Dakshinamoorthy||\n|J A N|23|3|Yakshagaanam|Madhu, Sheela|Sheela|M. S. Viswanathan||\n|J A N|30|4|Paalkkadal|Sheela, Sharada|T. K. Prasad|A. T. Ummer||\n|F E B|5|5|Amma|Madhu, Srividya|M. Krishnan Nair|M. K. Arjunan||\n|F E B|13|6|Appooppan|Thikkurissi Sukumaran Nair, Kamal Haasan|P. Bhaskaran|M. S. Baburaj||\n|F E B|20|7|Srishti|Chowalloor Krishnankutty, Ravi Alummoodu|K. T. Muhammad|M. S. Baburaj||\n|F E B|20|8|Vanadevatha|Prem Nazir, Madhubala|Yusufali Kechery|G. Devarajan||\n|F E B|27|9|Samasya|Madhu, Kamalahaasan|K. Thankappan|Shyam||\n|F E B|27|10|Yudhabhoomi|K. P. Ummer, Vidhubala|Crossbelt Mani|R. K. Shekhar||\n|M A R|5|11|Seemantha Puthran|Prem Nazir, Jayabharathi|A. B. Raj|M. K. Arjunan||\n|M A R|12|12|Swapnadanam|Rani Chandra, Dr. Mohandas|K. G. George|Bhaskar Chandavarkar||\n|M A R|19|13|Thulavarsham|Prem Nazir, sreedevi, Sudheer|N. Sankaran Nair|V. Dakshinamoorthy||\n|M A R|20|14|Aruthu|Kaviyoor Ponnamma, Kamalahasan|Ravi|G. Devarajan||\n|M A R|26|15|Swimming Pool|Kamal Haasan, M. G. Soman|J. Sasikumar|M. K. Arjunan||\n\n# Question\nWhat' s the content in the (1,1) cells\n" # noqa: E501 + + +def test_prefix(llm=None, sampling_params=None, prompts=None): + start_time = time.time() + + llm.generate(prompts, sampling_params=sampling_params) + + end_time = time.time() + print(f"cost time {end_time - start_time}") + + +@dataclasses.dataclass +class Request: + prompt: str + prompt_len: int + output_len: int + + +def sample_tokens(tokenizer: PreTrainedTokenizerBase, length: int) -> list[int]: + vocab = tokenizer.get_vocab() + all_special_ids = set(tokenizer.all_special_ids) + + # Remove the special tokens. + return random.choices( + [v for k, v in vocab.items() if k not in all_special_ids], + k=length, + ) + + +def sample_requests_from_dataset( + dataset_path: str, + num_requests: int, + tokenizer: PreTrainedTokenizerBase, + input_length_range: tuple[int, int], + fixed_output_len: Optional[int], +) -> list[Request]: + if fixed_output_len is not None and fixed_output_len < 4: + raise ValueError("output_len too small") + + # Load the dataset. + with open(dataset_path) as f: + dataset = json.load(f) + # Filter out the conversations with less than 2 turns. + dataset = [data for data in dataset if len(data["conversations"]) >= 2] + # Only keep the first two turns of each conversation. + dataset = [ + (data["conversations"][0]["value"], data["conversations"][1]["value"]) + for data in dataset + ] + + # Shuffle the dataset. + random.shuffle(dataset) + + min_len, max_len = input_length_range + assert min_len >= 0 and max_len >= min_len, "input_length_range too small" + + # Filter out sequences that are too long or too short + filtered_requests: list[Request] = [] + + for i in range(len(dataset)): + if len(filtered_requests) == num_requests: + break + + # Tokenize the prompts and completions. + prompt_token_ids = tokenizer(dataset[i][0]).input_ids + prompt = tokenizer.decode(prompt_token_ids) + completion = dataset[i][1] + completion_token_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_token_ids) + output_len = ( + len(completion_token_ids) if fixed_output_len is None else fixed_output_len + ) + if min_len <= prompt_len <= max_len: + filtered_requests.append(Request(prompt, prompt_len, output_len)) + + return filtered_requests + + +def sample_requests_from_random( + num_requests: int, + tokenizer: PreTrainedTokenizerBase, + input_length_range: tuple[int, int], + fixed_output_len: Optional[int], + prefix_len: int, +) -> list[Request]: + requests = [] + prefix_token_ids = sample_tokens(tokenizer, prefix_len) + min_len, max_len = input_length_range + + for i in range(num_requests): + unique_part_token_ids = sample_tokens( + tokenizer, random.randint(min_len - prefix_len, max_len - prefix_len) + ) + prompt_token_ids = prefix_token_ids + unique_part_token_ids + prompt = tokenizer.decode(prompt_token_ids) + prompt_len = len(prompt_token_ids) + assert min_len <= prompt_len <= max_len, ( + f"prompt_len {prompt_len} out of range {min_len}:{max_len}" + ) + requests.append(Request(prompt, prompt_len, fixed_output_len)) + return requests + + +def repeat_and_sort_requests( + requests: list[Request], repeat_count: int, sort: bool = False +) -> list[str]: + repeated_requests = requests * repeat_count + if sort: + repeated_requests.sort(key=lambda x: x[1]) + else: + random.shuffle(repeated_requests) + return [req.prompt for req in repeated_requests] + + +def main(args): + tokenizer = get_tokenizer(args.model, trust_remote_code=True) + input_length_range = tuple(map(int, args.input_length_range.split(":"))) + random.seed(args.seed) + if args.dataset_path is not None: + if args.prefix_len > 0: + raise ValueError( + "prefix-len is not supported when dataset-path is provided." + ) + print(f"Start to sample {args.num_prompts} prompts from {args.dataset_path}") + filtered_requests = sample_requests_from_dataset( + dataset_path=args.dataset_path, + num_requests=args.num_prompts, + tokenizer=tokenizer, + input_length_range=input_length_range, + fixed_output_len=args.output_len, + ) + else: + print(f"Start to sample {args.num_prompts} prompts from random") + filtered_requests = sample_requests_from_random( + num_requests=args.num_prompts, + tokenizer=tokenizer, + input_length_range=input_length_range, + fixed_output_len=args.output_len, + prefix_len=args.prefix_len, + ) + + # Print some helpful stats of the requests. + print(f"Sampled {len(filtered_requests)} requests.") + prompt_lens = [req.prompt_len for req in filtered_requests] + print(f"Average input length: {sum(prompt_lens) / len(prompt_lens)}") + print(f"P50 input length: {sorted(prompt_lens)[len(prompt_lens) // 2]}") + print(f"Min Prompt Length: {min(prompt_lens)}") + print(f"Max Prompt Length: {max(prompt_lens)}") + + engine_args = EngineArgs.from_cli_args(args) + + llm = LLM(**dataclasses.asdict(engine_args)) + + sampling_params = SamplingParams( + temperature=0, + max_tokens=args.output_len, + detokenize=not args.disable_detokenize, + ) + + print("Testing filtered requests") + prompts = repeat_and_sort_requests( + filtered_requests, repeat_count=args.repeat_count, sort=args.sort + ) + + print("------start generating------") + test_prefix( + llm=llm, + prompts=prompts, + sampling_params=sampling_params, + ) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the performance with or without " + "automatic prefix caching." + ) + parser.add_argument( + "--dataset-path", type=str, default=None, help="Path to the dataset." + ) + parser.add_argument("--output-len", type=int, default=10) + parser.add_argument( + "--num-prompts", + type=int, + required=True, + help="Number of the prompts sampled from dataset", + ) + parser.add_argument( + "--repeat-count", + type=int, + default=1, + help="Number of times to repeat each prompt", + ) + parser.add_argument( + "--sort", action="store_true", help="Sort prompts by input length" + ) + parser.add_argument( + "--input-length-range", + type=str, + required=True, + help="Range of input lengths for sampling prompts," + 'specified as "min:max" (e.g., "128:256").', + ) + parser.add_argument( + "--prefix-len", + type=int, + default=0, + help="Specifies the length of a common prefix to be " + "added to the input prompt. The input-length-range will " + "subtract this length when filtering prompts. Only used " + "when dataset-path is not provided.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize responses (i.e. do not include " + "detokenization time in the latency measurement)" + ), + ) + + parser = EngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/benchmark_prioritization.py b/online_apiserver_test_maxbs/benchmarks/benchmark_prioritization.py new file mode 100644 index 0000000000000000000000000000000000000000..bb453791c1862bdd0883f0a47af6fddeef764f8f --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/benchmark_prioritization.py @@ -0,0 +1,222 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Benchmark offline prioritization.""" + +import argparse +import dataclasses +import json +import random +import time +from typing import Optional + +from transformers import AutoTokenizer, PreTrainedTokenizerBase + +from vllm.engine.arg_utils import EngineArgs +from vllm.utils import FlexibleArgumentParser + + +# Select a equi-probable random priority +def get_random_flag(): + return 0 if random.random() < 0.5 else 1 + + +def sample_requests( + dataset_path: str, + num_requests: int, + tokenizer: PreTrainedTokenizerBase, + fixed_output_len: Optional[int], +) -> list[tuple[str, int, int, int]]: + if fixed_output_len is not None and fixed_output_len < 4: + raise ValueError("output_len too small") + + # Load the dataset. + with open(dataset_path) as f: + dataset = json.load(f) + # Filter out the conversations with less than 2 turns. + dataset = [data for data in dataset if len(data["conversations"]) >= 2] + # Only keep the first two turns of each conversation. + dataset = [ + (data["conversations"][0]["value"], data["conversations"][1]["value"]) + for data in dataset + ] + + # Shuffle the dataset. + random.shuffle(dataset) + + # Filter out sequences that are too long or too short + filtered_dataset: list[tuple[str, int, int]] = [] + for i in range(len(dataset)): + if len(filtered_dataset) == num_requests: + break + + # Tokenize the prompts and completions. + prompt = dataset[i][0] + prompt_token_ids = tokenizer(prompt).input_ids + completion = dataset[i][1] + completion_token_ids = tokenizer(completion).input_ids + prompt_len = len(prompt_token_ids) + output_len = ( + len(completion_token_ids) if fixed_output_len is None else fixed_output_len + ) + if prompt_len < 4 or output_len < 4: + # Prune too short sequences. + continue + if prompt_len > 1024 or prompt_len + output_len > 2048: + # Prune too long sequences. + continue + + priority = get_random_flag() + + filtered_dataset.append((prompt, prompt_len, output_len, priority)) + + return filtered_dataset + + +def run_vllm( + requests: list[tuple[str, int, int]], + n: int, + engine_args: EngineArgs, + disable_detokenize: bool = False, +) -> float: + from vllm import LLM, SamplingParams + + llm = LLM(**dataclasses.asdict(engine_args)) + + assert all( + llm.llm_engine.model_config.max_model_len >= (request[1] + request[2]) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of" + " input_len and output_len for all requests." + ) + + # Add the requests to the engine. + prompts = [] + sampling_params = [] + priority = [] + for prompt, _, output_len, _priority in requests: + prompts.append(prompt) + priority.append(_priority) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=output_len, + detokenize=not disable_detokenize, + ) + ) + + start = time.perf_counter() + llm.generate(prompts, sampling_params, priority=priority, use_tqdm=True) + end = time.perf_counter() + return end - start + + +def main(args: argparse.Namespace): + print(args) + random.seed(args.seed) + + # Sample the requests. + tokenizer = AutoTokenizer.from_pretrained( + args.tokenizer, trust_remote_code=args.trust_remote_code + ) + if args.dataset is None: + # Synthesize a prompt with the given input length. + prompt = "hi" * (args.input_len - 1) + requests = [ + (prompt, args.input_len, args.output_len, get_random_flag()) + for _ in range(args.num_prompts) + ] + else: + requests = sample_requests( + args.dataset, args.num_prompts, tokenizer, args.output_len + ) + + if args.backend == "vllm": + elapsed_time = run_vllm( + requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize + ) + else: + raise ValueError(f"Unknown backend: {args.backend}") + total_num_tokens = sum( + prompt_len + output_len for _, prompt_len, output_len, priority in requests + ) + print( + f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, " + f"{total_num_tokens / elapsed_time:.2f} tokens/s" + ) + + # Output JSON results if specified + if args.output_json: + results = { + "elapsed_time": elapsed_time, + "num_requests": len(requests), + "total_num_tokens": total_num_tokens, + "requests_per_second": len(requests) / elapsed_time, + "tokens_per_second": total_num_tokens / elapsed_time, + } + with open(args.output_json, "w") as f: + json.dump(results, f, indent=4) + + +def create_argument_parser(): + parser = FlexibleArgumentParser(description="Benchmark the throughput.") + parser.add_argument( + "--backend", type=str, choices=["vllm", "hf", "mii"], default="vllm" + ) + parser.add_argument( + "--dataset", type=str, default=None, help="Path to the dataset." + ) + parser.add_argument( + "--input-len", + type=int, + default=None, + help="Input prompt length for each request", + ) + parser.add_argument( + "--output-len", + type=int, + default=None, + help="Output length for each request. Overrides the " + "output length from the dataset.", + ) + parser.add_argument( + "--n", type=int, default=1, help="Number of generated sequences per prompt." + ) + parser.add_argument( + "--num-prompts", type=int, default=200, help="Number of prompts to process." + ) + parser.add_argument( + "--output-json", + type=str, + default=None, + help="Path to save the throughput results in JSON format.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize responses (i.e. do not include " + "detokenization time in the latency measurement)" + ), + ) + + parser = EngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + if args.tokenizer is None: + args.tokenizer = args.model + if args.dataset is None: + assert args.input_len is not None + assert args.output_len is not None + else: + assert args.input_len is None + + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/benchmark_serving.py b/online_apiserver_test_maxbs/benchmarks/benchmark_serving.py new file mode 100644 index 0000000000000000000000000000000000000000..9b235266dff1aac95d6fe0509f4a5f01641579cc --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/benchmark_serving.py @@ -0,0 +1,1378 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +r"""Benchmark online serving throughput. + +On the server side, run one of the following commands: + vLLM OpenAI API server + vllm serve \ + --swap-space 16 \ + --disable-log-requests + +On the client side, run: + python benchmarks/benchmark_serving.py \ + --backend \ + --model \ + --dataset-name sharegpt \ + --dataset-path \ + --request-rate \ # By default is inf + --num-prompts # By default is 1000 + + when using tgi backend, add + --endpoint /generate_stream + to the end of the command above. +""" + +import argparse +import asyncio +import gc +import json +import os +import random +import time +import warnings +from collections.abc import AsyncGenerator, Iterable +from dataclasses import dataclass +from datetime import datetime +from typing import Any, Literal, Optional + +import numpy as np +from tqdm.asyncio import tqdm +from transformers import PreTrainedTokenizerBase + +from backend_request_func import ( + ASYNC_REQUEST_FUNCS, + OPENAI_COMPATIBLE_BACKENDS, + RequestFuncInput, + RequestFuncOutput, +) + +try: + from vllm.transformers_utils.tokenizer import get_tokenizer +except ImportError: + from backend_request_func import get_tokenizer + +try: + from vllm.utils import FlexibleArgumentParser +except ImportError: + from argparse import ArgumentParser as FlexibleArgumentParser + +from benchmark_dataset import ( + AIMODataset, + ASRDataset, + BurstGPTDataset, + ConversationDataset, + CustomDataset, + HuggingFaceDataset, + InstructCoderDataset, + MTBenchDataset, + NextEditPredictionDataset, + RandomDataset, + SampleRequest, + ShareGPTDataset, + SonnetDataset, + VisionArenaDataset, +) +from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json + +MILLISECONDS_TO_SECONDS_CONVERSION = 1000 + + +@dataclass +class BenchmarkMetrics: + completed: int + total_input: int + total_output: int + request_throughput: float + request_goodput: float + output_throughput: float + total_token_throughput: float + mean_ttft_ms: float + median_ttft_ms: float + std_ttft_ms: float + percentiles_ttft_ms: list[tuple[float, float]] + mean_tpot_ms: float + median_tpot_ms: float + std_tpot_ms: float + percentiles_tpot_ms: list[tuple[float, float]] + mean_itl_ms: float + median_itl_ms: float + std_itl_ms: float + percentiles_itl_ms: list[tuple[float, float]] + # E2EL stands for end-to-end latency per request. + # It is the time taken on the client side from sending + # a request to receiving a complete response. + mean_e2el_ms: float + median_e2el_ms: float + std_e2el_ms: float + percentiles_e2el_ms: list[tuple[float, float]] + + +def _get_current_request_rate( + ramp_up_strategy: Optional[Literal["linear", "exponential"]], + ramp_up_start_rps: Optional[int], + ramp_up_end_rps: Optional[int], + request_index: int, + total_requests: int, + request_rate: float, +) -> float: + if ( + ramp_up_strategy + and ramp_up_start_rps is not None + and ramp_up_end_rps is not None + ): + progress = request_index / max(total_requests - 1, 1) + if ramp_up_strategy == "linear": + increase = (ramp_up_end_rps - ramp_up_start_rps) * progress + return ramp_up_start_rps + increase + elif ramp_up_strategy == "exponential": + ratio = ramp_up_end_rps / ramp_up_start_rps + return ramp_up_start_rps * (ratio**progress) + else: + raise ValueError(f"Unknown ramp-up strategy: {ramp_up_strategy}") + return request_rate + + +async def get_request( + input_requests: list[SampleRequest], + request_rate: float, + burstiness: float = 1.0, + ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None, + ramp_up_start_rps: Optional[int] = None, + ramp_up_end_rps: Optional[int] = None, +) -> AsyncGenerator[tuple[SampleRequest, float], None]: + """ + Asynchronously generates requests at a specified rate + with OPTIONAL burstiness and OPTIONAL ramp-up strategy. + + Args: + input_requests: + A list of input requests, each represented as a SampleRequest. + request_rate: + The rate at which requests are generated (requests/s). + burstiness (optional): + The burstiness factor of the request generation. + Only takes effect when request_rate is not inf. + Default value is 1, which follows a Poisson process. + Otherwise, the request intervals follow a gamma distribution. + A lower burstiness value (0 < burstiness < 1) results + in more bursty requests, while a higher burstiness value + (burstiness > 1) results in a more uniform arrival of requests. + ramp_up_strategy (optional): + The ramp-up strategy. Can be "linear" or "exponential". + If None, uses constant request rate (specified by request_rate). + ramp_up_start_rps (optional): + The starting request rate for ramp-up. + ramp_up_end_rps (optional): + The ending request rate for ramp-up. + """ + assert burstiness > 0, ( + f"A positive burstiness factor is expected, but given {burstiness}." + ) + # Convert to list to get length for ramp-up calculations + if isinstance(input_requests, Iterable) and not isinstance(input_requests, list): + input_requests = list(input_requests) + + total_requests = len(input_requests) + request_index = 0 + + for request in input_requests: + current_request_rate = _get_current_request_rate( + ramp_up_strategy, + ramp_up_start_rps, + ramp_up_end_rps, + request_index, + total_requests, + request_rate, + ) + + yield request, current_request_rate + + request_index += 1 + + if current_request_rate == float("inf"): + # If the request rate is infinity, then we don't need to wait. + continue + + theta = 1.0 / (current_request_rate * burstiness) + + # Sample the request interval from the gamma distribution. + # If burstiness is 1, it follows exponential distribution. + interval = np.random.gamma(shape=burstiness, scale=theta) + # The next request will be sent after the interval. + await asyncio.sleep(interval) + + +def calculate_metrics( + input_requests: list[SampleRequest], + outputs: list[RequestFuncOutput], + dur_s: float, + tokenizer: PreTrainedTokenizerBase, + selected_percentile_metrics: list[str], + selected_percentiles: list[float], + goodput_config_dict: dict[str, float], +) -> tuple[BenchmarkMetrics, list[int]]: + actual_output_lens: list[int] = [] + total_input = 0 + completed = 0 + good_completed = 0 + itls: list[float] = [] + tpots: list[float] = [] + all_tpots: list[float] = [] + ttfts: list[float] = [] + e2els: list[float] = [] + for i in range(len(outputs)): + if outputs[i].success: + output_len = outputs[i].output_tokens + + if not output_len: + # We use the tokenizer to count the number of output tokens + # for some serving backends instead of looking at + # len(outputs[i].itl) since multiple output tokens may be + # bundled together + # Note : this may inflate the output token count slightly + output_len = len( + tokenizer( + outputs[i].generated_text, add_special_tokens=False + ).input_ids + ) + actual_output_lens.append(output_len) + total_input += input_requests[i].prompt_len + tpot = 0 + if output_len > 1: + latency_minus_ttft = outputs[i].latency - outputs[i].ttft + tpot = latency_minus_ttft / (output_len - 1) + tpots.append(tpot) + # Note: if output_len <= 1, we regard tpot as 0 for goodput + all_tpots.append(tpot) + itls += outputs[i].itl + ttfts.append(outputs[i].ttft) + e2els.append(outputs[i].latency) + completed += 1 + else: + actual_output_lens.append(0) + + if goodput_config_dict: + valid_metrics = [] + slo_values = [] + + if "ttft" in goodput_config_dict: + valid_metrics.append(ttfts) + slo_values.append( + goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "tpot" in goodput_config_dict: + valid_metrics.append(all_tpots) + slo_values.append( + goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "e2el" in goodput_config_dict: + valid_metrics.append(e2els) + slo_values.append( + goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + + for req_metric in zip(*valid_metrics): + is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)]) + if is_good_req: + good_completed += 1 + + if completed == 0: + warnings.warn( + "All requests failed. This is likely due to a misconfiguration " + "on the benchmark arguments.", + stacklevel=2, + ) + metrics = BenchmarkMetrics( + completed=completed, + total_input=total_input, + total_output=sum(actual_output_lens), + request_throughput=completed / dur_s, + request_goodput=good_completed / dur_s, + output_throughput=sum(actual_output_lens) / dur_s, + total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s, + mean_ttft_ms=np.mean(ttfts or 0) + * 1000, # ttfts is empty if streaming is not supported by backend + std_ttft_ms=np.std(ttfts or 0) * 1000, + median_ttft_ms=np.median(ttfts or 0) * 1000, + percentiles_ttft_ms=[ + (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles + ], + mean_tpot_ms=np.mean(tpots or 0) * 1000, + std_tpot_ms=np.std(tpots or 0) * 1000, + median_tpot_ms=np.median(tpots or 0) * 1000, + percentiles_tpot_ms=[ + (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles + ], + mean_itl_ms=np.mean(itls or 0) * 1000, + std_itl_ms=np.std(itls or 0) * 1000, + median_itl_ms=np.median(itls or 0) * 1000, + percentiles_itl_ms=[ + (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles + ], + mean_e2el_ms=np.mean(e2els or 0) * 1000, + std_e2el_ms=np.std(e2els or 0) * 1000, + median_e2el_ms=np.median(e2els or 0) * 1000, + percentiles_e2el_ms=[ + (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles + ], + ) + + return metrics, actual_output_lens + + +async def benchmark( + backend: str, + api_url: str, + base_url: str, + model_id: str, + model_name: str, + tokenizer: PreTrainedTokenizerBase, + input_requests: list[SampleRequest], + logprobs: Optional[int], + request_rate: float, + burstiness: float, + disable_tqdm: bool, + profile: bool, + selected_percentile_metrics: list[str], + selected_percentiles: list[float], + ignore_eos: bool, + goodput_config_dict: dict[str, float], + max_concurrency: Optional[int], + lora_modules: Optional[Iterable[str]], + extra_body: Optional[dict], + ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None, + ramp_up_start_rps: Optional[int] = None, + ramp_up_end_rps: Optional[int] = None, +): + if backend in ASYNC_REQUEST_FUNCS: + request_func = ASYNC_REQUEST_FUNCS[backend] + else: + raise ValueError(f"Unknown backend: {backend}") + + print("Starting initial single prompt test run...") + test_prompt, test_prompt_len, test_output_len, test_mm_content = ( + input_requests[0].prompt, + input_requests[0].prompt_len, + input_requests[0].expected_output_len, + input_requests[0].multi_modal_data, + ) + + assert test_mm_content is None or isinstance(test_mm_content, dict) + test_input = RequestFuncInput( + model=model_id, + model_name=model_name, + prompt=test_prompt, + api_url=api_url, + prompt_len=test_prompt_len, + output_len=test_output_len, + logprobs=logprobs, + multi_modal_content=test_mm_content, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + + test_output = await request_func(request_func_input=test_input) + if not test_output.success: + raise ValueError( + "Initial test run failed - Please make sure benchmark arguments " + f"are correctly specified. Error: {test_output.error}" + ) + else: + print("Initial test run completed. Starting main benchmark run...") + + if lora_modules: + # For each input request, choose a LoRA module at random. + lora_modules = iter( + [random.choice(lora_modules) for _ in range(len(input_requests))] + ) + + if profile: + print("Starting profiler...") + profile_input = RequestFuncInput( + model=model_id, + model_name=model_name, + prompt=test_prompt, + api_url=base_url + "/start_profile", + prompt_len=test_prompt_len, + output_len=test_output_len, + logprobs=logprobs, + multi_modal_content=test_mm_content, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler started") + + distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution" + + if ramp_up_strategy is not None: + print( + f"Traffic ramp-up strategy: {ramp_up_strategy}. Will increase " + f"RPS from {ramp_up_start_rps} to {ramp_up_end_rps} RPS over " + "the duration of the benchmark." + ) + else: + print(f"Traffic request rate: {request_rate} RPS.") + + print(f"Burstiness factor: {burstiness} ({distribution})") + print(f"Maximum request concurrency: {max_concurrency}") + + pbar = None if disable_tqdm else tqdm(total=len(input_requests)) + + # This can be used once the minimum Python version is 3.10 or higher, + # and it will simplify the code in limited_request_func. + # semaphore = (asyncio.Semaphore(max_concurrency) + # if max_concurrency else contextlib.nullcontext()) + semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None + + async def limited_request_func(request_func_input, pbar): + if semaphore is None: + return await request_func(request_func_input=request_func_input, pbar=pbar) + async with semaphore: + return await request_func(request_func_input=request_func_input, pbar=pbar) + + benchmark_start_time = time.perf_counter() + tasks: list[asyncio.Task] = [] + + rps_change_events = [] + last_int_rps = -1 + if ramp_up_strategy is not None and ramp_up_start_rps is not None: + last_int_rps = ramp_up_start_rps + rps_change_events.append( + { + "rps": last_int_rps, + "timestamp": datetime.now().isoformat(), + } + ) + + async for request, current_request_rate in get_request( + input_requests, + request_rate, + burstiness, + ramp_up_strategy, + ramp_up_start_rps, + ramp_up_end_rps, + ): + if ramp_up_strategy is not None: + current_int_rps = int(current_request_rate) + if current_int_rps > last_int_rps: + timestamp = datetime.now().isoformat() + for rps_val in range(last_int_rps + 1, current_int_rps + 1): + rps_change_events.append({"rps": rps_val, "timestamp": timestamp}) + last_int_rps = current_int_rps + + prompt, prompt_len, output_len, mm_content = ( + request.prompt, + request.prompt_len, + request.expected_output_len, + request.multi_modal_data, + ) + req_model_id, req_model_name = model_id, model_name + if lora_modules: + req_lora_module = next(lora_modules) + req_model_id, req_model_name = req_lora_module, req_lora_module + + request_func_input = RequestFuncInput( + model=req_model_id, + model_name=req_model_name, + prompt=prompt, + api_url=api_url, + prompt_len=prompt_len, + output_len=output_len, + logprobs=logprobs, + multi_modal_content=mm_content, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + task = limited_request_func(request_func_input=request_func_input, pbar=pbar) + tasks.append(asyncio.create_task(task)) + outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks) + + if profile: + print("Stopping profiler...") + profile_input = RequestFuncInput( + model=model_id, + prompt=test_prompt, + api_url=base_url + "/stop_profile", + prompt_len=test_prompt_len, + output_len=test_output_len, + logprobs=logprobs, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler stopped") + + if pbar is not None: + pbar.close() + + benchmark_duration = time.perf_counter() - benchmark_start_time + + metrics, actual_output_lens = calculate_metrics( + input_requests=input_requests, + outputs=outputs, + dur_s=benchmark_duration, + tokenizer=tokenizer, + selected_percentile_metrics=selected_percentile_metrics, + selected_percentiles=selected_percentiles, + goodput_config_dict=goodput_config_dict, + ) + + print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="=")) + print("{:<40} {:<10}".format("Successful requests:", metrics.completed)) + print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration)) + print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input)) + print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output)) + print( + "{:<40} {:<10.2f}".format( + "Request throughput (req/s):", metrics.request_throughput + ) + ) + if goodput_config_dict: + print( + "{:<40} {:<10.2f}".format( + "Request goodput (req/s):", metrics.request_goodput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Output token throughput (tok/s):", metrics.output_throughput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Total Token throughput (tok/s):", metrics.total_token_throughput + ) + ) + + result = { + "duration": benchmark_duration, + "completed": metrics.completed, + "total_input_tokens": metrics.total_input, + "total_output_tokens": metrics.total_output, + "request_throughput": metrics.request_throughput, + "request_goodput": metrics.request_goodput if goodput_config_dict else None, + "output_throughput": metrics.output_throughput, + "total_token_throughput": metrics.total_token_throughput, + "input_lens": [output.prompt_len for output in outputs], + "output_lens": actual_output_lens, + "ttfts": [output.ttft for output in outputs], + "itls": [output.itl for output in outputs], + "generated_texts": [output.generated_text for output in outputs], + "errors": [output.error for output in outputs], + } + + if rps_change_events: + result["rps_change_events"] = rps_change_events + + def process_one_metric( + # E.g., "ttft" + metric_attribute_name: str, + # E.g., "TTFT" + metric_name: str, + # E.g., "Time to First Token" + metric_header: str, + ): + # This function prints and adds statistics of the specified + # metric. + if metric_attribute_name not in selected_percentile_metrics: + return + print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-")) + print( + "{:<40} {:<10.2f}".format( + f"Mean {metric_name} (ms):", + getattr(metrics, f"mean_{metric_attribute_name}_ms"), + ) + ) + print( + "{:<40} {:<10.2f}".format( + f"Median {metric_name} (ms):", + getattr(metrics, f"median_{metric_attribute_name}_ms"), + ) + ) + result[f"mean_{metric_attribute_name}_ms"] = getattr( + metrics, f"mean_{metric_attribute_name}_ms" + ) + result[f"median_{metric_attribute_name}_ms"] = getattr( + metrics, f"median_{metric_attribute_name}_ms" + ) + result[f"std_{metric_attribute_name}_ms"] = getattr( + metrics, f"std_{metric_attribute_name}_ms" + ) + for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"): + p_word = str(int(p)) if int(p) == p else str(p) + print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value)) + result[f"p{p_word}_{metric_attribute_name}_ms"] = value + + process_one_metric("ttft", "TTFT", "Time to First Token") + process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)") + process_one_metric("itl", "ITL", "Inter-token Latency") + process_one_metric("e2el", "E2EL", "End-to-end Latency") + + print("=" * 50) + + return result + + +def check_goodput_args(args): + # Check and parse goodput arguments + goodput_config_dict = {} + VALID_NAMES = ["ttft", "tpot", "e2el"] + if args.goodput: + goodput_config_dict = parse_goodput(args.goodput) + for slo_name, slo_val in goodput_config_dict.items(): + if slo_name not in VALID_NAMES: + raise ValueError( + f"Invalid metric name found, {slo_name}: {slo_val}. " + "The service level objective name should be one of " + f"{str(VALID_NAMES)}. " + ) + if slo_val < 0: + raise ValueError( + f"Invalid value found, {slo_name}: {slo_val}. " + "The service level objective value should be " + "non-negative." + ) + return goodput_config_dict + + +def parse_goodput(slo_pairs): + goodput_config_dict = {} + try: + for slo_pair in slo_pairs: + slo_name, slo_val = slo_pair.split(":") + goodput_config_dict[slo_name] = float(slo_val) + except ValueError as err: + raise argparse.ArgumentTypeError( + "Invalid format found for service level objectives. " + 'Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is a " + "number in milliseconds." + ) from err + return goodput_config_dict + + +def save_to_pytorch_benchmark_format( + args: argparse.Namespace, results: dict[str, Any], file_name: str +) -> None: + metrics = [ + "median_ttft_ms", + "mean_ttft_ms", + "std_ttft_ms", + "p99_ttft_ms", + "mean_tpot_ms", + "median_tpot_ms", + "std_tpot_ms", + "p99_tpot_ms", + "median_itl_ms", + "mean_itl_ms", + "std_itl_ms", + "p99_itl_ms", + ] + # These raw data might be useful, but they are rather big. They can be added + # later if needed + ignored_metrics = ["ttfts", "itls", "generated_texts", "errors"] + pt_records = convert_to_pytorch_benchmark_format( + args=args, + metrics={k: [results[k]] for k in metrics}, + extra_info={ + k: results[k] + for k in results + if k not in metrics and k not in ignored_metrics + }, + ) + if pt_records: + # Don't use json suffix here as we don't want CI to pick it up + pt_file = f"{os.path.splitext(file_name)[0]}.pytorch.json" + write_to_json(pt_file, pt_records) + + +def main(args: argparse.Namespace): + print(args) + random.seed(args.seed) + np.random.seed(args.seed) + + backend = args.backend + model_id = args.model + model_name = args.served_model_name + tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model + tokenizer_mode = args.tokenizer_mode + + # Validate ramp-up arguments + if args.ramp_up_strategy is not None: + if args.request_rate != float("inf"): + raise ValueError( + "When using ramp-up, do not specify --request-rate. " + "The request rate will be controlled by ramp-up parameters. " + "Please remove the --request-rate argument." + ) + if args.ramp_up_start_rps is None or args.ramp_up_end_rps is None: + raise ValueError( + "When using --ramp-up-strategy, both --ramp-up-start-rps and " + "--ramp-up-end-rps must be specified" + ) + if args.ramp_up_start_rps < 0 or args.ramp_up_end_rps < 0: + raise ValueError("Ramp-up start and end RPS must be non-negative") + if args.ramp_up_start_rps > args.ramp_up_end_rps: + raise ValueError("Ramp-up start RPS must be less than end RPS") + if args.ramp_up_strategy == "exponential" and args.ramp_up_start_rps == 0: + raise ValueError("For exponential ramp-up, the start RPS cannot be 0.") + + if args.base_url is not None: + api_url = f"{args.base_url}{args.endpoint}" + base_url = f"{args.base_url}" + else: + api_url = f"http://{args.host}:{args.port}{args.endpoint}" + base_url = f"http://{args.host}:{args.port}" + + tokenizer = get_tokenizer( + tokenizer_id, + tokenizer_mode=tokenizer_mode, + trust_remote_code=args.trust_remote_code, + ) + + if args.dataset_name is None: + raise ValueError( + "Please specify '--dataset-name' and the corresponding " + "'--dataset-path' if required." + ) + + if args.dataset_name == "custom": + dataset = CustomDataset(dataset_path=args.dataset_path) + input_requests = dataset.sample( + num_requests=args.num_prompts, + tokenizer=tokenizer, + output_len=args.custom_output_len, + skip_chat_template=args.custom_skip_chat_template, + ) + + elif args.dataset_name == "sonnet": + dataset = SonnetDataset(dataset_path=args.dataset_path) + # For the "sonnet" dataset, formatting depends on the backend. + if args.backend == "openai-chat": + input_requests = dataset.sample( + num_requests=args.num_prompts, + input_len=args.sonnet_input_len, + output_len=args.sonnet_output_len, + prefix_len=args.sonnet_prefix_len, + tokenizer=tokenizer, + return_prompt_formatted=False, + ) + else: + assert tokenizer.chat_template or tokenizer.default_chat_template, ( + "Tokenizer/model must have chat template for sonnet dataset." + ) + input_requests = dataset.sample( + num_requests=args.num_prompts, + input_len=args.sonnet_input_len, + output_len=args.sonnet_output_len, + prefix_len=args.sonnet_prefix_len, + tokenizer=tokenizer, + return_prompt_formatted=True, + ) + + elif args.dataset_name == "hf": + # all following datasets are implemented from the + # HuggingFaceDataset base class + if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS: + dataset_class = VisionArenaDataset + args.hf_split = "train" + args.hf_subset = None + elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS: + dataset_class = InstructCoderDataset + args.hf_split = "train" + elif args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS: + dataset_class = MTBenchDataset + args.hf_split = "train" + elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS: + dataset_class = ConversationDataset + elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS: + dataset_class = AIMODataset + args.hf_split = "train" + elif args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS: # noqa: E501 + dataset_class = NextEditPredictionDataset + args.hf_split = "train" + elif args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS: + dataset_class = ASRDataset + args.hf_split = "train" + else: + supported_datasets = set( + [ + dataset_name + for cls in HuggingFaceDataset.__subclasses__() + for dataset_name in cls.SUPPORTED_DATASET_PATHS + ] + ) + raise ValueError( + f"Unsupported dataset path: {args.dataset_path}. " + "Huggingface dataset only supports dataset_path" + f" from one of following: {supported_datasets}. " + "Please consider contributing if you would " + "like to add support for additional dataset formats." + ) + + if dataset_class.IS_MULTIMODAL and backend not in [ + "openai-chat", + "openai-audio", + ]: + # multi-modal benchmark is only available on OpenAI Chat backend. + raise ValueError( + "Multi-modal content is only supported on 'openai-chat' and " + "'openai-audio' backend." + ) + input_requests = dataset_class( + dataset_path=args.dataset_path, + dataset_subset=args.hf_subset, + dataset_split=args.hf_split, + random_seed=args.seed, + ).sample( + num_requests=args.num_prompts, + tokenizer=tokenizer, + output_len=args.hf_output_len, + ) + + else: + # For datasets that follow a similar structure, use a mapping. + dataset_mapping = { + "sharegpt": lambda: ShareGPTDataset( + random_seed=args.seed, dataset_path=args.dataset_path + ).sample( + tokenizer=tokenizer, + num_requests=args.num_prompts, + output_len=args.sharegpt_output_len, + ), + "burstgpt": lambda: BurstGPTDataset( + random_seed=args.seed, dataset_path=args.dataset_path + ).sample(tokenizer=tokenizer, num_requests=args.num_prompts), + "random": lambda: RandomDataset(dataset_path=args.dataset_path).sample( + tokenizer=tokenizer, + num_requests=args.num_prompts, + prefix_len=args.random_prefix_len, + input_len=args.random_input_len, + output_len=args.random_output_len, + range_ratio=args.random_range_ratio, + ), + } + + try: + input_requests = dataset_mapping[args.dataset_name]() + except KeyError as err: + raise ValueError(f"Unknown dataset: {args.dataset_name}") from err + goodput_config_dict = check_goodput_args(args) + + # Collect the sampling parameters. + sampling_params = { + k: v + for k, v in { + "top_p": args.top_p, + "top_k": args.top_k, + "min_p": args.min_p, + "temperature": args.temperature, + }.items() + if v is not None + } + + # Sampling parameters are only supported by openai-compatible backend. + if sampling_params and args.backend not in OPENAI_COMPATIBLE_BACKENDS: + raise ValueError( + "Sampling parameters are only supported by openai-compatible backends." + ) + + if "temperature" not in sampling_params: + sampling_params["temperature"] = 0.0 # Default to greedy decoding. + + if args.backend == "llama.cpp": + # Disable prompt caching in llama.cpp backend + sampling_params["cache_prompt"] = False + + # Avoid GC processing "static" data - reduce pause times. + gc.collect() + gc.freeze() + + benchmark_result = asyncio.run( + benchmark( + backend=backend, + api_url=api_url, + base_url=base_url, + model_id=model_id, + model_name=model_name, + tokenizer=tokenizer, + input_requests=input_requests, + logprobs=args.logprobs, + request_rate=args.request_rate, + burstiness=args.burstiness, + disable_tqdm=args.disable_tqdm, + profile=args.profile, + selected_percentile_metrics=args.percentile_metrics.split(","), + selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")], + ignore_eos=args.ignore_eos, + goodput_config_dict=goodput_config_dict, + max_concurrency=args.max_concurrency, + lora_modules=args.lora_modules, + extra_body=sampling_params, + ramp_up_strategy=args.ramp_up_strategy, + ramp_up_start_rps=args.ramp_up_start_rps, + ramp_up_end_rps=args.ramp_up_end_rps, + ) + ) + + # Save config and results to json + if args.save_result or args.append_result: + result_json: dict[str, Any] = {} + + # Setup + current_dt = datetime.now().strftime("%Y%m%d-%H%M%S") + result_json["date"] = current_dt + result_json["backend"] = backend + result_json["model_id"] = model_id + result_json["tokenizer_id"] = tokenizer_id + result_json["num_prompts"] = args.num_prompts + + # Metadata + if args.metadata: + for item in args.metadata: + if "=" in item: + kvstring = item.split("=") + result_json[kvstring[0].strip()] = kvstring[1].strip() + else: + raise ValueError( + "Invalid metadata format. Please use KEY=VALUE format." + ) + # Traffic + result_json["request_rate"] = ( + args.request_rate if args.request_rate < float("inf") else "inf" + ) + result_json["burstiness"] = args.burstiness + result_json["max_concurrency"] = args.max_concurrency + + if args.ramp_up_strategy is not None: + result_json["ramp_up_strategy"] = args.ramp_up_strategy + result_json["ramp_up_start_rps"] = args.ramp_up_start_rps + result_json["ramp_up_end_rps"] = args.ramp_up_end_rps + + # Merge with benchmark result + result_json = {**result_json, **benchmark_result} + + if not args.save_detailed: + # Remove fields with too many data points + for field in [ + "input_lens", + "output_lens", + "ttfts", + "itls", + "generated_texts", + "errors", + ]: + if field in result_json: + del result_json[field] + if field in benchmark_result: + del benchmark_result[field] + + # Save to file + base_model_id = model_id.split("/")[-1] + max_concurrency_str = ( + f"-concurrency{args.max_concurrency}" + if args.max_concurrency is not None + else "" + ) + if args.ramp_up_strategy is not None: + file_name = f"{backend}-ramp-up-{args.ramp_up_strategy}-{args.ramp_up_start_rps}qps-{args.ramp_up_end_rps}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json" # noqa + else: + file_name = f"{backend}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json" # noqa + if args.result_filename: + file_name = args.result_filename + if args.result_dir: + os.makedirs(args.result_dir, exist_ok=True) + file_name = os.path.join(args.result_dir, file_name) + with open( + file_name, mode="a+" if args.append_result else "w", encoding="utf-8" + ) as outfile: + # Append a newline. + if args.append_result and outfile.tell() != 0: + outfile.write("\n") + json.dump(result_json, outfile) + save_to_pytorch_benchmark_format(args, result_json, file_name) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the online serving throughput." + ) + parser.add_argument( + "--backend", + type=str, + default="vllm", + choices=list(ASYNC_REQUEST_FUNCS.keys()), + ) + parser.add_argument( + "--base-url", + type=str, + default=None, + help="Server or API base url if not using http host and port.", + ) + # Use 127.0.0.1 here instead of localhost to force the use of ipv4 + parser.add_argument("--host", type=str, default="127.0.0.1") + parser.add_argument("--port", type=int, default=8000) + parser.add_argument( + "--endpoint", + type=str, + default="/v1/completions", + help="API endpoint.", + ) + parser.add_argument( + "--dataset-name", + type=str, + default="sharegpt", + choices=["sharegpt", "burstgpt", "sonnet", "random", "hf", "custom"], + help="Name of the dataset to benchmark on.", + ) + parser.add_argument( + "--dataset-path", + type=str, + default=None, + help="Path to the sharegpt/sonnet dataset. " + "Or the huggingface dataset ID if using HF dataset.", + ) + parser.add_argument( + "--max-concurrency", + type=int, + default=None, + help="Maximum number of concurrent requests. This can be used " + "to help simulate an environment where a higher level component " + "is enforcing a maximum number of concurrent requests. While the " + "--request-rate argument controls the rate at which requests are " + "initiated, this argument will control how many are actually allowed " + "to execute at a time. This means that when used in combination, the " + "actual request rate may be lower than specified with --request-rate, " + "if the server is not processing requests fast enough to keep up.", + ) + + parser.add_argument( + "--model", + type=str, + required=True, + help="Name of the model.", + ) + parser.add_argument( + "--tokenizer", + type=str, + help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 + ) + parser.add_argument("--use-beam-search", action="store_true") + parser.add_argument( + "--num-prompts", + type=int, + default=1000, + help="Number of prompts to process.", + ) + parser.add_argument( + "--logprobs", + type=int, + default=None, + help=( + "Number of logprobs-per-token to compute & return as part of " + "the request. If unspecified, then either (1) if beam search " + "is disabled, no logprobs are computed & a single dummy " + "logprob is returned for each token; or (2) if beam search " + "is enabled 1 logprob per token is computed" + ), + ) + parser.add_argument( + "--request-rate", + type=float, + default=float("inf"), + help="Number of requests per second. If this is inf, " + "then all the requests are sent at time 0. " + "Otherwise, we use Poisson process or gamma distribution " + "to synthesize the request arrival times.", + ) + parser.add_argument( + "--burstiness", + type=float, + default=1.0, + help="Burstiness factor of the request generation. " + "Only take effect when request_rate is not inf. " + "Default value is 1, which follows Poisson process. " + "Otherwise, the request intervals follow a gamma distribution. " + "A lower burstiness value (0 < burstiness < 1) results in more " + "bursty requests. A higher burstiness value (burstiness > 1) " + "results in a more uniform arrival of requests.", + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument( + "--trust-remote-code", + action="store_true", + help="Trust remote code from huggingface", + ) + parser.add_argument( + "--disable-tqdm", + action="store_true", + help="Specify to disable tqdm progress bar.", + ) + parser.add_argument( + "--profile", + action="store_true", + help="Use Torch Profiler. The endpoint must be launched with " + "VLLM_TORCH_PROFILER_DIR to enable profiler.", + ) + parser.add_argument( + "--save-result", + action="store_true", + help="Specify to save benchmark results to a json file", + ) + parser.add_argument( + "--save-detailed", + action="store_true", + help="When saving the results, whether to include per request " + "information such as response, error, ttfs, tpots, etc.", + ) + parser.add_argument( + "--append-result", + action="store_true", + help="Append the benchmark result to the existing json file.", + ) + parser.add_argument( + "--metadata", + metavar="KEY=VALUE", + nargs="*", + help="Key-value pairs (e.g, --metadata version=0.3.3 tp=1) " + "for metadata of this run to be saved in the result JSON file " + "for record keeping purposes.", + ) + parser.add_argument( + "--result-dir", + type=str, + default=None, + help="Specify directory to save benchmark json results." + "If not specified, results are saved in the current directory.", + ) + parser.add_argument( + "--result-filename", + type=str, + default=None, + help="Specify the filename to save benchmark json results." + "If not specified, results will be saved in " + "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json" + " format.", + ) + parser.add_argument( + "--ignore-eos", + action="store_true", + help="Set ignore_eos flag when sending the benchmark request." + "Warning: ignore_eos is not supported in deepspeed_mii and tgi.", + ) + parser.add_argument( + "--percentile-metrics", + type=str, + default="ttft,tpot,itl", + help="Comma-separated list of selected metrics to report percentils. " + "This argument specifies the metrics to report percentiles. " + 'Allowed metric names are "ttft", "tpot", "itl", "e2el". ' + 'Default value is "ttft,tpot,itl".', + ) + parser.add_argument( + "--metric-percentiles", + type=str, + default="99", + help="Comma-separated list of percentiles for selected metrics. " + 'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". ' + 'Default value is "99". ' + 'Use "--percentile-metrics" to select metrics.', + ) + parser.add_argument( + "--goodput", + nargs="+", + required=False, + help='Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is in " + 'milliseconds. Multiple "KEY:VALUE" pairs can be provided, ' + "separated by spaces. Allowed request level metric names are " + '"ttft", "tpot", "e2el". For more context on the definition of ' + "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 " + "and the blog: https://hao-ai-lab.github.io/blogs/distserve", + ) + + # group for dataset specific arguments + custom_group = parser.add_argument_group("custom dataset options") + custom_group.add_argument( + "--custom-output-len", + type=int, + default=256, + help="Number of output tokens per request, used only for custom dataset.", + ) + custom_group.add_argument( + "--custom-skip-chat-template", + action="store_true", + help="Skip applying chat template to prompt, used only for custom dataset.", + ) + + sonnet_group = parser.add_argument_group("sonnet dataset options") + sonnet_group.add_argument( + "--sonnet-input-len", + type=int, + default=550, + help="Number of input tokens per request, used only for sonnet dataset.", + ) + sonnet_group.add_argument( + "--sonnet-output-len", + type=int, + default=150, + help="Number of output tokens per request, used only for sonnet dataset.", + ) + sonnet_group.add_argument( + "--sonnet-prefix-len", + type=int, + default=200, + help="Number of prefix tokens per request, used only for sonnet dataset.", + ) + + sharegpt_group = parser.add_argument_group("sharegpt dataset options") + sharegpt_group.add_argument( + "--sharegpt-output-len", + type=int, + default=None, + help="Output length for each request. Overrides the output length " + "from the ShareGPT dataset.", + ) + + random_group = parser.add_argument_group("random dataset options") + random_group.add_argument( + "--random-input-len", + type=int, + default=1024, + help="Number of input tokens per request, used only for random sampling.", + ) + random_group.add_argument( + "--random-output-len", + type=int, + default=128, + help="Number of output tokens per request, used only for random sampling.", + ) + random_group.add_argument( + "--random-range-ratio", + type=float, + default=0.0, + help="Range ratio for sampling input/output length, " + "used only for random sampling. Must be in the range [0, 1) to define " + "a symmetric sampling range" + "[length * (1 - range_ratio), length * (1 + range_ratio)].", + ) + random_group.add_argument( + "--random-prefix-len", + type=int, + default=0, + help=( + "Number of fixed prefix tokens before the random context " + "in a request. " + "The total input length is the sum of `random-prefix-len` and " + "a random " + "context length sampled from [input_len * (1 - range_ratio), " + "input_len * (1 + range_ratio)]." + ), + ) + + hf_group = parser.add_argument_group("hf dataset options") + hf_group.add_argument( + "--hf-subset", type=str, default=None, help="Subset of the HF dataset." + ) + hf_group.add_argument( + "--hf-split", type=str, default=None, help="Split of the HF dataset." + ) + hf_group.add_argument( + "--hf-output-len", + type=int, + default=None, + help="Output length for each request. Overrides the output lengths " + "from the sampled HF dataset.", + ) + + sampling_group = parser.add_argument_group("sampling parameters") + sampling_group.add_argument( + "--top-p", + type=float, + default=None, + help="Top-p sampling parameter. Only has effect on openai-compatible backends.", + ) + sampling_group.add_argument( + "--top-k", + type=int, + default=None, + help="Top-k sampling parameter. Only has effect on openai-compatible backends.", + ) + sampling_group.add_argument( + "--min-p", + type=float, + default=None, + help="Min-p sampling parameter. Only has effect on openai-compatible backends.", + ) + sampling_group.add_argument( + "--temperature", + type=float, + default=None, + help="Temperature sampling parameter. Only has effect on " + "openai-compatible backends. If not specified, default to greedy " + "decoding (i.e. temperature==0.0).", + ) + + parser.add_argument( + "--tokenizer-mode", + type=str, + default="auto", + choices=["auto", "slow", "mistral", "custom"], + help='The tokenizer mode.\n\n* "auto" will use the ' + 'fast tokenizer if available.\n* "slow" will ' + "always use the slow tokenizer. \n* " + '"mistral" will always use the `mistral_common` tokenizer. \n*' + '"custom" will use --tokenizer to select the preregistered tokenizer.', + ) + + parser.add_argument( + "--served-model-name", + type=str, + default=None, + help="The model name used in the API. " + "If not specified, the model name will be the " + "same as the ``--model`` argument. ", + ) + + parser.add_argument( + "--lora-modules", + nargs="+", + default=None, + help="A subset of LoRA module names passed in when " + "launching the server. For each request, the " + "script chooses a LoRA module at random.", + ) + + parser.add_argument( + "--ramp-up-strategy", + type=str, + default=None, + choices=["linear", "exponential"], + help="The ramp-up strategy. This would be used to " + "ramp up the request rate from initial RPS to final " + "RPS rate (specified by --ramp-up-start-rps and --ramp-up-end-rps). " + "over the duration of the benchmark.", + ) + parser.add_argument( + "--ramp-up-start-rps", + type=int, + default=None, + help="The starting request rate for ramp-up (RPS). " + "Needs to be specified when --ramp-up-strategy is used.", + ) + parser.add_argument( + "--ramp-up-end-rps", + type=int, + default=None, + help="The ending request rate for ramp-up (RPS). " + "Needs to be specified when --ramp-up-strategy is used.", + ) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/benchmark_serving_structured_output.py b/online_apiserver_test_maxbs/benchmarks/benchmark_serving_structured_output.py new file mode 100644 index 0000000000000000000000000000000000000000..e23a5a9e2233df685998fbae0ae8cc8910b572cf --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/benchmark_serving_structured_output.py @@ -0,0 +1,1043 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +r"""Benchmark online serving throughput with structured outputs. + +On the server side, run one of the following commands: + (vLLM OpenAI API server) + vllm serve --disable-log-requests + +On the client side, run: + python benchmarks/benchmark_serving_structured_output.py \ + --backend \ + --model \ + --dataset json \ + --structured-output-ratio 1.0 \ + --request-rate 10 \ + --num-prompts 1000 + + when using tgi backend, add + --endpoint /generate_stream + to the end of the command above. +""" + +import argparse +import asyncio +import copy +import dataclasses +import json +import os +import random +import time +import uuid +import warnings +from collections.abc import AsyncGenerator +from dataclasses import dataclass +from typing import Optional + +import datasets +import numpy as np +import pandas as pd +from tqdm.asyncio import tqdm +from transformers import PreTrainedTokenizerBase + +from backend_request_func import ( + ASYNC_REQUEST_FUNCS, + RequestFuncInput, + RequestFuncOutput, +) + +try: + from vllm.transformers_utils.tokenizer import get_tokenizer +except ImportError: + from backend_request_func import get_tokenizer + +try: + from vllm.utils import FlexibleArgumentParser +except ImportError: + from argparse import ArgumentParser as FlexibleArgumentParser + +from vllm.v1.structured_output.backend_xgrammar import ( + has_xgrammar_unsupported_json_features, +) + +MILLISECONDS_TO_SECONDS_CONVERSION = 1000 + + +@dataclass +class BenchmarkMetrics: + completed: int + total_input: int + total_output: int + request_throughput: float + request_goodput: float + output_throughput: float + total_token_throughput: float + mean_ttft_ms: float + median_ttft_ms: float + std_ttft_ms: float + percentiles_ttft_ms: list[tuple[float, float]] + mean_tpot_ms: float + median_tpot_ms: float + std_tpot_ms: float + percentiles_tpot_ms: list[tuple[float, float]] + mean_itl_ms: float + median_itl_ms: float + std_itl_ms: float + percentiles_itl_ms: list[tuple[float, float]] + # E2EL stands for end-to-end latency per request. + # It is the time taken on the client side from sending + # a request to receiving a complete response. + mean_e2el_ms: float + median_e2el_ms: float + std_e2el_ms: float + percentiles_e2el_ms: list[tuple[float, float]] + + +@dataclasses.dataclass +class SampleRequest: + """A class representing a single inference request for benchmarking. + + Attributes: + prompt: The input text prompt for the model. + multi_modal_data: Optional dictionary containing multi-modal data (e.g. + images). + prompt_len: The length of the prompt in tokens. + expected_output_len: The expected length of the output in tokens. + """ + + prompt: str + prompt_len: int + expected_output_len: int + schema: dict + structure_type: str + completion: str = None + + +def sample_requests( + tokenizer: PreTrainedTokenizerBase, args: argparse.Namespace +) -> list[SampleRequest]: + if args.dataset == "json" or args.dataset == "json-unique": + if args.json_schema_path is None: + dir_path = os.path.dirname(os.path.realpath(__file__)) + args.json_schema_path = os.path.join( + dir_path, "structured_schemas", "structured_schema_1.json" + ) + json_schemas = [] + with open(args.json_schema_path) as f: + schema = json.load(f) + + if args.dataset == "json-unique": + json_schemas = [copy.deepcopy(schema) for _ in range(args.num_prompts)] + for i in range(len(json_schemas)): + if "properties" not in json_schemas[i]: + json_schemas[i]["properties"] = {} + json_schemas[i]["properties"][f"__optional_field_{uuid.uuid4()}"] = { + "type": "string", + "description": "An unique optional field to avoid cached schemas", + } + else: + json_schemas = [schema] * args.num_prompts + + def gen_prompt(index: int): + return f"Generate an example of a brief user profile given the following schema: {json.dumps(get_schema(index))}" # noqa: E501 + + def get_schema(index: int): + return json_schemas[index % len(json_schemas)] + + requests = [ + SampleRequest( + prompt=gen_prompt(i), + prompt_len=len(tokenizer(gen_prompt(i)).input_ids), + expected_output_len=args.output_len, + schema=get_schema(i), + structure_type=args.structure_type, + ) + for i in range(args.num_prompts) + ] + + elif args.dataset == "grammar": + schema = """ + root ::= select_statement + + select_statement ::= "SELECT " column " from " table " where " condition + + column ::= "col_1 " | "col_2 " + + table ::= "table_1 " | "table_2 " + + condition ::= column "= " number + + number ::= "1 " | "2 " + """ + prompt = "Generate an SQL query to show the 'username' \ + and 'email' from the 'users' table." + + input_len = len(tokenizer(prompt).input_ids) + print(f"Input length of the prompt: {input_len} tokens") + requests = [ + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=schema, + structure_type=args.structure_type, + ) + for _ in range(args.num_prompts) + ] + + elif args.dataset == "regex": + regex = r"\w+@\w+\.com\n" + args.regex = regex + prompt = "Generate an email address for Alan Turing, \ + who works in Enigma. End in .com and new line. \ + Example result: alan.turing@enigma.com\n" + + input_len = len(tokenizer(prompt).input_ids) + print(f"Input length of the prompt: {input_len} tokens") + requests = [ + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=regex, + structure_type=args.structure_type, + ) + for _ in range(args.num_prompts) + ] + + elif args.dataset == "choice": + choice = ["Positive", "Negative"] + args.choice = choice + prompt = "Classify this sentiment: vLLM is wonderful!" + input_len = len(tokenizer(prompt).input_ids) + print(f"Input length of the prompt: {input_len} tokens") + requests = [ + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=choice, + structure_type=args.structure_type, + ) + for _ in range(args.num_prompts) + ] + + elif args.dataset == "xgrammar_bench": + requests: list[SampleRequest] = [] + dataset = datasets.load_dataset("NousResearch/json-mode-eval", split="train") + full_dataset_len = len(dataset) + + def _filter_func(item): + import json + + schema = json.loads(item["schema"]) + return not has_xgrammar_unsupported_json_features(schema) + + dataset = dataset.filter(_filter_func) + num_filtered_out = full_dataset_len - len(dataset) + print( + f"dataset has {len(dataset)} entries after filtering " + f"out {num_filtered_out} entries with unsupported features" + ) + len_dataset = len(dataset) + for data_point_idx in range(args.num_prompts): + idx = data_point_idx + while idx >= len_dataset: + idx -= len_dataset + schema = dataset["schema"][idx] + prompt = tokenizer.apply_chat_template( + dataset["prompt"][idx], tokenize=False, add_generation_prompt=True + ) + input_len = len(tokenizer(prompt).input_ids) + completion = dataset["completion"][idx] + + requests.append( + SampleRequest( + prompt=prompt, + prompt_len=input_len, + expected_output_len=args.output_len, + schema=schema, + structure_type=args.structure_type, + completion=completion, + ) + ) + + return requests + + +async def get_request( + input_requests: list[SampleRequest], + request_rate: float, + burstiness: float = 1.0, +) -> AsyncGenerator[tuple[int, SampleRequest], None]: + """ + Asynchronously generates requests at a specified rate + with OPTIONAL burstiness. + + Args: + input_requests: + A list of input requests, each represented as a tuple. + request_rate: + The rate at which requests are generated (requests/s). + burstiness (optional): + The burstiness factor of the request generation. + Only takes effect when request_rate is not inf. + Default value is 1, which follows a Poisson process. + Otherwise, the request intervals follow a gamma distribution. + A lower burstiness value (0 < burstiness < 1) results + in more bursty requests, while a higher burstiness value + (burstiness > 1) results in a more uniform arrival of requests. + """ + input_requests = iter(input_requests) + + # Calculate scale parameter theta to maintain the desired request_rate. + assert burstiness > 0, ( + f"A positive burstiness factor is expected, but given {burstiness}." + ) + theta = 1.0 / (request_rate * burstiness) + + for i, request in enumerate(input_requests): + yield i, request + + if request_rate == float("inf"): + # If the request rate is infinity, then we don't need to wait. + continue + + # Sample the request interval from the gamma distribution. + # If burstiness is 1, it follows exponential distribution. + interval = np.random.gamma(shape=burstiness, scale=theta) + # The next request will be sent after the interval. + await asyncio.sleep(interval) + + +def calculate_metrics( + input_requests: list[tuple[str, int, int]], + outputs: list[RequestFuncOutput], + dur_s: float, + tokenizer: PreTrainedTokenizerBase, + selected_percentile_metrics: list[str], + selected_percentiles: list[float], + goodput_config_dict: Optional[dict[str, float]] = None, +) -> tuple[BenchmarkMetrics, list[int]]: + actual_output_lens: list[int] = [] + total_input = 0 + completed = 0 + good_completed = 0 + itls: list[float] = [] + tpots: list[float] = [] + all_tpots: list[float] = [] + ttfts: list[float] = [] + e2els: list[float] = [] + for i in range(len(outputs)): + if outputs[i].success: + # We use the tokenizer to count the number of output tokens for all + # serving backends instead of looking at len(outputs[i].itl) since + # multiple output tokens may be bundled together + # Note : this may inflate the output token count slightly + output_len = len( + tokenizer(outputs[i].generated_text, add_special_tokens=False).input_ids + ) + actual_output_lens.append(output_len) + total_input += input_requests[i].prompt_len + tpot = 0 + if output_len > 1: + latency_minus_ttft = outputs[i].latency - outputs[i].ttft + tpot = latency_minus_ttft / (output_len - 1) + tpots.append(tpot) + outputs[i].tpot = tpot + # Note: if output_len <= 1, we regard tpot as 0 for goodput + all_tpots.append(tpot) + itls += outputs[i].itl + ttfts.append(outputs[i].ttft) + e2els.append(outputs[i].latency) + completed += 1 + else: + actual_output_lens.append(0) + + if goodput_config_dict: + valid_metrics = [] + slo_values = [] + + if "ttft" in goodput_config_dict: + valid_metrics.append(ttfts) + slo_values.append( + goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "tpot" in goodput_config_dict: + valid_metrics.append(all_tpots) + slo_values.append( + goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + if "e2el" in goodput_config_dict: + valid_metrics.append(e2els) + slo_values.append( + goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION + ) + + for req_metric in zip(*valid_metrics): + is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)]) + if is_good_req: + good_completed += 1 + + if completed == 0: + warnings.warn( + "All requests failed. This is likely due to a misconfiguration " + "on the benchmark arguments.", + stacklevel=2, + ) + metrics = BenchmarkMetrics( + completed=completed, + total_input=total_input, + total_output=sum(actual_output_lens), + request_throughput=completed / dur_s, + request_goodput=good_completed / dur_s, + output_throughput=sum(actual_output_lens) / dur_s, + total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s, + mean_ttft_ms=np.mean(ttfts or 0) + * 1000, # ttfts is empty if streaming is not supported by backend + std_ttft_ms=np.std(ttfts or 0) * 1000, + median_ttft_ms=np.median(ttfts or 0) * 1000, + percentiles_ttft_ms=[ + (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles + ], + mean_tpot_ms=np.mean(tpots or 0) * 1000, + std_tpot_ms=np.std(tpots or 0) * 1000, + median_tpot_ms=np.median(tpots or 0) * 1000, + percentiles_tpot_ms=[ + (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles + ], + mean_itl_ms=np.mean(itls or 0) * 1000, + std_itl_ms=np.std(itls or 0) * 1000, + median_itl_ms=np.median(itls or 0) * 1000, + percentiles_itl_ms=[ + (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles + ], + mean_e2el_ms=np.mean(e2els or 0) * 1000, + std_e2el_ms=np.std(e2els or 0) * 1000, + median_e2el_ms=np.median(e2els or 0) * 1000, + percentiles_e2el_ms=[ + (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles + ], + ) + + return metrics, actual_output_lens + + +async def benchmark( + backend: str, + api_url: str, + base_url: str, + model_id: str, + tokenizer: PreTrainedTokenizerBase, + input_requests: list[SampleRequest], + request_rate: float, + burstiness: float, + disable_tqdm: bool, + profile: bool, + selected_percentile_metrics: list[str], + selected_percentiles: list[str], + ignore_eos: bool, + max_concurrency: Optional[int], + structured_output_ratio: float, + goodput_config_dict: Optional[dict[str, float]] = None, +): + if backend in ASYNC_REQUEST_FUNCS: + request_func = ASYNC_REQUEST_FUNCS[backend] + else: + raise ValueError(f"Unknown backend: {backend}") + + def prepare_extra_body(request) -> dict: + extra_body = {} + # Add the schema to the extra_body + extra_body[request.structure_type] = request.schema + return extra_body + + print("Starting initial single prompt test run...") + structured_output_req_idx = random.sample( + range(len(input_requests)), int(len(input_requests) * structured_output_ratio) + ) + + test_request = input_requests[0] + test_req_extra_body = ( + prepare_extra_body(test_request) if 0 in structured_output_req_idx else None + ) + test_input = RequestFuncInput( + model=model_id, + prompt=test_request.prompt, + api_url=api_url, + prompt_len=test_request.prompt_len, + output_len=test_request.expected_output_len, + ignore_eos=ignore_eos, + extra_body=test_req_extra_body, + ) + test_output = await request_func(request_func_input=test_input) + if not test_output.success: + raise ValueError( + "Initial test run failed - Please make sure benchmark arguments " + f"are correctly specified. Error: {test_output.error}" + ) + else: + print("Initial test run completed. Starting main benchmark run...") + + if profile: + print("Starting profiler...") + profile_input = RequestFuncInput( + model=model_id, + prompt=test_request.prompt, + api_url=base_url + "/start_profile", + prompt_len=test_request.prompt_len, + output_len=test_request.expected_output_len, + ignore_eos=ignore_eos, + extra_body=test_req_extra_body, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler started") + + distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution" + + print(f"Traffic request rate: {request_rate}") + print(f"Burstiness factor: {burstiness} ({distribution})") + print(f"Maximum request concurrency: {max_concurrency}") + + pbar = None if disable_tqdm else tqdm(total=len(input_requests)) + + # This can be used once the minimum Python version is 3.10 or higher, + # and it will simplify the code in limited_request_func. + # semaphore = (asyncio.Semaphore(max_concurrency) + # if max_concurrency else contextlib.nullcontext()) + semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None + + async def limited_request_func(request_func_input, pbar): + if semaphore is None: + return await request_func(request_func_input=request_func_input, pbar=pbar) + async with semaphore: + return await request_func(request_func_input=request_func_input, pbar=pbar) + + benchmark_start_time = time.perf_counter() + tasks: list[asyncio.Task] = [] + expected: list[str] = [] + async for i, request in get_request(input_requests, request_rate, burstiness): + extra_body = ( + prepare_extra_body(request) if i in structured_output_req_idx else None + ) + request_func_input = RequestFuncInput( + model=model_id, + prompt=request.prompt, + api_url=api_url, + prompt_len=request.prompt_len, + output_len=request.expected_output_len, + ignore_eos=ignore_eos, + extra_body=extra_body, + ) + expected.append(request.completion) + tasks.append( + asyncio.create_task( + limited_request_func(request_func_input=request_func_input, pbar=pbar) + ) + ) + outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks) + + if profile: + print("Stopping profiler...") + profile_input = RequestFuncInput( + model=model_id, + prompt=test_request.prompt, + api_url=base_url + "/stop_profile", + prompt_len=test_request.prompt_len, + output_len=test_request.expected_output_len, + extra_body={test_request.structure_type: test_request.schema}, + ) + profile_output = await request_func(request_func_input=profile_input) + if profile_output.success: + print("Profiler stopped") + + if pbar is not None: + pbar.close() + + benchmark_duration = time.perf_counter() - benchmark_start_time + + metrics, actual_output_lens = calculate_metrics( + input_requests=input_requests, + outputs=outputs, + dur_s=benchmark_duration, + tokenizer=tokenizer, + selected_percentile_metrics=selected_percentile_metrics, + selected_percentiles=selected_percentiles, + goodput_config_dict=goodput_config_dict, + ) + + print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="=")) + print("{:<40} {:<10}".format("Successful requests:", metrics.completed)) + print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration)) + print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input)) + print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output)) + print( + "{:<40} {:<10.2f}".format( + "Request throughput (req/s):", metrics.request_throughput + ) + ) + if goodput_config_dict: + print( + "{:<40} {:<10.2f}".format( + "Request goodput (req/s):", metrics.request_goodput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Output token throughput (tok/s):", metrics.output_throughput + ) + ) + print( + "{:<40} {:<10.2f}".format( + "Total Token throughput (tok/s):", metrics.total_token_throughput + ) + ) + + result = { + "duration": benchmark_duration, + "completed": metrics.completed, + "total_input_tokens": metrics.total_input, + "total_output_tokens": metrics.total_output, + "request_throughput": metrics.request_throughput, + "output_throughput": metrics.output_throughput, + "total_token_throughput": metrics.total_token_throughput, + "ttft_description": pd.Series([output.ttft for output in outputs]) + .describe() + .to_dict(), + "tpot_description": pd.Series([output.tpot for output in outputs]) + .describe() + .to_dict(), + "input_lens": [output.prompt_len for output in outputs], + "output_lens": actual_output_lens, + "ttfts": [output.ttft for output in outputs], + "itls": [output.itl for output in outputs], + "errors": [output.error for output in outputs], + } + + ret = [ + {"generated": output.generated_text, "expected": gt} + for output, gt in zip(outputs, expected) + ] + + def process_one_metric( + # E.g., "ttft" + metric_attribute_name: str, + # E.g., "TTFT" + metric_name: str, + # E.g., "Time to First Token" + metric_header: str, + ): + # This function prints and adds statistics of the specified + # metric. + if metric_attribute_name not in selected_percentile_metrics: + return + print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-")) + print( + "{:<40} {:<10.2f}".format( + f"Mean {metric_name} (ms):", + getattr(metrics, f"mean_{metric_attribute_name}_ms"), + ) + ) + print( + "{:<40} {:<10.2f}".format( + f"Median {metric_name} (ms):", + getattr(metrics, f"median_{metric_attribute_name}_ms"), + ) + ) + result[f"mean_{metric_attribute_name}_ms"] = getattr( + metrics, f"mean_{metric_attribute_name}_ms" + ) + result[f"median_{metric_attribute_name}_ms"] = getattr( + metrics, f"median_{metric_attribute_name}_ms" + ) + result[f"std_{metric_attribute_name}_ms"] = getattr( + metrics, f"std_{metric_attribute_name}_ms" + ) + for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"): + p_word = str(int(p)) if int(p) == p else str(p) + print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value)) + result[f"p{p_word}_{metric_attribute_name}_ms"] = value + + process_one_metric("ttft", "TTFT", "Time to First Token") + process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)") + process_one_metric("itl", "ITL", "Inter-token Latency") + process_one_metric("e2el", "E2EL", "End-to-end Latency") + + print("=" * 50) + + return result, ret + + +def evaluate(ret, args): + def _eval_correctness_json(expected, actual): + # extract json string from string using regex + import regex as re + + actual = actual.replace("\n", "").replace(" ", "").strip() + try: + actual = re.search(r"\{.*\}", actual).group() + actual = json.loads(actual) + except Exception: + return False + + return True + + def _eval_correctness_choice(expected, actual): + return actual in args.choice + + def _eval_correctness_regex(expected, actual): + import regex as re + + return re.match(args.regex, actual) is not None + + def _eval_correctness(expected, actual): + if args.structure_type == "guided_json": + return _eval_correctness_json(expected, actual) + elif args.structure_type == "guided_regex": + return _eval_correctness_regex(expected, actual) + elif args.structure_type == "guided_choice": + return _eval_correctness_choice(expected, actual) + else: + return None + + scores = [] + for res in ret: + score = _eval_correctness(res["expected"], res["generated"]) + res["correctness"] = score + scores.append(score) + + not_none_scores = [score for score in scores if score is not None] + + return ( + (sum(not_none_scores) / len(not_none_scores) * 100) + if len(not_none_scores) > 0 + else None + ) + + +def parse_goodput(slo_pairs): + goodput_config_dict = {} + try: + for slo_pair in slo_pairs: + slo_name, slo_val = slo_pair.split(":") + goodput_config_dict[slo_name] = float(slo_val) + except ValueError as err: + raise argparse.ArgumentTypeError( + "Invalid format found for service level objectives. " + 'Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is a " + "number in milliseconds." + ) from err + return goodput_config_dict + + +def check_goodput_args(args): + goodput_config_dict = {} + VALID_NAMES = ["ttft", "tpot", "e2el"] + if args.goodput: + goodput_config_dict = parse_goodput(args.goodput) + for slo_name, slo_val in goodput_config_dict.items(): + if slo_name not in VALID_NAMES: + raise ValueError( + f"Invalid metric name found, {slo_name}: {slo_val}. " + "The service level objective name should be one of " + f"{str(VALID_NAMES)}. " + ) + if slo_val < 0: + raise ValueError( + f"Invalid value found, {slo_name}: {slo_val}. " + "The service level objective value should be " + "non-negative." + ) + return goodput_config_dict + + +def main(args: argparse.Namespace): + print(args) + random.seed(args.seed) + np.random.seed(args.seed) + + backend = args.backend + model_id = args.model + tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model + + if args.base_url is not None: + api_url = f"{args.base_url}{args.endpoint}" + base_url = f"{args.base_url}" + else: + api_url = f"http://{args.host}:{args.port}{args.endpoint}" + base_url = f"http://{args.host}:{args.port}" + + tokenizer = get_tokenizer( + tokenizer_id, + trust_remote_code=args.trust_remote_code, + tokenizer_mode=args.tokenizer_mode, + ) + + if args.dataset == "grammar": + args.structure_type = "guided_grammar" + elif args.dataset == "regex": + args.structure_type = "guided_regex" + elif args.dataset == "choice": + args.structure_type = "guided_choice" + else: + args.structure_type = "guided_json" + + if args.no_structured_output: + args.structured_output_ratio = 0 + if args.save_results: + result_file_name = f"{args.structured_output_ratio}guided" + result_file_name += f"_{backend}" + result_file_name += f"_{args.request_rate}qps" + result_file_name += f"_{args.model.split('/')[-1]}" + result_file_name += f"_{args.dataset}" + result_file_name += f"_{args.num_prompts}" + result_file_name += f"_out{args.output_len}" + result_file_name += ".txt" + else: + result_file_name = None + + input_requests = sample_requests(tokenizer, args) + + goodput_config_dict = check_goodput_args(args) + + benchmark_result, ret = asyncio.run( + benchmark( + backend=backend, + api_url=api_url, + base_url=base_url, + model_id=model_id, + tokenizer=tokenizer, + input_requests=input_requests, + request_rate=args.request_rate, + burstiness=args.burstiness, + disable_tqdm=args.disable_tqdm, + profile=args.profile, + selected_percentile_metrics=args.percentile_metrics.split(","), + selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")], + ignore_eos=args.ignore_eos, + max_concurrency=args.max_concurrency, + structured_output_ratio=args.structured_output_ratio, + goodput_config_dict=goodput_config_dict, + ) + ) + + # Save config and results to json + score = evaluate(ret, args) + print("correct_rate(%)", score, "\n") + if args.save_results: + results = { + "backend": backend, + "model_id": model_id, + "tokenizer_id": tokenizer_id, + "num_prompts": args.num_prompts, + "request_rate": args.request_rate + if args.request_rate < float("inf") + else "inf", + "burstiness": args.burstiness, + "max_concurrency": args.max_concurrency, + "correct_rate(%)": score, + } + results = {"outputs": ret, **results, **benchmark_result} + + # Save to file + if args.result_filename: + result_file_name = args.result_filename + if args.result_dir: + result_file_name = os.path.join(args.result_dir, result_file_name) + with open(result_file_name, "w", encoding="utf-8") as outfile: + json.dump(results, outfile, indent=4) + + +def create_argument_parser(): + parser = FlexibleArgumentParser( + description="Benchmark the online serving throughput." + ) + parser.add_argument( + "--backend", + type=str, + default="vllm", + choices=list(ASYNC_REQUEST_FUNCS.keys()), + ) + parser.add_argument( + "--base-url", + type=str, + default=None, + help="Server or API base url if not using http host and port.", + ) + # Use 127.0.0.1 here instead of localhost to force the use of ipv4 + parser.add_argument("--host", type=str, default="127.0.0.1") + parser.add_argument("--port", type=int, default=8000) + parser.add_argument( + "--endpoint", + type=str, + default="/v1/completions", + help="API endpoint.", + ) + parser.add_argument( + "--dataset", + default="json", + choices=["json", "json-unique", "grammar", "regex", "choice", "xgrammar_bench"], + ) + parser.add_argument( + "--json-schema-path", type=str, default=None, help="Path to json schema." + ) + parser.add_argument( + "--max-concurrency", + type=int, + default=None, + help="Maximum number of concurrent requests. This can be used " + "to help simulate an environment where a higher level component " + "is enforcing a maximum number of concurrent requests. While the " + "--request-rate argument controls the rate at which requests are " + "initiated, this argument will control how many are actually allowed " + "to execute at a time. This means that when used in combination, the " + "actual request rate may be lower than specified with --request-rate, " + "if the server is not processing requests fast enough to keep up.", + ) + parser.add_argument( + "--model", + type=str, + required=True, + help="Name of the model.", + ) + parser.add_argument( + "--tokenizer", + type=str, + help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 + ) + parser.add_argument( + "--tokenizer-mode", + type=str, + default="auto", + help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 + ) + parser.add_argument( + "--num-prompts", + type=int, + default=1000, + help="Number of prompts to process.", + ) + parser.add_argument( + "--output-len", + type=int, + default=128, + help="Number of output tokens.", + ) + parser.add_argument( + "--request-rate", + type=float, + default=float("inf"), + help="Number of requests per second. If this is inf, " + "then all the requests are sent at time 0. " + "Otherwise, we use Poisson process or gamma distribution " + "to synthesize the request arrival times.", + ) + parser.add_argument( + "--burstiness", + type=float, + default=1.0, + help="Burstiness factor of the request generation. " + "Only take effect when request_rate is not inf. " + "Default value is 1, which follows Poisson process. " + "Otherwise, the request intervals follow a gamma distribution. " + "A lower burstiness value (0 < burstiness < 1) results in more " + "bursty requests. A higher burstiness value (burstiness > 1) " + "results in a more uniform arrival of requests.", + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument( + "--trust-remote-code", + action="store_true", + help="Trust remote code from huggingface", + ) + parser.add_argument( + "--disable-tqdm", + action="store_true", + help="Specify to disable tqdm progress bar.", + ) + parser.add_argument( + "--save-results", + action="store_true", + help="Specify to save benchmark results to a json file", + ) + parser.add_argument( + "--profile", + action="store_true", + help="Use Torch Profiler. The endpoint must be launched with " + "VLLM_TORCH_PROFILER_DIR to enable profiler.", + ) + parser.add_argument( + "--result-dir", + type=str, + default=None, + help="Specify directory to save benchmark json results." + "If not specified, results are saved in the current directory.", + ) + parser.add_argument( + "--result-filename", + type=str, + default=None, + help="Specify the filename to save benchmark json results." + "If not specified, results will be saved in " + "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json" + " format.", + ) + parser.add_argument( + "--ignore-eos", + action="store_true", + help="Set ignore_eos flag when sending the benchmark request." + "Warning: ignore_eos is not supported in deepspeed_mii and tgi.", + ) + parser.add_argument( + "--percentile-metrics", + type=str, + default="ttft,tpot,itl", + help="Comma-separated list of selected metrics to report percentils. " + "This argument specifies the metrics to report percentiles. " + 'Allowed metric names are "ttft", "tpot", "itl", "e2el". ' + 'Default value is "ttft,tpot,itl".', + ) + parser.add_argument( + "--metric-percentiles", + type=str, + default="99", + help="Comma-separated list of percentiles for selected metrics. " + 'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". ' + 'Default value is "99". ' + 'Use "--percentile-metrics" to select metrics.', + ) + parser.add_argument( + "--goodput", + nargs="+", + required=False, + help='Specify service level objectives for goodput as "KEY:VALUE" ' + "pairs, where the key is a metric name, and the value is in " + 'milliseconds. Multiple "KEY:VALUE" pairs can be provided, ' + "separated by spaces. Allowed request level metric names are " + '"ttft", "tpot", "e2el". For more context on the definition of ' + "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 " + "and the blog: https://hao-ai-lab.github.io/blogs/distserve", + ) + + parser.add_argument( + "--no-structured-output", + action="store_true", + default=False, + help="Whether to disable JSON decoding or not.", + ) + parser.add_argument( + "--structured-output-ratio", + type=float, + default=1.0, + help="Ratio of Structured Outputs requests", + ) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/benchmark_throughput.py b/online_apiserver_test_maxbs/benchmarks/benchmark_throughput.py new file mode 100644 index 0000000000000000000000000000000000000000..0ded34c70badd28d972b3c993ddb7881b7b94aff --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/benchmark_throughput.py @@ -0,0 +1,730 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Benchmark offline inference throughput.""" + +import argparse +import dataclasses +import json +import os +import random +import time +import warnings +from typing import Any, Optional, Union + +import torch +import uvloop +from tqdm import tqdm +from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase + +from benchmark_dataset import ( + AIMODataset, + BurstGPTDataset, + ConversationDataset, + InstructCoderDataset, + RandomDataset, + SampleRequest, + ShareGPTDataset, + SonnetDataset, + VisionArenaDataset, +) +from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json +from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs +from vllm.entrypoints.openai.api_server import ( + build_async_engine_client_from_engine_args, +) +from vllm.inputs import TextPrompt, TokensPrompt +from vllm.lora.request import LoRARequest +from vllm.outputs import RequestOutput +from vllm.sampling_params import BeamSearchParams +from vllm.utils import FlexibleArgumentParser, merge_async_iterators + + +def run_vllm( + requests: list[SampleRequest], + n: int, + engine_args: EngineArgs, + disable_detokenize: bool = False, +) -> tuple[float, Optional[list[RequestOutput]]]: + from vllm import LLM, SamplingParams + + llm = LLM(**dataclasses.asdict(engine_args)) + assert all( + llm.llm_engine.model_config.max_model_len + >= (request.prompt_len + request.expected_output_len) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of" + " prompt_len and expected_output_len for all requests." + ) + # Add the requests to the engine. + prompts: list[Union[TextPrompt, TokensPrompt]] = [] + sampling_params: list[SamplingParams] = [] + for request in requests: + prompts.append( + TokensPrompt( + prompt_token_ids=request.prompt["prompt_token_ids"], + multi_modal_data=request.multi_modal_data, + ) + if "prompt_token_ids" in request.prompt + else TextPrompt( + prompt=request.prompt, multi_modal_data=request.multi_modal_data + ) + ) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=request.expected_output_len, + detokenize=not disable_detokenize, + ) + ) + lora_requests: Optional[list[LoRARequest]] = None + if engine_args.enable_lora: + lora_requests = [request.lora_request for request in requests] + + use_beam_search = False + + outputs = None + if not use_beam_search: + start = time.perf_counter() + outputs = llm.generate( + prompts, sampling_params, lora_request=lora_requests, use_tqdm=True + ) + end = time.perf_counter() + else: + assert lora_requests is None, "BeamSearch API does not support LoRA" + prompts = [request.prompt for request in requests] + # output_len should be the same for all requests. + output_len = requests[0].expected_output_len + for request in requests: + assert request.expected_output_len == output_len + start = time.perf_counter() + llm.beam_search( + prompts, + BeamSearchParams( + beam_width=n, + max_tokens=output_len, + ignore_eos=True, + ), + ) + end = time.perf_counter() + return end - start, outputs + + +def run_vllm_chat( + requests: list[SampleRequest], + n: int, + engine_args: EngineArgs, + disable_detokenize: bool = False, +) -> tuple[float, list[RequestOutput]]: + """ + Run vLLM chat benchmark. This function is recommended ONLY for benchmarking + multimodal models as it properly handles multimodal inputs and chat + formatting. For non-multimodal models, use run_vllm() instead. + """ + from vllm import LLM, SamplingParams + + llm = LLM(**dataclasses.asdict(engine_args)) + + assert all( + llm.llm_engine.model_config.max_model_len + >= (request.prompt_len + request.expected_output_len) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of " + "prompt_len and expected_output_len for all requests." + ) + + prompts = [] + sampling_params: list[SamplingParams] = [] + for request in requests: + prompts.append(request.prompt) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=request.expected_output_len, + detokenize=not disable_detokenize, + ) + ) + start = time.perf_counter() + outputs = llm.chat(prompts, sampling_params, use_tqdm=True) + end = time.perf_counter() + return end - start, outputs + + +async def run_vllm_async( + requests: list[SampleRequest], + n: int, + engine_args: AsyncEngineArgs, + disable_frontend_multiprocessing: bool = False, + disable_detokenize: bool = False, +) -> float: + from vllm import SamplingParams + + async with build_async_engine_client_from_engine_args( + engine_args, disable_frontend_multiprocessing + ) as llm: + model_config = await llm.get_model_config() + assert all( + model_config.max_model_len + >= (request.prompt_len + request.expected_output_len) + for request in requests + ), ( + "Please ensure that max_model_len is greater than the sum of" + " prompt_len and expected_output_len for all requests." + ) + + # Add the requests to the engine. + prompts: list[Union[TextPrompt, TokensPrompt]] = [] + sampling_params: list[SamplingParams] = [] + lora_requests: list[Optional[LoRARequest]] = [] + for request in requests: + prompts.append( + TokensPrompt( + prompt_token_ids=request.prompt["prompt_token_ids"], + multi_modal_data=request.multi_modal_data, + ) + if "prompt_token_ids" in request.prompt + else TextPrompt( + prompt=request.prompt, multi_modal_data=request.multi_modal_data + ) + ) + sampling_params.append( + SamplingParams( + n=n, + temperature=1.0, + top_p=1.0, + ignore_eos=True, + max_tokens=request.expected_output_len, + detokenize=not disable_detokenize, + ) + ) + lora_requests.append(request.lora_request) + + generators = [] + start = time.perf_counter() + for i, (prompt, sp, lr) in enumerate( + zip(prompts, sampling_params, lora_requests) + ): + generator = llm.generate(prompt, sp, lora_request=lr, request_id=f"test{i}") + generators.append(generator) + all_gens = merge_async_iterators(*generators) + async for i, res in all_gens: + pass + end = time.perf_counter() + return end - start + + +def run_hf( + requests: list[SampleRequest], + model: str, + tokenizer: PreTrainedTokenizerBase, + n: int, + max_batch_size: int, + trust_remote_code: bool, + disable_detokenize: bool = False, +) -> float: + llm = AutoModelForCausalLM.from_pretrained( + model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code + ) + if llm.config.model_type == "llama": + # To enable padding in the HF backend. + tokenizer.pad_token = tokenizer.eos_token + llm = llm.cuda() + + pbar = tqdm(total=len(requests)) + start = time.perf_counter() + batch: list[str] = [] + max_prompt_len = 0 + max_output_len = 0 + for i in range(len(requests)): + prompt = requests[i].prompt + prompt_len = requests[i].prompt_len + output_len = requests[i].expected_output_len + # Add the prompt to the batch. + batch.append(prompt) + max_prompt_len = max(max_prompt_len, prompt_len) + max_output_len = max(max_output_len, output_len) + if len(batch) < max_batch_size and i != len(requests) - 1: + # Check if we can add more requests to the batch. + next_prompt_len = requests[i + 1].prompt_len + next_output_len = requests[i + 1].expected_output_len + if ( + max(max_prompt_len, next_prompt_len) + + max(max_output_len, next_output_len) + ) <= 2048: + # We can add more requests to the batch. + continue + + # Generate the sequences. + input_ids = tokenizer(batch, return_tensors="pt", padding=True).input_ids + llm_outputs = llm.generate( + input_ids=input_ids.cuda(), + do_sample=True, + num_return_sequences=n, + temperature=1.0, + top_p=1.0, + use_cache=True, + max_new_tokens=max_output_len, + ) + if not disable_detokenize: + # Include the decoding time. + tokenizer.batch_decode(llm_outputs, skip_special_tokens=True) + pbar.update(len(batch)) + + # Clear the batch. + batch = [] + max_prompt_len = 0 + max_output_len = 0 + end = time.perf_counter() + return end - start + + +def run_mii( + requests: list[SampleRequest], + model: str, + tensor_parallel_size: int, + output_len: int, +) -> float: + from mii import client, serve + + llm = serve(model, tensor_parallel=tensor_parallel_size) + prompts = [request.prompt for request in requests] + + start = time.perf_counter() + llm.generate(prompts, max_new_tokens=output_len) + end = time.perf_counter() + client = client(model) + client.terminate_server() + return end - start + + +def save_to_pytorch_benchmark_format( + args: argparse.Namespace, results: dict[str, Any] +) -> None: + pt_records = convert_to_pytorch_benchmark_format( + args=args, + metrics={ + "requests_per_second": [results["requests_per_second"]], + "tokens_per_second": [results["tokens_per_second"]], + }, + extra_info={ + k: results[k] for k in ["elapsed_time", "num_requests", "total_num_tokens"] + }, + ) + if pt_records: + # Don't use json suffix here as we don't want CI to pick it up + pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json" + write_to_json(pt_file, pt_records) + + +def get_requests(args, tokenizer): + # Common parameters for all dataset types. + common_kwargs = { + "dataset_path": args.dataset_path, + "random_seed": args.seed, + } + sample_kwargs = { + "tokenizer": tokenizer, + "lora_path": args.lora_path, + "max_loras": args.max_loras, + "num_requests": args.num_prompts, + "input_len": args.input_len, + "output_len": args.output_len, + } + + if args.dataset_path is None or args.dataset_name == "random": + sample_kwargs["range_ratio"] = args.random_range_ratio + sample_kwargs["prefix_len"] = args.prefix_len + dataset_cls = RandomDataset + elif args.dataset_name == "sharegpt": + dataset_cls = ShareGPTDataset + if args.backend == "vllm-chat": + sample_kwargs["enable_multimodal_chat"] = True + elif args.dataset_name == "sonnet": + assert tokenizer.chat_template or tokenizer.default_chat_template, ( + "Tokenizer/model must have chat template for sonnet dataset." + ) + dataset_cls = SonnetDataset + sample_kwargs["prefix_len"] = args.prefix_len + sample_kwargs["return_prompt_formatted"] = True + elif args.dataset_name == "burstgpt": + dataset_cls = BurstGPTDataset + elif args.dataset_name == "hf": + if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS: + dataset_cls = VisionArenaDataset + common_kwargs["dataset_subset"] = None + common_kwargs["dataset_split"] = "train" + sample_kwargs["enable_multimodal_chat"] = True + elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS: + dataset_cls = InstructCoderDataset + common_kwargs["dataset_split"] = "train" + elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS: + dataset_cls = ConversationDataset + common_kwargs["dataset_subset"] = args.hf_subset + common_kwargs["dataset_split"] = args.hf_split + sample_kwargs["enable_multimodal_chat"] = True + elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS: + dataset_cls = AIMODataset + common_kwargs["dataset_subset"] = None + common_kwargs["dataset_split"] = "train" + else: + raise ValueError(f"Unknown dataset name: {args.dataset_name}") + # Remove None values + sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None} + return dataset_cls(**common_kwargs).sample(**sample_kwargs) + + +def main(args: argparse.Namespace): + if args.seed is None: + args.seed = 0 + print(args) + random.seed(args.seed) + # Sample the requests. + tokenizer = AutoTokenizer.from_pretrained( + args.tokenizer, trust_remote_code=args.trust_remote_code + ) + requests = get_requests(args, tokenizer) + is_multi_modal = any(request.multi_modal_data is not None for request in requests) + request_outputs: Optional[list[RequestOutput]] = None + if args.backend == "vllm": + if args.async_engine: + elapsed_time = uvloop.run( + run_vllm_async( + requests, + args.n, + AsyncEngineArgs.from_cli_args(args), + args.disable_frontend_multiprocessing, + args.disable_detokenize, + ) + ) + else: + elapsed_time, request_outputs = run_vllm( + requests, + args.n, + EngineArgs.from_cli_args(args), + args.disable_detokenize, + ) + elif args.backend == "hf": + assert args.tensor_parallel_size == 1 + elapsed_time = run_hf( + requests, + args.model, + tokenizer, + args.n, + args.hf_max_batch_size, + args.trust_remote_code, + args.disable_detokenize, + ) + elif args.backend == "mii": + elapsed_time = run_mii( + requests, args.model, args.tensor_parallel_size, args.output_len + ) + elif args.backend == "vllm-chat": + elapsed_time, request_outputs = run_vllm_chat( + requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize + ) + else: + raise ValueError(f"Unknown backend: {args.backend}") + + if request_outputs: + # Note: with the vllm and vllm-chat backends, + # we have request_outputs, which we use to count tokens. + total_prompt_tokens = 0 + total_output_tokens = 0 + for ro in request_outputs: + if not isinstance(ro, RequestOutput): + continue + total_prompt_tokens += ( + len(ro.prompt_token_ids) if ro.prompt_token_ids else 0 + ) + total_output_tokens += sum(len(o.token_ids) for o in ro.outputs if o) + total_num_tokens = total_prompt_tokens + total_output_tokens + else: + total_num_tokens = sum(r.prompt_len + r.expected_output_len for r in requests) + total_output_tokens = sum(r.expected_output_len for r in requests) + total_prompt_tokens = total_num_tokens - total_output_tokens + + if is_multi_modal and args.backend != "vllm-chat": + print( + "\033[91mWARNING\033[0m: Multi-modal request with " + f"{args.backend} backend detected. The " + "following metrics are not accurate because image tokens are not" + " counted. See vllm-project/vllm/issues/9778 for details." + ) + # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length. + # vllm-chat backend counts the image tokens now + + print( + f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, " + f"{total_num_tokens / elapsed_time:.2f} total tokens/s, " + f"{total_output_tokens / elapsed_time:.2f} output tokens/s" + ) + print(f"Total num prompt tokens: {total_prompt_tokens}") + print(f"Total num output tokens: {total_output_tokens}") + + # Output JSON results if specified + if args.output_json: + results = { + "elapsed_time": elapsed_time, + "num_requests": len(requests), + "total_num_tokens": total_num_tokens, + "requests_per_second": len(requests) / elapsed_time, + "tokens_per_second": total_num_tokens / elapsed_time, + } + with open(args.output_json, "w") as f: + json.dump(results, f, indent=4) + save_to_pytorch_benchmark_format(args, results) + + +def validate_args(args): + """ + Validate command-line arguments. + """ + + # === Deprecation and Defaulting === + if args.dataset is not None: + warnings.warn( + "The '--dataset' argument will be deprecated in the next release. " + "Please use '--dataset-name' and '--dataset-path' instead.", + stacklevel=2, + ) + args.dataset_path = args.dataset + + if not getattr(args, "tokenizer", None): + args.tokenizer = args.model + + # === Backend Validation === + valid_backends = {"vllm", "hf", "mii", "vllm-chat"} + if args.backend not in valid_backends: + raise ValueError(f"Unsupported backend: {args.backend}") + + # === Dataset Configuration === + if not args.dataset and not args.dataset_path: + print("When dataset path is not set, it will default to random dataset") + args.dataset_name = "random" + if args.input_len is None: + raise ValueError("input_len must be provided for a random dataset") + + # === Dataset Name Specific Checks === + # --hf-subset and --hf-split: only used + # when dataset_name is 'hf' + if args.dataset_name != "hf" and ( + getattr(args, "hf_subset", None) is not None + or getattr(args, "hf_split", None) is not None + ): + warnings.warn( + "--hf-subset and --hf-split will be ignored \ + since --dataset-name is not 'hf'.", + stacklevel=2, + ) + elif args.dataset_name == "hf": + if args.dataset_path in ( + VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys() + | ConversationDataset.SUPPORTED_DATASET_PATHS + ): + assert args.backend == "vllm-chat", ( + f"{args.dataset_path} needs to use vllm-chat as the backend." + ) # noqa: E501 + elif args.dataset_path in ( + InstructCoderDataset.SUPPORTED_DATASET_PATHS + | AIMODataset.SUPPORTED_DATASET_PATHS + ): + assert args.backend == "vllm", ( + f"{args.dataset_path} needs to use vllm as the backend." + ) # noqa: E501 + else: + raise ValueError(f"{args.dataset_path} is not supported by hf dataset.") + + # --random-range-ratio: only used when dataset_name is 'random' + if args.dataset_name != "random" and args.random_range_ratio is not None: + warnings.warn( + "--random-range-ratio will be ignored since \ + --dataset-name is not 'random'.", + stacklevel=2, + ) + + # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not + # set. + if ( + args.dataset_name not in {"random", "sonnet", None} + and args.prefix_len is not None + ): + warnings.warn( + "--prefix-len will be ignored since --dataset-name\ + is not 'random', 'sonnet', or not set.", + stacklevel=2, + ) + + # === LoRA Settings === + if getattr(args, "enable_lora", False) and args.backend != "vllm": + raise ValueError("LoRA benchmarking is only supported for vLLM backend") + if getattr(args, "enable_lora", False) and args.lora_path is None: + raise ValueError("LoRA path must be provided when enable_lora is True") + + # === Backend-specific Validations === + if args.backend == "hf" and args.hf_max_batch_size is None: + raise ValueError("HF max batch size is required for HF backend") + if args.backend != "hf" and args.hf_max_batch_size is not None: + raise ValueError("HF max batch size is only for HF backend.") + + if ( + args.backend in {"hf", "mii"} + and getattr(args, "quantization", None) is not None + ): + raise ValueError("Quantization is only for vLLM backend.") + + if args.backend == "mii" and args.dtype != "auto": + raise ValueError("dtype must be auto for MII backend.") + if args.backend == "mii" and args.n != 1: + raise ValueError("n must be 1 for MII backend.") + if args.backend == "mii" and args.tokenizer != args.model: + raise ValueError("Tokenizer must be the same as the model for MII backend.") + + # --data-parallel is not supported currently. + # https://github.com/vllm-project/vllm/issues/16222 + if args.data_parallel_size > 1: + raise ValueError( + "Data parallel is not supported in offline benchmark, \ + please use benchmark serving instead" + ) + + +def create_argument_parser(): + parser = FlexibleArgumentParser(description="Benchmark the throughput.") + parser.add_argument( + "--backend", + type=str, + choices=["vllm", "hf", "mii", "vllm-chat"], + default="vllm", + ) + parser.add_argument( + "--dataset-name", + type=str, + choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"], + help="Name of the dataset to benchmark on.", + default="sharegpt", + ) + parser.add_argument( + "--dataset", + type=str, + default=None, + help="Path to the ShareGPT dataset, will be deprecated in\ + the next release. The dataset is expected to " + "be a json in form of list[dict[..., conversations: " + "list[dict[..., value: ]]]]", + ) + parser.add_argument( + "--dataset-path", type=str, default=None, help="Path to the dataset" + ) + parser.add_argument( + "--input-len", + type=int, + default=None, + help="Input prompt length for each request", + ) + parser.add_argument( + "--output-len", + type=int, + default=None, + help="Output length for each request. Overrides the " + "output length from the dataset.", + ) + parser.add_argument( + "--n", type=int, default=1, help="Number of generated sequences per prompt." + ) + parser.add_argument( + "--num-prompts", type=int, default=1000, help="Number of prompts to process." + ) + parser.add_argument( + "--hf-max-batch-size", + type=int, + default=None, + help="Maximum batch size for HF backend.", + ) + parser.add_argument( + "--output-json", + type=str, + default=None, + help="Path to save the throughput results in JSON format.", + ) + parser.add_argument( + "--async-engine", + action="store_true", + default=False, + help="Use vLLM async engine rather than LLM class.", + ) + parser.add_argument( + "--disable-frontend-multiprocessing", + action="store_true", + default=False, + help="Disable decoupled async engine frontend.", + ) + parser.add_argument( + "--disable-detokenize", + action="store_true", + help=( + "Do not detokenize the response (i.e. do not include " + "detokenization time in the measurement)" + ), + ) + # LoRA + parser.add_argument( + "--lora-path", + type=str, + default=None, + help="Path to the LoRA adapters to use. This can be an absolute path, " + "a relative path, or a Hugging Face model identifier.", + ) + parser.add_argument( + "--prefix-len", + type=int, + default=None, + help=f"Number of prefix tokens to be used in RandomDataset " + "and SonnetDataset. For RandomDataset, the total input " + "length is the sum of prefix-len (default: " + f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length " + "sampled from [input_len * (1 - range_ratio), " + "input_len * (1 + range_ratio)]. For SonnetDataset, " + f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) " + "controls how much of the input is fixed lines versus " + "random lines, but the total input length remains approximately " + "input_len tokens.", + ) + # random dataset + parser.add_argument( + "--random-range-ratio", + type=float, + default=None, + help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) " + "for sampling input/output length, " + "used only for RandomDataset. Must be in the range [0, 1) to " + "define a symmetric sampling range " + "[length * (1 - range_ratio), length * (1 + range_ratio)].", + ) + + # hf dtaset + parser.add_argument( + "--hf-subset", type=str, default=None, help="Subset of the HF dataset." + ) + parser.add_argument( + "--hf-split", type=str, default=None, help="Split of the HF dataset." + ) + + parser = AsyncEngineArgs.add_cli_args(parser) + + return parser + + +if __name__ == "__main__": + parser = create_argument_parser() + args = parser.parse_args() + if args.tokenizer is None: + args.tokenizer = args.model + validate_args(args) + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/benchmark_utils.py b/online_apiserver_test_maxbs/benchmarks/benchmark_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..283f938df50af101fa2ff82f3c46e123b25e5524 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/benchmark_utils.py @@ -0,0 +1,74 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import json +import math +import os +from typing import Any + + +def convert_to_pytorch_benchmark_format( + args: argparse.Namespace, metrics: dict[str, list], extra_info: dict[str, Any] +) -> list: + """ + Save the benchmark results in the format used by PyTorch OSS benchmark with + on metric per record + https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database + """ + records = [] + if not os.environ.get("SAVE_TO_PYTORCH_BENCHMARK_FORMAT", False): + return records + + for name, benchmark_values in metrics.items(): + record = { + "benchmark": { + "name": "vLLM benchmark", + "extra_info": { + "args": vars(args), + }, + }, + "model": { + "name": args.model, + }, + "metric": { + "name": name, + "benchmark_values": benchmark_values, + "extra_info": extra_info, + }, + } + + tp = record["benchmark"]["extra_info"]["args"].get("tensor_parallel_size") + # Save tensor_parallel_size parameter if it's part of the metadata + if not tp and "tensor_parallel_size" in extra_info: + record["benchmark"]["extra_info"]["args"]["tensor_parallel_size"] = ( + extra_info["tensor_parallel_size"] + ) + + records.append(record) + + return records + + +class InfEncoder(json.JSONEncoder): + def clear_inf(self, o: Any): + if isinstance(o, dict): + return {k: self.clear_inf(v) for k, v in o.items()} + elif isinstance(o, list): + return [self.clear_inf(v) for v in o] + elif isinstance(o, float) and math.isinf(o): + return "inf" + return o + + def iterencode(self, o: Any, *args, **kwargs) -> Any: + return super().iterencode(self.clear_inf(o), *args, **kwargs) + + +def write_to_json(filename: str, records: list) -> None: + with open(filename, "w") as f: + json.dump( + records, + f, + cls=InfEncoder, + default=lambda o: f"<{type(o).__name__} object is not JSON serializable>", + ) diff --git a/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/sparse_benchmarks.py b/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/sparse_benchmarks.py new file mode 100644 index 0000000000000000000000000000000000000000..9ec270bbd2e988b459f99d876215984bf125d0f9 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/sparse_benchmarks.py @@ -0,0 +1,516 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import itertools +import pickle as pkl +import time +from collections.abc import Iterable +from typing import Callable + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import make_rand_sparse_tensors +from weight_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512] +DEFAULT_TP_SIZES = [1] + + +# bench +def bench_fn( + label: str, sub_label: str, description: str, fn: Callable, *args, **kwargs +) -> TMeasurement: + min_run_time = 1 + + globals = { + "args": args, + "kwargs": kwargs, + "fn": fn, + } + return TBenchmark.Timer( + stmt="fn(*args, **kwargs)", + globals=globals, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + +def bench_int8( + dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str +) -> Iterable[TMeasurement]: + assert dtype == torch.int8 + b_compressed, e, a, b = make_rand_sparse_tensors(torch.int8, m, n, k) + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + + out = ops.cutlass_scaled_sparse_mm( + a, b_compressed, e, scale_a, scale_b, torch.bfloat16 + ) + out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16) + + if not torch.allclose(out, out_ref): + print("Incorrect results") + print(out) + print(out_ref) + else: + print("Correct results") + + timers = [] + # pytorch impl - bfloat16 + timers.append( + bench_fn( + label, + sub_label, + "pytorch_bf16_bf16_bf16_matmul-no-scales", + torch.mm, + a.to(dtype=torch.bfloat16), + b.to(dtype=torch.bfloat16), + ) + ) + + # pytorch impl - float16 + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp16_fp16_fp16_matmul-no-scales", + torch.mm, + a.to(dtype=torch.float16), + b.to(dtype=torch.float16), + ) + ) + + # cutlass impl + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_mm", + ops.cutlass_scaled_mm, + a, + b, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_mm_bias", + ops.cutlass_scaled_mm, + a, + b, + scale_a, + scale_b, + torch.bfloat16, + bias, + ) + ) + + # cutlass sparse impl + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_sparse_mm", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass sparse with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_i8_i8_bf16_scaled_sparse_mm_bias", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + bias, + ) + ) + + return timers + + +def bench_fp8( + dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str +) -> Iterable[TMeasurement]: + assert dtype == torch.float8_e4m3fn + b_compressed, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n, k) + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + + out = ops.cutlass_scaled_sparse_mm( + a, b_compressed, e, scale_a, scale_b, torch.bfloat16 + ) + out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16) + + if not torch.allclose(out, out_ref): + print("Incorrect results") + print(out) + print(out_ref) + else: + print("Correct results") + + timers = [] + + # pytorch impl w. bf16 + timers.append( + bench_fn( + label, + sub_label, + "pytorch_bf16_bf16_bf16_matmul-no-scales", + torch.mm, + a.to(dtype=torch.bfloat16, device="cuda"), + b.to(dtype=torch.bfloat16, device="cuda"), + ) + ) + + # pytorch impl: bf16 output, without fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_bf16_scaled_mm", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.bfloat16, + ) + ) + + # pytorch impl: bf16 output, with fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.bfloat16, + use_fast_accum=True, + ) + ) + + # pytorch impl: fp16 output, without fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_fp16_scaled_mm", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.float16, + ) + ) + + # pytorch impl: fp16 output, with fp8 fast accum + timers.append( + bench_fn( + label, + sub_label, + "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum", + torch._scaled_mm, + a, + b, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=torch.float16, + use_fast_accum=True, + ) + ) + + # cutlass impl: bf16 output + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_bf16_scaled_mm", + ops.cutlass_scaled_mm, + a, + b, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass impl: bf16 output + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_bf16_scaled_sparse_mm", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + ) + ) + + # cutlass impl: fp16 output + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_fp16_scaled_sparse_mm", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.float16, + ) + ) + + # cutlass impl: bf16 output, with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_bf16_scaled_sparse_mm_bias", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.bfloat16, + bias, + ) + ) + + # cutlass impl: fp16 output, with bias + timers.append( + bench_fn( + label, + sub_label, + "cutlass_fp8_fp8_fp16_scaled_sparse_mm_bias", + ops.cutlass_scaled_sparse_mm, + a, + b_compressed, + e, + scale_a, + scale_b, + torch.float16, + bias.to(dtype=torch.float16), + ) + ) + + return timers + + +def bench( + dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str +) -> Iterable[TMeasurement]: + if dtype == torch.int8: + return bench_int8(dtype, m, k, n, label, sub_label) + if dtype == torch.float8_e4m3fn: + return bench_fp8(dtype, m, k, n, label, sub_label) + raise ValueError("unsupported type") + + +# runner +def print_timers(timers: Iterable[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def run( + dtype: torch.dtype, MKNs: Iterable[tuple[int, int, int]] +) -> Iterable[TMeasurement]: + results = [] + for m, k, n in MKNs: + timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm", f"MKN=({m}x{k}x{n})") + print_timers(timers) + results.extend(timers) + + return results + + +# output makers +def make_output( + data: Iterable[TMeasurement], + MKNs: Iterable[tuple[int, int, int]], + base_description: str, + timestamp=None, +): + print(f"== All Results {base_description} ====") + print_timers(data) + + # pickle all the results + timestamp = int(time.time()) if timestamp is None else timestamp + with open(f"{base_description}-{timestamp}.pkl", "wb") as f: + pkl.dump(data, f) + + +# argparse runners + + +def run_square_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment)) + MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes)) + data = run(args.dtype, MKNs) + + make_output(data, MKNs, f"square_bench-{args.dtype}") + + +def run_range_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment)) + n = len(dim_sizes) + Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes + Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes + Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes + MKNs = list(zip(Ms, Ks, Ns)) + data = run(args.dtype, MKNs) + + make_output(data, MKNs, f"range_bench-{args.dtype}") + + +def run_model_bench(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]: + KNs = [] + for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]): + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + KNs.append(KN) + return KNs + + model_bench_data = [] + models_tps = list(itertools.product(args.models, args.tp_sizes)) + for model, tp_size in models_tps: + Ms = args.batch_sizes + KNs = model_shapes(model, tp_size) + MKNs = [] + for m in Ms: + for k, n in KNs: + MKNs.append((m, k, n)) + + data = run(args.dtype, MKNs) + model_bench_data.append(data) + + # Print all results + for data, model_tp in zip(model_bench_data, models_tps): + model, tp_size = model_tp + print(f"== Results {args.dtype} {model}-TP{tp_size} ====") + print_timers(data) + + timestamp = int(time.time()) + + all_data = [] + for d in model_bench_data: + all_data.extend(d) + # pickle all data + with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f: + pkl.dump(all_data, f) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "int8": + return torch.int8 + if dt == "fp8": + return torch.float8_e4m3fn + raise ValueError("unsupported dtype") + + parser = FlexibleArgumentParser( + description=""" +Benchmark Cutlass GEMM. + + To run square GEMMs: + python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64 + + To run constant N and K and sweep M: + python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384 + + To run dimensions from a model: + python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1 + + Output: + - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs. + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + + parser.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['int8', 'fp8']", + ) + subparsers = parser.add_subparsers(dest="cmd") + + square_parser = subparsers.add_parser("square_bench") + square_parser.add_argument("--dim-start", type=int, required=True) + square_parser.add_argument("--dim-end", type=int, required=True) + square_parser.add_argument("--dim-increment", type=int, required=True) + square_parser.set_defaults(func=run_square_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--dim-start", type=int, required=True) + range_parser.add_argument("--dim-end", type=int, required=True) + range_parser.add_argument("--dim-increment", type=int, required=True) + range_parser.add_argument("--m-constant", type=int, default=None) + range_parser.add_argument("--n-constant", type=int, default=None) + range_parser.add_argument("--k-constant", type=int, default=None) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/utils.py b/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b4f3c6bf94eda0e1bf1d253def3b17d273415dcc --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/utils.py @@ -0,0 +1,100 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +# Cutlass bench utils +from collections.abc import Iterable + +import torch + +import vllm._custom_ops as ops + + +def to_fp8(tensor: torch.Tensor) -> torch.Tensor: + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to( + dtype=torch.float8_e4m3fn + ) + + +def to_int8(tensor: torch.Tensor) -> torch.Tensor: + return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8) + + +def to_bf16(tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(dtype=torch.bfloat16) + + +def to_fp16(tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(dtype=torch.float16) + + +def make_rand_tensors( + dtype: torch.dtype, m: int, n: int, k: int +) -> tuple[torch.Tensor, torch.Tensor]: + a = torch.randn((m, k), device="cuda") * 5 + b = torch.randn((n, k), device="cuda").t() * 5 + + if dtype == torch.int8: + return to_int8(a), to_int8(b) + if dtype == torch.float8_e4m3fn: + return to_fp8(a), to_fp8(b) + + raise ValueError("unsupported dtype") + + +def prune_to_2_4(tensor): + # Reshape tensor to [N, 4] where N is number of groups of 4 + original_shape = tensor.shape + reshaped = tensor.reshape(-1, 4) + + # Get indices of top 2 absolute values in each group of 4 + _, indices = torch.topk(torch.abs(reshaped), k=2, dim=1) + + # Create binary mask + mask = torch.zeros_like(reshaped) + mask.scatter_(dim=1, index=indices, src=torch.ones_like(indices, dtype=mask.dtype)) + + # Apply mask and reshape back + pruned = reshaped * mask + + # Turn all -0.0 to 0.0 + pruned[pruned == -0.0] = 0.0 + + return pruned.reshape(original_shape) + + +def make_rand_sparse_tensors( + dtype: torch.dtype, m: int, n: int, k: int +) -> tuple[torch.Tensor, torch.Tensor]: + a = torch.randn((m, k), device="cuda") * 5 + b = torch.randn((n, k), device="cuda").t() * 5 + + b = prune_to_2_4(b.t()).t() + + if dtype == torch.int8: + a, b = to_int8(a), to_int8(b) + elif dtype == torch.float8_e4m3fn: + a, b = to_fp8(a), to_fp8(b) + elif dtype == torch.float16: + a, b = to_fp16(a), to_fp16(b) + elif dtype == torch.bfloat16: + a, b = to_bf16(a), to_bf16(b) + else: + raise ValueError("unsupported dtype") + + b_compressed, e = ops.cutlass_sparse_compress(b.t()) + + # Compressed B, Metadata, Original A, B + return b_compressed, e, a, b + + +def make_n_rand_sparse_tensors( + num_tensors: int, dtype: torch.dtype, m: int, n: int, k: int +) -> tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]: + ABs = [] + for _ in range(num_tensors): + b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k) + if b_comp is not None: + ABs.append(make_rand_sparse_tensors(dtype, m, n, k)) + BComps, Es, As, Bs = zip(*ABs) + return list(BComps), list(Es), list(As), list(Bs) diff --git a/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py b/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py new file mode 100644 index 0000000000000000000000000000000000000000..a5a5b52f603977cc0c967e455724bc7c6a2bf585 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py @@ -0,0 +1,372 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import itertools +import pickle as pkl +import time +from collections.abc import Iterable +from typing import Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import make_rand_tensors +from weight_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + w8a8_block_fp8_matmul, +) +from vllm.utils import FlexibleArgumentParser, cdiv + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512] +DEFAULT_TP_SIZES = [1] + + +# bench +def bench_fn( + label: str, sub_label: str, description: str, fn: Callable, *args, **kwargs +) -> TMeasurement: + min_run_time = 1 + + globals = { + "args": args, + "kwargs": kwargs, + "fn": fn, + } + return TBenchmark.Timer( + stmt="fn(*args, **kwargs)", + globals=globals, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + +def bench_int8( + dtype: torch.dtype, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + """Benchmark INT8-based kernels.""" + assert dtype == torch.int8 + a, b = make_rand_tensors(torch.int8, m, n, k) + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + azp = torch.zeros((m,), device="cuda", dtype=torch.int32) + azp_adj = torch.zeros((n,), device="cuda", dtype=torch.int32) + + bench_fns = { + "pytorch_bf16_bf16_bf16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16) + ), + "pytorch_fp16_fp16_fp16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.float16), b.to(dtype=torch.float16) + ), + "cutlass_i8_i8_bf16_scaled_mm": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16 + ), + "cutlass_i8_i8_bf16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16, bias + ), + "cutlass_i8_i8_bf16_scaled_mm_azp": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj + ), + "cutlass_i8_i8_bf16_scaled_mm_azp_bias": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj, None, bias + ), + "cutlass_i8_i8_bf16_scaled_mm_azp_pt": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj, azp + ), + "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias": lambda: ops.cutlass_scaled_mm_azp( + a, b, scale_a, scale_b, torch.bfloat16, azp_adj, azp, bias + ), + } + + timers = [] + for name, fn in bench_fns.items(): + # If bench_kernels is None, run all. Otherwise, run only exact matches. + if bench_kernels is None or name in bench_kernels: + print(f"Running {name}") + timers.append(bench_fn(label, sub_label, name, fn)) + + return timers + + +def bench_fp8( + dtype: torch.dtype, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + """Benchmark FP8-based kernels.""" + assert dtype == torch.float8_e4m3fn + a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k) + a_cont = a.contiguous() + scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32) + scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32) + + block_scale_a = torch.rand((m, cdiv(k, 128)), device="cuda", dtype=torch.float32) + block_scale_b = torch.rand( + cdiv(k, 128), cdiv(n, 128), device="cuda", dtype=torch.float32 + ) + block_scale_a_M_major = block_scale_a.t().contiguous().t() + block_scale_b_K_major = block_scale_b.t().contiguous().t() + bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16) + + print(m, k, n) + + bench_fns = { + "pytorch_bf16_bf16_bf16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16) + ), + "pytorch_fp16_fp16_fp16_matmul-no-scales": lambda: torch.mm( + a.to(dtype=torch.float16), b.to(dtype=torch.float16) + ), + "pytorch_fp8_fp8_fp16_scaled_mm": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.float16 + ), + "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.float16, use_fast_accum=True + ), + "pytorch_fp8_fp8_bf16_scaled_mm": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.bfloat16 + ), + "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum": lambda: torch._scaled_mm( + a, b, scale_a, scale_b, out_dtype=torch.bfloat16, use_fast_accum=True + ), + "cutlass_fp8_fp8_bf16_scaled_mm": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16 + ), + "cutlass_fp8_fp8_fp16_scaled_mm": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.float16 + ), + "cutlass_fp8_fp8_bf16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.bfloat16, bias + ), + "cutlass_fp8_fp8_fp16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm( + a, b, scale_a, scale_b, torch.float16, bias.to(dtype=torch.float16) + ), + "triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_block_fp8_matmul( + a_cont, b.t(), block_scale_a, block_scale_b.t(), (128, 128) + ), + "cutlass_fp8_fp8_fp16_scaled_mm_blockwise": lambda: ops.cutlass_scaled_mm( + a, b, block_scale_a_M_major, block_scale_b_K_major, torch.float16 + ), + } + + timers = [] + for name, fn in bench_fns.items(): + # If bench_kernels is None, run all. Otherwise, run only exact matches. + if bench_kernels is None or name in bench_kernels: + print(f"Running {name}") + timers.append(bench_fn(label, sub_label, name, fn)) + + return timers + + +def bench( + dtype: torch.dtype, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + if dtype == torch.int8: + return bench_int8(dtype, m, k, n, label, sub_label, bench_kernels) + if dtype == torch.float8_e4m3fn: + return bench_fp8(dtype, m, k, n, label, sub_label, bench_kernels) + raise ValueError("unsupported type") + + +# runner +def print_timers(timers: Iterable[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def run( + dtype: torch.dtype, + MKNs: Iterable[tuple[int, int, int]], + bench_kernels: Optional[list[str]] = None, +) -> Iterable[TMeasurement]: + results = [] + for m, k, n in MKNs: + timers = bench( + dtype, + m, + k, + n, + f"scaled-{dtype}-gemm", + f"MKN=({m}x{k}x{n})", + bench_kernels=bench_kernels, + ) + print_timers(timers) + results.extend(timers) + return results + + +def make_output( + data: Iterable[TMeasurement], + MKNs: Iterable[tuple[int, int, int]], + base_description: str, + timestamp=None, +): + print(f"== All Results {base_description} ====") + print_timers(data) + + # pickle all the results + timestamp = int(time.time()) if timestamp is None else timestamp + with open(f"{base_description}-{timestamp}.pkl", "wb") as f: + pkl.dump(data, f) + + +def run_square_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment)) + MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes)) + data = run(args.dtype, MKNs, bench_kernels=args.kernels) + make_output(data, MKNs, f"square_bench-{args.dtype}") + + +def run_range_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment)) + n = len(dim_sizes) + Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes + Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes + Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes + MKNs = list(zip(Ms, Ks, Ns)) + data = run(args.dtype, MKNs, bench_kernels=args.kernels) + make_output(data, MKNs, f"range_bench-{args.dtype}") + + +def run_model_bench(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]: + KNs = [] + for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]): + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + KNs.append(KN) + return KNs + + model_bench_data = [] + models_tps = list(itertools.product(args.models, args.tp_sizes)) + for model, tp_size in models_tps: + Ms = args.batch_sizes + KNs = model_shapes(model, tp_size) + MKNs = [] + for m in Ms: + for k, n in KNs: + MKNs.append((m, k, n)) + + data = run(args.dtype, MKNs, bench_kernels=args.kernels) + model_bench_data.append(data) + + # Print all results + for data, model_tp in zip(model_bench_data, models_tps): + model, tp_size = model_tp + print(f"== Results {args.dtype} {model}-TP{tp_size} ====") + print_timers(data) + + timestamp = int(time.time()) + + all_data = [] + for d in model_bench_data: + all_data.extend(d) + # pickle all data + with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f: + pkl.dump(all_data, f) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "int8": + return torch.int8 + if dt == "fp8": + return torch.float8_e4m3fn + raise ValueError("unsupported dtype") + + parser = FlexibleArgumentParser( + description=""" +Benchmark Cutlass GEMM. + + To run square GEMMs: + python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64 + + To run constant N and K and sweep M: + python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384 + + To run dimensions from a model: + python3 ./benchmarks/cutlass_benchmarks/w8a8_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1 + + Output: + - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs. + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + + parser.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['int8', 'fp8']", + ) + parser.add_argument( + "--kernels", + nargs="+", + type=str, + default=None, + help="Exact names of the kernels to benchmark. If not set, runs all kernels.", + ) + + subparsers = parser.add_subparsers(dest="cmd") + + square_parser = subparsers.add_parser("square_bench") + square_parser.add_argument("--dim-start", type=int, required=True) + square_parser.add_argument("--dim-end", type=int, required=True) + square_parser.add_argument("--dim-increment", type=int, required=True) + square_parser.set_defaults(func=run_square_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--dim-start", type=int, required=True) + range_parser.add_argument("--dim-end", type=int, required=True) + range_parser.add_argument("--dim-increment", type=int, required=True) + range_parser.add_argument("--m-constant", type=int, default=None) + range_parser.add_argument("--n-constant", type=int, default=None) + range_parser.add_argument("--k-constant", type=int, default=None) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/weight_shapes.py b/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/weight_shapes.py new file mode 100644 index 0000000000000000000000000000000000000000..25b96ef56620ea7dbd97846cdd57ab0e97a6dfd1 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/cutlass_benchmarks/weight_shapes.py @@ -0,0 +1,46 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +# Weight Shapes are in the format +# ([K, N], TP_SPLIT_DIM) +# Example: +# A shape of ([14336, 4096], 0) indicates the following GEMM shape, +# - TP1 : K = 14336, N = 4096 +# - TP2 : K = 7168, N = 4096 +# A shape of ([4096, 6144], 1) indicates the following GEMM shape, +# - TP1 : K = 4096, N = 6144 +# - TP4 : K = 4096, N = 1536 + +# TP1 shapes +WEIGHT_SHAPES = { + "mistralai/Mistral-7B-v0.1": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-7b-hf": [ + ([4096, 12288], 1), + ([4096, 4096], 0), + ([4096, 22016], 1), + ([11008, 4096], 0), + ], + "meta-llama/Llama-3-8b": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-13b-hf": [ + ([5120, 15360], 1), + ([5120, 5120], 0), + ([5120, 27648], 1), + ([13824, 5120], 0), + ], + "meta-llama/Llama-2-70b-hf": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 57344], 1), + ([28672, 8192], 0), + ], +} diff --git a/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..94999630bae122fa86d68c77cf9fc3941fbe4b14 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh @@ -0,0 +1,145 @@ +#!/bin/bash + +# benchmark the overhead of disaggregated prefill. +# methodology: +# - send all request to prefill vLLM instance. It will buffer KV cache. +# - then send all request to decode instance. +# - The TTFT of decode instance is the overhead. + +set -ex + +kill_gpu_processes() { + # kill all processes on GPU. + pgrep pt_main_thread | xargs -r kill -9 + pgrep python3 | xargs -r kill -9 + sleep 10 + + # remove vllm config file + rm -rf ~/.config/vllm + + # Print the GPU memory usage + # so that we know if all GPU processes are killed. + gpu_memory_usage=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0) + # The memory usage should be 0 MB. + echo "GPU 0 Memory Usage: $gpu_memory_usage MB" +} + +wait_for_server() { + # wait for vllm server to start + # return 1 if vllm server crashes + local port=$1 + timeout 1200 bash -c " + until curl -s localhost:${port}/v1/completions > /dev/null; do + sleep 1 + done" && return 0 || return 1 +} + + +benchmark() { + + export VLLM_LOGGING_LEVEL=DEBUG + export VLLM_HOST_IP=$(hostname -I | awk '{print $1}') + + # compare chunked prefill with disaggregated prefill + + results_folder="./results" + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + dataset_name="sonnet" + dataset_path="../sonnet_4x.txt" + num_prompts=10 + qps=$1 + prefix_len=50 + input_len=2048 + output_len=$2 + + + CUDA_VISIBLE_DEVICES=0 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8100 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + + CUDA_VISIBLE_DEVICES=1 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8200 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + wait_for_server 8100 + wait_for_server 8200 + + # let the prefill instance finish prefill + python3 ../benchmark_serving.py \ + --backend vllm \ + --model $model \ + --dataset-name $dataset_name \ + --dataset-path $dataset_path \ + --sonnet-input-len $input_len \ + --sonnet-output-len "$output_len" \ + --sonnet-prefix-len $prefix_len \ + --num-prompts $num_prompts \ + --port 8100 \ + --save-result \ + --result-dir $results_folder \ + --result-filename disagg_prefill_tp1.json \ + --request-rate "inf" + + + # send the request to decode. + # The TTFT of this command will be the overhead of disagg prefill impl. + python3 ../benchmark_serving.py \ + --backend vllm \ + --model $model \ + --dataset-name $dataset_name \ + --dataset-path $dataset_path \ + --sonnet-input-len $input_len \ + --sonnet-output-len "$output_len" \ + --sonnet-prefix-len $prefix_len \ + --num-prompts $num_prompts \ + --port 8200 \ + --save-result \ + --result-dir $results_folder \ + --result-filename disagg_prefill_tp1_overhead.json \ + --request-rate "$qps" + kill_gpu_processes + +} + + +main() { + + (which wget && which curl) || (apt-get update && apt-get install -y wget curl) + (which jq) || (apt-get -y install jq) + (which socat) || (apt-get -y install socat) + + pip install quart httpx datasets + + cd "$(dirname "$0")" + + cd .. + # create sonnet-4x.txt + echo "" > sonnet_4x.txt + for _ in {1..4} + do + cat sonnet.txt >> sonnet_4x.txt + done + cd disagg_benchmarks + + rm -rf results + mkdir results + + default_qps=1 + default_output_len=1 + benchmark $default_qps $default_output_len + +} + + +main "$@" diff --git a/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..eb5d891d0d4a58b75f975ae43655ad539bc6a71e --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh @@ -0,0 +1,163 @@ +#!/bin/bash + +# Requirement: 2x GPUs. + + +# Model: meta-llama/Meta-Llama-3.1-8B-Instruct +# Query: 1024 input tokens, 6 output tokens, QPS 2/4/6/8, 100 requests +# Resource: 2x GPU +# Approaches: +# 2. Chunked prefill: 2 vllm instance with tp=4, equivalent to 1 tp=4 instance with QPS 4 +# 3. Disaggregated prefill: 1 prefilling instance and 1 decoding instance +# Prefilling instance: max_output_token=1 +# Decoding instance: force the input tokens be the same across requests to bypass prefilling + +set -ex + +kill_gpu_processes() { + # kill all processes on GPU. + pgrep pt_main_thread | xargs -r kill -9 + pgrep python3 | xargs -r kill -9 + for port in 8000 8100 8200; do lsof -t -i:$port | xargs -r kill -9; done + sleep 1 +} + +wait_for_server() { + # wait for vllm server to start + # return 1 if vllm server crashes + local port=$1 + timeout 1200 bash -c " + until curl -s localhost:${port}/v1/completions > /dev/null; do + sleep 1 + done" && return 0 || return 1 +} + + +launch_chunked_prefill() { + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + # disagg prefill + CUDA_VISIBLE_DEVICES=0 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8100 \ + --max-model-len 10000 \ + --enable-chunked-prefill \ + --gpu-memory-utilization 0.6 & + CUDA_VISIBLE_DEVICES=1 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8200 \ + --max-model-len 10000 \ + --enable-chunked-prefill \ + --gpu-memory-utilization 0.6 & + wait_for_server 8100 + wait_for_server 8200 + python3 round_robin_proxy.py & + sleep 1 +} + + +launch_disagg_prefill() { + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + # disagg prefill + CUDA_VISIBLE_DEVICES=0 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8100 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + CUDA_VISIBLE_DEVICES=1 python3 \ + -m vllm.entrypoints.openai.api_server \ + --model $model \ + --port 8200 \ + --max-model-len 10000 \ + --gpu-memory-utilization 0.6 \ + --kv-transfer-config \ + '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' & + + wait_for_server 8100 + wait_for_server 8200 + python3 disagg_prefill_proxy_server.py & + sleep 1 +} + + +benchmark() { + results_folder="./results" + model="meta-llama/Meta-Llama-3.1-8B-Instruct" + dataset_name="sonnet" + dataset_path="../sonnet_4x.txt" + num_prompts=100 + qps=$1 + prefix_len=50 + input_len=1024 + output_len=$2 + tag=$3 + + python3 ../benchmark_serving.py \ + --backend vllm \ + --model $model \ + --dataset-name $dataset_name \ + --dataset-path $dataset_path \ + --sonnet-input-len $input_len \ + --sonnet-output-len "$output_len" \ + --sonnet-prefix-len $prefix_len \ + --num-prompts $num_prompts \ + --port 8000 \ + --save-result \ + --result-dir $results_folder \ + --result-filename "$tag"-qps-"$qps".json \ + --request-rate "$qps" + + sleep 2 +} + + +main() { + + (which wget && which curl) || (apt-get update && apt-get install -y wget curl) + (which jq) || (apt-get -y install jq) + (which socat) || (apt-get -y install socat) + (which lsof) || (apt-get -y install lsof) + + pip install quart httpx matplotlib aiohttp datasets + + cd "$(dirname "$0")" + + cd .. + # create sonnet-4x.txt so that we can sample 2048 tokens for input + echo "" > sonnet_4x.txt + for _ in {1..4} + do + cat sonnet.txt >> sonnet_4x.txt + done + cd disagg_benchmarks + + rm -rf results + mkdir results + + default_output_len=6 + + export VLLM_HOST_IP=$(hostname -I | awk '{print $1}') + + launch_chunked_prefill + for qps in 2 4 6 8; do + benchmark $qps $default_output_len chunked_prefill + done + kill_gpu_processes + + launch_disagg_prefill + for qps in 2 4 6 8; do + benchmark $qps $default_output_len disagg_prefill + done + kill_gpu_processes + + python3 visualize_benchmark_results.py + +} + + +main "$@" diff --git a/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py new file mode 100644 index 0000000000000000000000000000000000000000..f62d8102e2d9f29c62f177d62453be5d8496cc1e --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py @@ -0,0 +1,63 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import os + +import aiohttp +from quart import Quart, make_response, request + +AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60) + +app = Quart(__name__) + + +async def forward_request(url, data): + async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session: + headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"} + async with session.post(url=url, json=data, headers=headers) as response: + if response.status == 200: + # if response.headers.get('Transfer-Encoding') == 'chunked': + if True: + async for chunk_bytes in response.content.iter_chunked(1024): + yield chunk_bytes + else: + content = await response.read() + yield content + + +@app.route("/v1/completions", methods=["POST"]) +async def handle_request(): + try: + original_request_data = await request.get_json() + + prefill_request = original_request_data.copy() + # change max_tokens = 1 to let it only do prefill + prefill_request["max_tokens"] = 1 + + # finish prefill + async for _ in forward_request( + "http://localhost:8100/v1/completions", prefill_request + ): + continue + + # return decode + generator = forward_request( + "http://localhost:8200/v1/completions", original_request_data + ) + response = await make_response(generator) + response.timeout = None + + return response + + except Exception as e: + import sys + import traceback + + exc_info = sys.exc_info() + print("Error occurred in disagg prefill proxy server") + print(e) + print("".join(traceback.format_exception(*exc_info))) + + +if __name__ == "__main__": + app.run(port=8000) diff --git a/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/round_robin_proxy.py b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/round_robin_proxy.py new file mode 100644 index 0000000000000000000000000000000000000000..b1df2f255822dad046f5dfcdc1d6538006463510 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/round_robin_proxy.py @@ -0,0 +1,63 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import asyncio +import itertools + +import aiohttp +from aiohttp import web + + +class RoundRobinProxy: + def __init__(self, target_ports): + self.target_ports = target_ports + self.port_cycle = itertools.cycle(self.target_ports) + + async def handle_request(self, request): + target_port = next(self.port_cycle) + target_url = f"http://localhost:{target_port}{request.path_qs}" + + async with aiohttp.ClientSession() as session: + try: + # Forward the request + async with session.request( + method=request.method, + url=target_url, + headers=request.headers, + data=request.content, + ) as response: + # Start sending the response + resp = web.StreamResponse( + status=response.status, headers=response.headers + ) + await resp.prepare(request) + + # Stream the response content + async for chunk in response.content.iter_any(): + await resp.write(chunk) + + await resp.write_eof() + return resp + + except Exception as e: + return web.Response(text=f"Error: {str(e)}", status=500) + + +async def main(): + proxy = RoundRobinProxy([8100, 8200]) + app = web.Application() + app.router.add_route("*", "/{path:.*}", proxy.handle_request) + + runner = web.AppRunner(app) + await runner.setup() + site = web.TCPSite(runner, "localhost", 8000) + await site.start() + + print("Proxy server started on http://localhost:8000") + + # Keep the server running + await asyncio.Event().wait() + + +if __name__ == "__main__": + asyncio.run(main()) diff --git a/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/visualize_benchmark_results.py b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/visualize_benchmark_results.py new file mode 100644 index 0000000000000000000000000000000000000000..74fa56d076cf14bc066468be24f6053b45166001 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/disagg_benchmarks/visualize_benchmark_results.py @@ -0,0 +1,47 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import json + +import matplotlib.pyplot as plt +import pandas as pd + +if __name__ == "__main__": + data = [] + for name in ["disagg_prefill", "chunked_prefill"]: + for qps in [2, 4, 6, 8]: + with open(f"results/{name}-qps-{qps}.json") as f: + x = json.load(f) + x["name"] = name + x["qps"] = qps + data.append(x) + + df = pd.DataFrame.from_dict(data) + dis_df = df[df["name"] == "disagg_prefill"] + chu_df = df[df["name"] == "chunked_prefill"] + + plt.style.use("bmh") + plt.rcParams["font.size"] = 20 + + for key in [ + "mean_ttft_ms", + "median_ttft_ms", + "p99_ttft_ms", + "mean_itl_ms", + "median_itl_ms", + "p99_itl_ms", + ]: + fig, ax = plt.subplots(figsize=(11, 7)) + plt.plot( + dis_df["qps"], dis_df[key], label="disagg_prefill", marker="o", linewidth=4 + ) + plt.plot( + chu_df["qps"], chu_df[key], label="chunked_prefill", marker="o", linewidth=4 + ) + ax.legend() + + ax.set_xlabel("QPS") + ax.set_ylabel(key) + ax.set_ylim(bottom=0) + fig.savefig(f"results/{key}.png") + plt.close(fig) diff --git a/online_apiserver_test_maxbs/benchmarks/fused_kernels/layernorm_rms_benchmarks.py b/online_apiserver_test_maxbs/benchmarks/fused_kernels/layernorm_rms_benchmarks.py new file mode 100644 index 0000000000000000000000000000000000000000..901524214469e8677cbb58c3078e8b1d724585bb --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/fused_kernels/layernorm_rms_benchmarks.py @@ -0,0 +1,228 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import pickle as pkl +import time +from collections.abc import Iterable +from dataclasses import dataclass +from itertools import product +from typing import Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from tqdm import tqdm + +import vllm._custom_ops as ops +from vllm.model_executor.layers.layernorm import RMSNorm + + +@dataclass +class bench_params_t: + num_tokens: int + hidden_size: int + add_residual: bool + dtype: torch.dtype + + def description(self): + return ( + f"N {self.num_tokens} " + f"x D {self.hidden_size} " + f"x R {self.add_residual} " + f"x DT {self.dtype}" + ) + + +def get_bench_params() -> list[bench_params_t]: + ## Test Fixtures + NUM_TOKENS = [2**x for x in range(11)] + HIDDEN_SIZES = list(range(1024, 8129, 1024)) + ADD_RESIDUAL = [True, False] + DTYPES = [torch.bfloat16, torch.float] + + combinations = product(NUM_TOKENS, HIDDEN_SIZES, ADD_RESIDUAL, DTYPES) + bench_params = list( + map(lambda x: bench_params_t(x[0], x[1], x[2], x[3]), combinations) + ) + return bench_params + + +# Reference impls +def unfused_int8_impl( + rms_norm_layer: RMSNorm, + x: torch.Tensor, + residual: Optional[torch.Tensor], + quant_dtype: torch.dtype, +): + # Norm + torch_out = None + if residual is None: + torch_out = rms_norm_layer.forward_cuda(x, residual) + else: + torch_out, _ = rms_norm_layer.forward_cuda(x, residual) + + # Quant + torch_out, _, _ = ops.scaled_int8_quant(torch_out) + + +def unfused_fp8_impl( + rms_norm_layer: RMSNorm, + x: torch.Tensor, + residual: Optional[torch.Tensor], + quant_dtype: torch.dtype, +): + # Norm + torch_out = None + if residual is None: + torch_out = rms_norm_layer.forward_cuda(x, residual) + else: + torch_out, _ = rms_norm_layer.forward_cuda(x, residual) + + # Quant + torch_out, _ = ops.scaled_fp8_quant(torch_out) + + +def fused_impl( + rms_norm_layer: RMSNorm, # this stores the weights + x: torch.Tensor, + residual: Optional[torch.Tensor], + quant_dtype: torch.dtype, +): + out, _ = ops.rms_norm_dynamic_per_token_quant( + x, rms_norm_layer.weight, 1e-6, quant_dtype, residual=residual + ) + + +# Bench functions +def bench_fn( + rms_norm_layer: RMSNorm, + x: torch.Tensor, + residual: torch.Tensor, + quant_dtype: torch.dtype, + label: str, + sub_label: str, + fn: Callable, + description: str, +) -> TMeasurement: + min_run_time = 1 + + globals = { + "rms_norm_layer": rms_norm_layer, + "x": x, + "residual": residual, + "quant_dtype": quant_dtype, + "fn": fn, + } + return TBenchmark.Timer( + stmt="fn(rms_norm_layer, x, residual, quant_dtype)", + globals=globals, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + +def bench(params: bench_params_t, label: str, sub_label: str) -> Iterable[TMeasurement]: + # Make inputs + layer = RMSNorm(params.hidden_size, 1e-6).to(dtype=params.dtype) + # Make weights + layer.weight.data.normal_(mean=1.0, std=0.1) + # Make inputs + scale = 1 / params.hidden_size + x = ( + torch.randn( + params.num_tokens, params.hidden_size, dtype=params.dtype, device="cuda" + ) + * scale + ) + residual = ( + (torch.randn_like(x) * scale).to(device="cuda") if params.add_residual else None + ) + + timers = [] + + # unfused int8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.int8, + label, + sub_label, + unfused_int8_impl, + "unfused_int8_impl", + ) + ) + + # unfused fp8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.float8_e4m3fn, + label, + sub_label, + unfused_fp8_impl, + "unfused_fp8_impl", + ) + ) + + # fused int8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.int8, + label, + sub_label, + fused_impl, + "fused_int8_impl", + ) + ) + + # fused fp8 impl. + timers.append( + bench_fn( + layer, + x, + residual, + torch.float8_e4m3fn, + label, + sub_label, + fused_impl, + "fused_fp8_impl", + ) + ) + + print_timers(timers) + + return timers + + +# launch bench +# runner +def print_timers(timers: Iterable[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def main(): + torch.set_default_device("cuda") + bench_params = get_bench_params() + + timers = [] + for bp in tqdm(bench_params): + timers.extend(bench(bp, "rms-norm-dynamic-per-token-quant", bp.description())) + print_timers(timers) + + # pickle all the results + timestamp = int(time.time()) + with open(f"rms_norm_dpt_quant-{timestamp}.pkl", "wb") as f: + pkl.dump(timers, f) + + +if __name__ == "__main__": + main() diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/bench_fp8_gemm.py b/online_apiserver_test_maxbs/benchmarks/kernels/bench_fp8_gemm.py new file mode 100644 index 0000000000000000000000000000000000000000..920961899038061c13125bdd37f9843b5e4d548a --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/bench_fp8_gemm.py @@ -0,0 +1,159 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import argparse +import copy +import itertools + +import torch +from weight_shapes import WEIGHT_SHAPES + +from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm +from vllm._custom_ops import scaled_fp8_quant as vllm_scaled_fp8_quant +from vllm.triton_utils import triton + +PROVIDER_CFGS = { + "torch-bf16": dict(enabled=True), + "fp8-tensor-w-token-a": dict( + w="tensor", a="token", no_a_quant=False, enabled=False + ), + "fp8-tensor-w-tensor-a": dict( + w="tensor", a="tensor", no_a_quant=False, enabled=True + ), + "fp8-channel-w-token-a": dict( + w="channel", a="token", no_a_quant=False, enabled=True + ), + "fp8-channel-w-tensor-a": dict( + w="channel", a="tensor", no_a_quant=False, enabled=False + ), + "fp8-tensor-w-token-a-noquant": dict( + w="tensor", a="token", no_a_quant=True, enabled=False + ), + "fp8-tensor-w-tensor-a-noquant": dict( + w="tensor", a="tensor", no_a_quant=True, enabled=True + ), + "fp8-channel-w-token-a-noquant": dict( + w="channel", a="token", no_a_quant=True, enabled=True + ), + "fp8-channel-w-tensor-a-noquant": dict( + w="channel", a="tensor", no_a_quant=True, enabled=False + ), +} + +_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]] + + +def _quant_weight_fp8(b: torch.Tensor, w_type: str, device: str): + if w_type == "tensor": + scale_b = torch.ones(1, device=device, dtype=torch.float32) + b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b) + else: + b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, use_per_token_if_dynamic=True) + return b_fp8.t(), scale_b_fp8 + + +def build_fp8_runner(cfg, a, b, dtype, device): + b_fp8, scale_b_fp8 = _quant_weight_fp8(b, cfg["w"], device) + + scale_a_const = ( + torch.ones(1, device=device, dtype=torch.float32) + if cfg["a"] == "tensor" + else None + ) + + if cfg["no_a_quant"]: + if cfg["a"] == "tensor": + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const) + else: + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True) + + def run(): + return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype) + + return run + + if cfg["a"] == "tensor": + + def run(): + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const) + return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype) + + else: + + def run(): + a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True) + return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype) + + return run + + +@triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["batch_size"], + x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384], + x_log=False, + line_arg="provider", + line_vals=_enabled, + line_names=_enabled, + ylabel="TFLOP/s (larger is better)", + plot_name="BF16 vs FP8 GEMMs", + args={}, + ) +) +def benchmark(batch_size, provider, N, K): + M = batch_size + device = "cuda" + dtype = torch.bfloat16 + + a = torch.randn((M, K), device=device, dtype=dtype) + b = torch.randn((N, K), device=device, dtype=dtype) + + quantiles = [0.5, 0.2, 0.8] + + if provider == "torch-bf16": + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: torch.nn.functional.linear(a, b), quantiles=quantiles + ) + else: + cfg = PROVIDER_CFGS[provider] + run_quant = build_fp8_runner(cfg, a, b, dtype, device) + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: run_quant(), quantiles=quantiles + ) + + to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3) + return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms) + + +def prepare_shapes(args): + out = [] + for model, tp_size in itertools.product(args.models, args.tp_sizes): + for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]): + KN[tp_dim] //= tp_size + KN.append(model) + out.append(KN) + return out + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--models", + nargs="+", + type=str, + default=["meta-llama/Llama-3.1-8B-Instruct"], + choices=list(WEIGHT_SHAPES.keys()), + ) + parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1]) + args = parser.parse_args() + + for K, N, model in prepare_shapes(args): + print(f"{model}, N={N} K={K}, BF16 vs FP8 GEMMs TFLOP/s:") + benchmark.run( + print_data=True, + show_plots=True, + save_path=f"bench_fp8_res_n{N}_k{K}", + N=N, + K=K, + ) + + print("Benchmark finished!") diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/bench_int8_gemm.py b/online_apiserver_test_maxbs/benchmarks/kernels/bench_int8_gemm.py new file mode 100644 index 0000000000000000000000000000000000000000..e9c6d64404d0dc6958aeb675bf6b893623649ffa --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/bench_int8_gemm.py @@ -0,0 +1,169 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import argparse +import copy +import itertools + +import torch +from weight_shapes import WEIGHT_SHAPES + +from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm +from vllm._custom_ops import scaled_int8_quant as vllm_scaled_int8_quant +from vllm.triton_utils import triton + +PROVIDER_CFGS = { + "torch-bf16": dict(enabled=True), + "int8-tensor-w-token-a": dict( + w="tensor", a="token", no_a_quant=False, enabled=False + ), + "int8-tensor-w-tensor-a": dict( + w="tensor", a="tensor", no_a_quant=False, enabled=True + ), + "int8-channel-w-token-a": dict( + w="channel", a="token", no_a_quant=False, enabled=True + ), + "int8-channel-w-tensor-a": dict( + w="channel", a="tensor", no_a_quant=False, enabled=False + ), + "int8-tensor-w-token-a-noquant": dict( + w="tensor", a="token", no_a_quant=True, enabled=False + ), + "int8-tensor-w-tensor-a-noquant": dict( + w="tensor", a="tensor", no_a_quant=True, enabled=True + ), + "int8-channel-w-token-a-noquant": dict( + w="channel", a="token", no_a_quant=True, enabled=True + ), + "int8-channel-w-tensor-a-noquant": dict( + w="channel", a="tensor", no_a_quant=True, enabled=False + ), +} + + +def _quant_weight(b, w_type, device): + if w_type == "tensor": + scale_b = torch.ones(1, device=device, dtype=torch.float32) + b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b, scale_b) + assert scale_b_int8.numel() == 1 + else: # channel + b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b) + assert scale_b_int8.numel() == b.shape[0] + return b_int8.t(), scale_b_int8 + + +def build_int8_runner(cfg, a, b, dtype, device): + # quant before running the kernel + b_int8, scale_b_int8 = _quant_weight(b, cfg["w"], device) + + scale_a_const = None + if cfg["a"] == "tensor": + scale_a_const = torch.ones(1, device=device, dtype=torch.float32) + + # no quant, create activation ahead + if cfg["no_a_quant"]: + if cfg["a"] == "tensor": + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const) + else: # token + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a) + + def run_quant(): + return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype) + + return run_quant + + # dynamic quant, create activation inside + if cfg["a"] == "tensor": + + def run_quant(): + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const) + return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype) + + else: # token + + def run_quant(): + a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a) + return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype) + + return run_quant + + +_enabled = [k for k, v in PROVIDER_CFGS.items() if v.get("enabled")] + + +@triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["batch_size"], + x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384], + x_log=False, + line_arg="provider", + line_vals=_enabled, + line_names=[k for k in _enabled], + ylabel="TFLOP/s (larger is better)", + plot_name="BF16 vs INT8 GEMMs", + args={}, + ) +) +def benchmark(batch_size, provider, N, K): + M = batch_size + device = "cuda" + dtype = torch.bfloat16 + a = torch.randn((M, K), device=device, dtype=dtype) + b = torch.randn((N, K), device=device, dtype=dtype) + + quantiles = [0.5, 0.2, 0.8] + + if provider == "torch-bf16": + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: torch.nn.functional.linear(a, b), quantiles=quantiles + ) + else: + cfg = PROVIDER_CFGS[provider] + run_quant = build_int8_runner(cfg, a, b, dtype, device) + ms, min_ms, max_ms = triton.testing.do_bench_cudagraph( + lambda: run_quant(), quantiles=quantiles + ) + + to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3) + return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms) + + +def prepare_shapes(args): + KN_model_names = [] + for model, tp_size in itertools.product(args.models, args.tp_sizes): + for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]): + KN[tp_dim] //= tp_size + KN.append(model) + KN_model_names.append(KN) + return KN_model_names + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--models", + nargs="+", + type=str, + default=["meta-llama/Llama-3.1-8B-Instruct"], + choices=list(WEIGHT_SHAPES.keys()), + help="List of models to benchmark", + ) + parser.add_argument( + "--tp-sizes", + nargs="+", + type=int, + default=[1], + help="List of tensor parallel sizes", + ) + args = parser.parse_args() + + for K, N, model in prepare_shapes(args): + print(f"{model}, N={N} K={K}, BF16 vs INT8 GEMMs TFLOP/s:") + benchmark.run( + print_data=True, + show_plots=True, + save_path=f"bench_int8_res_n{N}_k{K}", + N=N, + K=K, + ) + + print("Benchmark finished!") diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_aqlm.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_aqlm.py new file mode 100644 index 0000000000000000000000000000000000000000..42de062b08e424619b010f9170725b6785a35f58 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_aqlm.py @@ -0,0 +1,345 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import os +import sys +from typing import Optional + +import torch +import torch.nn.functional as F + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.aqlm import ( + dequantize_weight, + generic_dequantize_gemm, + get_int_dtype, + optimized_dequantize_gemm, +) +from vllm.utils import FlexibleArgumentParser + +os.environ["CUDA_VISIBLE_DEVICES"] = "0" + + +def torch_mult( + # [..., in_features] + input: torch.Tensor, + weights: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, +) -> torch.Tensor: + output = F.linear(input, weights) + return output + + +def dequant_out_scale( + # [..., in_features] + input: torch.Tensor, + # [num_out_groups, num_in_groups, num_codebooks] + codes: torch.IntTensor, + # [num_codebooks, codebook_size, out_group_size, in_group_size] + codebooks: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, + output_partition_sizes: torch.IntTensor, + bias: Optional[torch.Tensor], +) -> torch.Tensor: + weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes) + + if bias is None: + output = F.linear(input, weights, bias) + orig_shape = output.shape + flattened_output = output.view(-1, output.size(-1)) + f_scales = scales.view(-1, scales.shape[0]) + b_scales = f_scales.expand(flattened_output.shape[0], -1) + flattened_output *= b_scales + return flattened_output.view(orig_shape) + else: + b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1]) + weights *= b_scales + return F.linear(input, weights, bias) + + +def dequant_weight_scale( + # [..., in_features] + input: torch.Tensor, + # [num_out_groups, num_in_groups, num_codebooks] + codes: torch.IntTensor, + # [num_codebooks, codebook_size, out_group_size, in_group_size] + codebooks: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, + output_partition_sizes: torch.IntTensor, + bias: Optional[torch.Tensor], +) -> torch.Tensor: + weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes) + + b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1]) + weights *= b_scales + return F.linear(input, weights, bias) + + +def dequant_no_scale( + # [..., in_features] + input: torch.Tensor, + # [num_out_groups, num_in_groups, num_codebooks] + codes: torch.IntTensor, + # [num_codebooks, codebook_size, out_group_size, in_group_size] + codebooks: torch.Tensor, + # [num_out_groups, 1, 1, 1] + scales: torch.Tensor, + output_partition_sizes: torch.IntTensor, + bias: Optional[torch.Tensor], +) -> torch.Tensor: + weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes) + + return F.linear(input, weights, bias) + + +# Compare the optimized 1x16 and 2x8 cuda decompression/dequant kernels against +# the generic pytorch version. +# Just visual comparison. +def dequant_test(k: int, parts: torch.Tensor, nbooks: int, bits: int) -> None: + n = int(parts.sum().item()) + + device = torch.device("cuda:0") + + code_range = (1 << bits) // 2 + ingroups = 8 + + codes = torch.randint( + -code_range, + code_range, + size=(n, k // ingroups, nbooks), + dtype=get_int_dtype(bits), + device=device, + ) + + codebooks = torch.randn( + size=(parts.shape[0] * nbooks, 1 << bits, 1, 8), + dtype=torch.float16, + device=device, + ) + + count = 0 + for index in range(16): + for i in range(8): + for book in range(nbooks): + codebooks[book, index, 0, i] = count * (10**book) + count += 1 + + print("codes shape", codes.shape) + + for i in range(16): + for book in range(nbooks): + codes[0, i, book] = i + codes[0, -i, book] = i + + weights = dequantize_weight(codes, codebooks, None) + weights2 = ops.aqlm_dequant(codes, codebooks, parts) + + print("weights shape:", weights.shape) + print("weights2 shape:", weights2.shape) + + print("weights are:", weights) + print("weights2 are:", weights2) + + print("first 128 weights are", weights[0, 0:128].to(torch.int32)) + print("first 128 weights2 are:", weights2[0, 0:128].to(torch.int32)) + + print("last 128 weights are", weights[0, -128:]) + print("last 128 weights2 are:", weights2[0, -128:]) + + +def main(): + parser = FlexibleArgumentParser(description="Benchmark aqlm performance.") + + # Add arguments + parser.add_argument( + "--nbooks", type=int, default=1, help="Number of codebooks (default: 1)" + ) + parser.add_argument( + "--bits", + type=int, + default=16, + help="Number of bits per code element (default: 16)", + ) + parser.add_argument( + "--test", + type=bool, + default=False, + help="Run the decompression/dequant tester rather than benchmarking " + "(default: False)", + ) + + # Parse the arguments + args = parser.parse_args() + + # Extract values + nbooks = args.nbooks + bits = args.bits + + if args.test: + dequant_test(4096, torch.tensor((4096,)), nbooks, bits) + return + + # Otherwise, benchmark. + methods = [ + ops.aqlm_gemm, + dequant_out_scale, + generic_dequantize_gemm, + optimized_dequantize_gemm, + dequant_weight_scale, + torch_mult, + dequant_no_scale, + ] + + filename = f"./aqlm_benchmark_{nbooks}x{bits}.csv" + print(f"writing benchmarks to file {filename}") + with open(filename, "w") as f: + sys.stdout = f + + print("m | k | n | n parts", end="") + for method in methods: + print(f" | {method.__name__.replace('_', ' ')} (µs)", end="") + print("") + + # These are reasonable prefill sizes. + ksandpartions = ( + (4096, (4096, 4096, 4096)), + (4096, (4096,)), + (4096, (11008, 11008)), + (11008, (4096,)), + ) + + # reasonable ranges for m. + for m in [ + 1, + 2, + 4, + 8, + 10, + 12, + 14, + 16, + 24, + 32, + 48, + 52, + 56, + 64, + 96, + 112, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ]: + print(f"{m}", file=sys.__stdout__) + for ksp in ksandpartions: + run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits, methods) + + sys.stdout = sys.__stdout__ + + +def run_grid(m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, methods): + # I didn't see visible improvements from increasing these, but feel free :) + num_warmup_trials = 1 + num_trials = 1 + + num_calls = 100 + + # warmup. + for method in methods: + for _ in range(num_warmup_trials): + run_timing( + num_calls=num_calls, + m=m, + k=k, + parts=parts, + nbooks=nbooks, + bits=bits, + method=method, + ) + + n = parts.sum().item() + print(f"{m} | {k} | {n} | {parts.tolist()}", end="") + + for method in methods: + best_time_us = 1e20 + for _ in range(num_trials): + kernel_dur_ms = run_timing( + num_calls=num_calls, + m=m, + k=k, + parts=parts, + nbooks=nbooks, + bits=bits, + method=method, + ) + + kernel_dur_us = 1000 * kernel_dur_ms + + if kernel_dur_us < best_time_us: + best_time_us = kernel_dur_us + + print(f" | {kernel_dur_us:.0f}", end="") + + print("") + + +def run_timing( + num_calls: int, m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, method +) -> float: + n = int(parts.sum().item()) + + device = torch.device("cuda:0") + + input = torch.randn((1, m, k), dtype=torch.float16, device=device) + + code_range = (1 << bits) // 2 + ingroups = 8 + + codes = torch.randint( + -code_range, + code_range, + size=(n, k // ingroups, nbooks), + dtype=get_int_dtype(bits), + device=device, + ) + + codebooks = torch.randn( + size=(parts.shape[0] * nbooks, 1 << bits, 1, 8), + dtype=torch.float16, + device=device, + ) + + scales = torch.randn(size=(n, 1, 1, 1), dtype=torch.float16, device=device) + + # for comparison to just a pytorch mult. + weights = torch.randn((n, k), dtype=torch.float16, device=device) + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + start_event.record() + + if method is torch_mult: + for i in range(num_calls): + torch_mult(input, weights, scales) + else: + for i in range(num_calls): + method(input, codes, codebooks, scales, parts, None) + + end_event.record() + end_event.synchronize() + + dur_ms = start_event.elapsed_time(end_event) / num_calls + return dur_ms + + +if __name__ == "__main__": + sys.exit(main()) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_bitblas.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_bitblas.py new file mode 100644 index 0000000000000000000000000000000000000000..97ee060341373022b06c6ce62600d5c23c771c37 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_bitblas.py @@ -0,0 +1,242 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +# Copyright (c) Microsoft Corporation. +# Licensed under the MIT License. + +from vllm.model_executor.layers.quantization.utils.bitblas_utils import ( + MINIMUM_BITBLAS_VERSION, +) + +try: + import bitblas + + if bitblas.__version__ < MINIMUM_BITBLAS_VERSION: + raise ImportError( + "bitblas version is wrong. Please " + f"install bitblas>={MINIMUM_BITBLAS_VERSION}" + ) +except ImportError as e: + bitblas_import_exception = e + raise ValueError( + "Trying to use the bitblas backend, but could not import" + f"with the following error: {bitblas_import_exception}. " + "Please install bitblas through the following command: " + f"`pip install bitblas>={MINIMUM_BITBLAS_VERSION}`" + ) from bitblas_import_exception + +from bitblas import Matmul, MatmulConfig, auto_detect_nvidia_target + +from vllm.utils import FlexibleArgumentParser + +parser = FlexibleArgumentParser( + description="Benchmark BitBLAS int4 on a specific target." +) + +# Add arguments to the parser +parser.add_argument( + "--target", + type=str, + default=auto_detect_nvidia_target(), + help="Specify the target device for benchmarking.", +) +parser.add_argument( + "--group_size", type=int, default=None, help="Group size for grouped quantization." +) +parser.add_argument( + "--A_dtype", + type=str, + default="float16", + choices=["float16", "float32", "float64", "int32", "int8"], + help="Data type of activation A.", +) +parser.add_argument( + "--W_dtype", + type=str, + default="int4", + choices=[ + "float16", + "float32", + "float64", + "int32", + "int8", + "int4", + "int2", + "int1", + "nf4", + "fp4_e2m1", + ], + help="Data type of weight W.", +) +parser.add_argument( + "--accum_dtype", + type=str, + default="float16", + choices=["float16", "int32"], + help="Data type for accumulation.", +) +parser.add_argument( + "--out_dtype", + type=str, + default="float16", + choices=["float16", "float32", "int32", "int8"], + help="Data type for output.", +) +parser.add_argument( + "--layout", + type=str, + default="nt", + choices=["nt", "nn"], + help="Matrix layout, 'nt' for non-transpose A and transpose W.", +) +parser.add_argument( + "--with_bias", action="store_true", help="Include bias in the benchmark." +) +parser.add_argument( + "--with_scaling", + action="store_true", + help="Include scaling factor in the quantization.", +) +parser.add_argument( + "--with_zeros", action="store_true", help="Include zeros in the quantization." +) +parser.add_argument( + "--zeros_mode", + type=str, + default=None, + choices=["original", "rescale", "quantized"], + help="Specify the mode for calculating zeros.", +) + +# Parse the arguments +args = parser.parse_args() + +# Assign arguments to variables +target = args.target +A_dtype = args.A_dtype +W_dtype = args.W_dtype +accum_dtype = args.accum_dtype +out_dtype = args.out_dtype +layout = args.layout +with_bias = args.with_bias +group_size = args.group_size +with_scaling = args.with_scaling +with_zeros = args.with_zeros +zeros_mode = args.zeros_mode + +# Define a list of shared arguments that repeat in every config +shared_args = [ + A_dtype, + W_dtype, + out_dtype, + accum_dtype, + layout, + with_bias, + group_size, + with_scaling, + with_zeros, + zeros_mode, +] + +# Define just the (M, K, N) shapes in a more compact list +shapes = [ + # square test + (1, 16384, 16384), + # BLOOM-176B + (1, 43008, 14336), + (1, 14336, 14336), + (1, 57344, 14336), + (1, 14336, 57344), + # OPT-65B + (1, 9216, 9216), + (1, 36864, 9216), + (1, 9216, 36864), + (1, 22016, 8192), + # LLAMA-70B/65B + (1, 8192, 22016), + (1, 8192, 8192), + (1, 28672, 8192), + (1, 8192, 28672), + # square test + (16384, 16384, 16384), + # BLOOM-176B + (8192, 43008, 14336), + (8192, 14336, 14336), + (8192, 57344, 14336), + (8192, 14336, 57344), + # OPT-65B + (8192, 9216, 9216), + (8192, 36864, 9216), + (8192, 9216, 36864), + (8192, 22016, 8192), + # LLAMA-70B/65B + (8192, 8192, 22016), + (8192, 8192, 8192), + (8192, 28672, 8192), + (8192, 8192, 28672), +] + +# Build test shapes with all the shared arguments +test_shapes = [(MatmulConfig, Matmul, (*shape, *shared_args)) for shape in shapes] + +benchmark_sets = [] +benchmark_sets.extend(test_shapes) + +benchmark_results = {} +for config_class, operator, input_args in benchmark_sets: + config = config_class(*input_args) + matmul = operator(config, target=target, enable_tuning=True) + kernel_latency = matmul.profile_latency() + + print("Time cost is: {:.3f} ms".format(kernel_latency)) + + profile_config = { + f"{operator.__name__}-{'-'.join([str(i) for i in input_args])}": { + "BitBLAS_top20_latency": kernel_latency, + } + } + + benchmark_results.update(profile_config) + +# Define headers for the table +headers = [ + "PrimFunc", + "Input Arguments", + "BitBLAS Top20 Latency", +] + +# Calculate column widths for pretty printing +col_widths = [0, 0, 0] +for config_key, values in benchmark_results.items(): + args_split = config_key.split("-") + func_name = args_split[0] + input_args_str = "-".join(args_split[1:]) + col_widths[0] = max(col_widths[0], len(func_name) + 2, len(headers[0]) + 2) + col_widths[1] = max(col_widths[1], len(input_args_str) + 2, len(headers[1]) + 2) + col_widths[2] = max( + col_widths[2], + len(f"{values['BitBLAS_top20_latency']:.3f} ms") + 2, + len(headers[2]) + 2, + ) + # break only if you want to measure widths from a single example; + # otherwise, let it loop over all items. + +# Print header +for i, header in enumerate(headers): + headers[i] = header.ljust(col_widths[i]) +print("".join(headers)) +print("-" * sum(col_widths)) + +# Print rows +for config_key, values in benchmark_results.items(): + args_split = config_key.split("-") + func_name = args_split[0] + input_args_str = "-".join(args_split[1:]) + row = [ + func_name, + input_args_str, + f"{values['BitBLAS_top20_latency']:.3f} ms", + ] + row_str = "".join( + [str(cell).ljust(col_widths[idx]) for idx, cell in enumerate(row)] + ) + print(row_str) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_cutlass_fp4_moe.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_cutlass_fp4_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..35c20ee41b9a94c10f1367a45fe391cc15e9ffec --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_cutlass_fp4_moe.py @@ -0,0 +1,490 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +""" +Benchmark the performance of the cutlass_moe_fp4 kernel vs the triton_moe +kernel. The cutlass_moe_fp4 kernel takes in fp4 quantized weights and 16-bit +activations. The triton_moe kernel takes in fp8 weights(tensor scaled to fp8) +and 16-bit activations. +""" + +import nvtx +import torch +import torch.utils.benchmark as benchmark + +from vllm import _custom_ops as ops +from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config +from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4 +from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk +from vllm.scalar_type import scalar_types +from vllm.utils import FlexibleArgumentParser + +WEIGHT_SHAPES_MOE = { + "nvidia/DeepSeek-R1-FP4": [ + [256, 8, 2048, 7168], + ], +} + +DEFAULT_MODELS = [ + "nvidia/DeepSeek-R1-FP4", +] + +DEFAULT_BATCH_SIZES = [4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048] +DEFAULT_TP_SIZES = [1] + +PER_ACT_TOKEN_OPTS = [False] +PER_OUT_CH_OPTS = [False] +FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max() +FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max + + +def to_fp8(tensor: torch.Tensor): + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to( + dtype=torch.float8_e4m3fn + ) + + +def bench_run( + results: list[benchmark.Measurement], + model: str, + num_experts: int, + topk: int, + per_act_token: bool, + per_out_ch: bool, + mkn: tuple[int, int, int], +): + label = "NVFP4 Blockscaled CUTLASS MOE vs FP8 Tensor Scaled Triton" + + sub_label = ( + "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, MKN=({})".format( + model, num_experts, topk, per_act_token, per_out_ch, mkn + ) + ) + + print(f"Testing: {sub_label}") + + (m, k, n) = mkn + + dtype = torch.half + device = "cuda" + a = torch.randn((m, k), device=device, dtype=dtype) / 10 + w1 = torch.randn((num_experts, 2 * n, k), device=device, dtype=dtype) / 10 + w2 = torch.randn((num_experts, k, n), device=device, dtype=dtype) / 10 + + _, a_fp8_scale = ops.scaled_fp8_quant(a) + + w1_fp8q = torch.empty( + (num_experts, 2 * n, k), device=device, dtype=torch.float8_e4m3fn + ) + w2_fp8q = torch.empty((num_experts, k, n), device=device, dtype=torch.float8_e4m3fn) + w1_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32) + w2_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32) + + for expert in range(num_experts): + w1_fp8q[expert], w1_fp8scale[expert] = ops.scaled_fp8_quant(w1[expert]) + w2_fp8q[expert], w2_fp8scale[expert] = ops.scaled_fp8_quant(w2[expert]) + + w1_fp8q_notransp = w1_fp8q.clone() + w2_fp8q_notransp = w2_fp8q.clone() + w1_fp8q = w1_fp8q.transpose(1, 2) + w2_fp8q = w2_fp8q.transpose(1, 2) + + score = torch.randn((m, num_experts), device=device, dtype=dtype) + + topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False) + + quant_blocksize = 16 + w1_blockscale = torch.empty( + (num_experts, 2 * n, k // quant_blocksize), + device=device, + dtype=torch.float8_e4m3fn, + ) + w2_blockscale = torch.empty( + (num_experts, k, n // quant_blocksize), device=device, dtype=torch.float8_e4m3fn + ) + + # n_b_scales = 2 * n if per_out_ch else 1 + # k_b_scales = k if per_out_ch else 1 + w1_fp4 = torch.empty((num_experts, 2 * n, k // 2), device=device, dtype=torch.uint8) + w2_fp4 = torch.empty((num_experts, k, n // 2), device=device, dtype=torch.uint8) + + w1_gs = torch.empty((num_experts,), device=device, dtype=torch.float32) + w2_gs = torch.empty((num_experts,), device=device, dtype=torch.float32) + a1_gs = torch.ones((num_experts,), device=device, dtype=torch.float32) + a2_gs = torch.ones((num_experts,), device=device, dtype=torch.float32) + + for expert in range(num_experts): + w1_e = w1[expert] + w2_e = w2[expert] + w1_amax = torch.abs(w1_e).max().to(torch.float32) + w2_amax = torch.abs(w2_e).max().to(torch.float32) + w1_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w1_amax + w2_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w2_amax + + w1_fp4[expert], w1_blockscale[expert] = ops.scaled_fp4_quant( + w1_e, w1_gs[expert] + ) + + w2_fp4[expert], w2_blockscale[expert] = ops.scaled_fp4_quant( + w2_e, w2_gs[expert] + ) + + def run_triton_moe( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_fp8_scale: torch.Tensor, + num_repeats: int, + ): + for _ in range(num_repeats): + fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_fp8_scale, + ) + + def run_cutlass_moe_fp4( + a: torch.Tensor, + w1_fp4: torch.Tensor, + w2_fp4: torch.Tensor, + w1_blockscale: torch.Tensor, + w2_blockscale: torch.Tensor, + w1_gs: torch.Tensor, + w2_gs: torch.Tensor, + a1_gs: torch.Tensor, + a2_gs: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + m: int, + n: int, + k: int, + e: int, + device: torch.device, + num_repeats: int, + ): + for _ in range(num_repeats): + with nvtx.annotate("cutlass_moe_fp4", color="green"): + cutlass_moe_fp4( + a=a, + a1_gscale=a1_gs, + a2_gscale=a2_gs, + w1_fp4=w1_fp4, + w1_blockscale=w1_blockscale, + w1_alphas=w1_gs, + w2_fp4=w2_fp4, + w2_blockscale=w2_blockscale, + w2_alphas=w2_gs, + topk_weights=topk_weights, + topk_ids=topk_ids, + m=m, + n=n, + k=k, + e=num_experts, + device=device, + ) + + def run_cutlass_from_graph( + a: torch.Tensor, + a1_gscale: torch.Tensor, + w1_fp4: torch.Tensor, + w1_blockscale: torch.Tensor, + w1_alphas: torch.Tensor, + a2_gscale: torch.Tensor, + w2_fp4: torch.Tensor, + w2_blockscale: torch.Tensor, + w2_alphas: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + m: int, + n: int, + k: int, + e: int, + device: torch.device, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return cutlass_moe_fp4( + a=a, + a1_gscale=a1_gs, + w1_fp4=w1_fp4, + w1_blockscale=w1_blockscale, + w1_alphas=w1_alphas, + a2_gscale=a2_gs, + w2_fp4=w2_fp4, + w2_blockscale=w2_blockscale, + w2_alphas=w2_alphas, + topk_weights=topk_weights, + topk_ids=topk_ids, + m=m, + n=n, + k=k, + e=num_experts, + device=device, + ) + + def run_triton_from_graph( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_fp8_scale: torch.Tensor, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_fp8_scale, + ) + + def replay_graph(graph, num_repeats): + for _ in range(num_repeats): + graph.replay() + torch.cuda.synchronize() + + cutlass_stream = torch.cuda.Stream() + cutlass_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(cutlass_graph, stream=cutlass_stream): + run_cutlass_from_graph( + a=a, + a1_gscale=a1_gs, + w1_fp4=w1_fp4, + w1_blockscale=w1_blockscale, + w1_alphas=w1_gs, + a2_gscale=a2_gs, + w2_fp4=w2_fp4, + w2_blockscale=w2_blockscale, + w2_alphas=w2_gs, + topk_weights=topk_weights, + topk_ids=topk_ids, + m=m, + n=n, + k=k, + e=num_experts, + device=device, + ) + torch.cuda.synchronize() + + triton_stream = torch.cuda.Stream() + triton_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(triton_graph, stream=triton_stream): + run_triton_from_graph( + a, + w1_fp8q_notransp, + w2_fp8q_notransp, + topk_weights, + topk_ids, + w1_fp8scale, + w2_fp8scale, + a_fp8_scale, + ) + torch.cuda.synchronize() + + min_run_time = 5 + num_warmup = 5 + num_runs = 25 + + globals = { + # Baseline params + "w1": w1, + "w2": w2, + "score": score, + "topk": topk, + "w1_fp8q_notransp": w1_fp8q_notransp, + "w2_fp8q_notransp": w2_fp8q_notransp, + "w1_fp8scale": w1_fp8scale, + "w2_fp8scale": w2_fp8scale, + "a_fp8_scale": a_fp8_scale, + # Cutlass params + "a": a, + "a1_gscale": a1_gs, + "w1_fp4": w1_fp4, + "w1_blockscale": w1_blockscale, + "w1_alphas": w1_gs, + "a2_gscale": a2_gs, + "w2_fp4": w2_fp4, + "w2_blockscale": w2_blockscale, + "w2_alphas": w2_gs, + "topk_weights": topk_weights, + "topk_ids": topk_ids, + "m": m, + "n": n, + "k": k, + "e": num_experts, + "device": device, + # cuda graph params + "cutlass_graph": cutlass_graph, + "triton_graph": triton_graph, + # Gen params + "num_runs": num_runs, + # Kernels + "run_triton_moe": run_triton_moe, + "run_cutlass_moe_fp4": run_cutlass_moe_fp4, + "replay_graph": replay_graph, + } + + # Warmup + run_triton_moe( + a, + w1_fp8q_notransp, + w2_fp8q_notransp, + topk_weights, + topk_ids, + w1_fp8scale, + w2_fp8scale, + a_fp8_scale, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_triton_moe(a, w1_fp8q_notransp, w2_fp8q_notransp, topk_weights, topk_ids, w1_fp8scale, w2_fp8scale, a_fp8_scale, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(triton_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(triton_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + + run_cutlass_moe_fp4( + a, + w1_fp4, + w2_fp4, + w1_blockscale, + w2_blockscale, + w1_gs, + w2_gs, + a1_gs, + a2_gs, + topk_weights, + topk_ids, + m, + n, + k, + num_experts, + device, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_cutlass_moe_fp4(a, w1_fp4, w2_fp4, w1_blockscale, w2_blockscale, w1_alphas, w2_alphas, a1_gscale, a2_gscale, topk_weights, topk_ids, m, n, k, e, device, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="cutlass_moe_fp4", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(cutlass_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(cutlass_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="cutlass_moe_fp4_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + +def main(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + results: list[benchmark.Measurement] = [] + + for model in args.models: + for tp in args.tp_sizes: + for layer in WEIGHT_SHAPES_MOE[model]: + num_experts = layer[0] + topk = layer[1] + size_k = layer[2] + size_n = layer[3] // tp + + if len(args.limit_k) > 0 and size_k not in args.limit_k: + continue + + if len(args.limit_n) > 0 and size_n not in args.limit_n: + continue + + for per_act_token in PER_ACT_TOKEN_OPTS: + for per_out_ch in PER_OUT_CH_OPTS: + for size_m in args.batch_sizes: + mkn = (size_m, size_k, size_n) + bench_run( + results, + model, + num_experts, + topk, + per_act_token, + per_out_ch, + mkn, + ) + + compare = benchmark.Compare(results) + compare.print() + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark NVFP4 CUTLASS MOE across specified models/shapes/batches" + ) + parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES_MOE.keys(), + ) + parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES) + parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + parser.add_argument("--limit-k", nargs="+", type=int, default=[]) + parser.add_argument("--limit-n", nargs="+", type=int, default=[]) + parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-act-token", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[]) + + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py new file mode 100644 index 0000000000000000000000000000000000000000..1d4e730f99ae911535caa4cca17122656db06a4b --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py @@ -0,0 +1,383 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import torch +import torch.utils.benchmark as benchmark +from benchmark_shapes import WEIGHT_SHAPES_MOE + +from vllm import _custom_ops as ops +from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config +from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8 +from vllm.model_executor.layers.fused_moe.fused_moe import ( + fused_experts, + fused_topk, +) +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = [ + "nm-testing/Mixtral-8x7B-Instruct-v0.1", + "nm-testing/deepseekv2-lite", + "ibm-granite/granite-3.0-1b-a400m", + "ibm-granite/granite-3.0-3b-a800m", +] +DEFAULT_BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512] +DEFAULT_TP_SIZES = [1] + +PER_ACT_TOKEN_OPTS = [False] +PER_OUT_CH_OPTS = [False] + + +def to_fp8(tensor: torch.Tensor): + finfo = torch.finfo(torch.float8_e4m3fn) + return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to( + dtype=torch.float8_e4m3fn + ) + + +def bench_run( + results: list[benchmark.Measurement], + model: str, + num_experts: int, + topk: int, + per_act_token: bool, + per_out_ch: bool, + mkn: tuple[int, int, int], +): + label = "Quant Matmul" + + sub_label = ( + "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, MKN=({})".format( + model, num_experts, topk, per_act_token, per_out_ch, mkn + ) + ) + + print(f"Testing: {sub_label}") + + (m, k, n) = mkn + + dtype = torch.half + + a = torch.randn((m, k), device="cuda", dtype=dtype) / 10 + w1 = torch.randn((num_experts, 2 * n, k), device="cuda", dtype=dtype) / 10 + w2 = torch.randn((num_experts, k, n), device="cuda", dtype=dtype) / 10 + + _, a_scale = ops.scaled_fp8_quant(a) + + w1_q = torch.empty( + (num_experts, 2 * n, k), device="cuda", dtype=torch.float8_e4m3fn + ) + w2_q = torch.empty((num_experts, k, n), device="cuda", dtype=torch.float8_e4m3fn) + w1_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32) + w2_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32) + + for expert in range(num_experts): + w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(w1[expert]) + w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(w2[expert]) + + score = torch.randn((m, num_experts), device="cuda", dtype=dtype) + + topk_weights, topk_ids, token_expert_indices = fused_topk( + a, score, topk, renormalize=False + ) + + def run_triton_moe( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_scale: torch.Tensor, + num_repeats: int, + ): + for _ in range(num_repeats): + fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_scale, + ) + + def run_cutlass_moe( + a: torch.Tensor, + a_scale: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + per_act_token: bool, + num_repeats: int, + ): + for _ in range(num_repeats): + cutlass_moe_fp8( + a, + w1, + w2, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + per_act_token, + a1_scale=None, + ) + + def run_cutlass_from_graph( + a: torch.Tensor, + a_scale: torch.Tensor, + w1_q: torch.Tensor, + w2_q: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return cutlass_moe_fp8( + a, + w1_q, + w2_q, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + per_act_token, + a1_scale=None, + ) + + def run_triton_from_graph( + a: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + a_scale: torch.Tensor, + ): + with set_current_vllm_config( + VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) + ): + return fused_experts( + a, + w1, + w2, + topk_weights, + topk_ids, + use_fp8_w8a8=True, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a_scale, + ) + + def replay_graph(graph, num_repeats): + for _ in range(num_repeats): + graph.replay() + torch.cuda.synchronize() + + cutlass_stream = torch.cuda.Stream() + cutlass_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(cutlass_graph, stream=cutlass_stream): + run_cutlass_from_graph( + a, + a_scale, + w1_q, + w2_q, + w1_scale, + w2_scale, + topk_weights, + topk_ids, + ) + torch.cuda.synchronize() + + triton_stream = torch.cuda.Stream() + triton_graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(triton_graph, stream=triton_stream): + run_triton_from_graph( + a, + w1_q, + w2_q, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + a_scale, + ) + torch.cuda.synchronize() + + min_run_time = 5 + num_warmup = 5 + num_runs = 25 + + globals = { + # Baseline params + "w1": w1, + "w2": w2, + "score": score, + "topk": topk, + # Cutlass params + "a_scale": a_scale, + "w1_q": w1_q, + "w2_q": w2_q, + "w1_scale": w1_scale, + "w2_scale": w2_scale, + "per_act_token": per_act_token, + # cuda graph params + "cutlass_graph": cutlass_graph, + "triton_graph": triton_graph, + # Gen params + "a": a, + "topk_weights": topk_weights, + "topk_ids": topk_ids, + "num_runs": num_runs, + # Kernels + "run_triton_moe": run_triton_moe, + "run_cutlass_moe": run_cutlass_moe, + "replay_graph": replay_graph, + } + + # Warmup + run_triton_moe( + a, + w1_q, + w2_q, + topk_weights, + topk_ids, + w1_scale, + w2_scale, + a_scale, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_triton_moe(a, w1_q, w2_q, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(triton_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(triton_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="triton_moe_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + run_cutlass_moe( + a, + a_scale, + w1_q, + w2_q, + w1_scale, + w2_scale, + topk_weights, + topk_ids, + per_act_token, + num_warmup, + ) + + results.append( + benchmark.Timer( + stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, per_act_token, num_runs)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="grouped_gemm_moe", + ).blocked_autorange(min_run_time=min_run_time) + ) + + # Warmup + replay_graph(cutlass_graph, num_warmup) + + results.append( + benchmark.Timer( + stmt="replay_graph(cutlass_graph, num_runs)", + globals=globals, + label=label, + sub_label=sub_label, + description="grouped_gemm_moe_cuda_graphs", + ).blocked_autorange(min_run_time=min_run_time) + ) + + +def main(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + results: list[benchmark.Measurement] = [] + + for model in args.models: + for tp in args.tp_sizes: + for layer in WEIGHT_SHAPES_MOE[model]: + num_experts = layer[0] + topk = layer[1] + size_k = layer[2] + size_n = layer[3] // tp + + if len(args.limit_k) > 0 and size_k not in args.limit_k: + continue + + if len(args.limit_n) > 0 and size_n not in args.limit_n: + continue + + for per_act_token in PER_ACT_TOKEN_OPTS: + for per_out_ch in PER_OUT_CH_OPTS: + for size_m in DEFAULT_BATCH_SIZES: + mkn = (size_m, size_k, size_n) + bench_run( + results, + model, + num_experts, + topk, + per_act_token, + per_out_ch, + mkn, + ) + + compare = benchmark.Compare(results) + compare.print() + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark Marlin across specified models/shapes/batches" + ) + parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES_MOE.keys(), + ) + parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES) + parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + parser.add_argument("--limit-k", nargs="+", type=int, default=[]) + parser.add_argument("--limit-n", nargs="+", type=int, default=[]) + parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-act-token", nargs="+", type=int, default=[]) + parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[]) + + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_layernorm.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_layernorm.py new file mode 100644 index 0000000000000000000000000000000000000000..69978ec6b23e94ce21d8f3dfb633317b3803e6ef --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_layernorm.py @@ -0,0 +1,93 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import time + +import torch + +from vllm.model_executor.layers.layernorm import RMSNorm +from vllm.platforms import current_platform +from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser + + +@torch.inference_mode() +def main( + num_tokens: int, + hidden_size: int, + add_residual: bool, + dtype: torch.dtype, + seed: int = 0, + do_profile: bool = False, + num_warmup_iters: int = 5, + num_iters: int = 100, +) -> None: + current_platform.seed_everything(seed) + torch.set_default_device("cuda") + + layer = RMSNorm(hidden_size).to(dtype=dtype) + layer.weight.data.normal_(mean=1.0, std=0.1) + scale = 1 / (2 * hidden_size) + x = torch.randn(num_tokens, hidden_size, dtype=dtype) + x *= scale + residual = torch.randn_like(x) * scale if add_residual else None + + def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: + torch.cuda.synchronize() + if profile: + torch.cuda.cudart().cudaProfilerStart() + start_time = time.perf_counter() + + for _ in range(num_iters): + layer(x, residual) + torch.cuda.synchronize() + + end_time = time.perf_counter() + if profile: + torch.cuda.cudart().cudaProfilerStop() + return (end_time - start_time) / num_iters + + # Warmup. + print("Warming up...") + run_benchmark = run_cuda_benchmark + run_benchmark(num_iters=num_warmup_iters, profile=False) + + # Benchmark. + if do_profile: + latency = run_benchmark(num_iters=1, profile=True) + else: + latency = run_benchmark(num_iters=num_iters, profile=False) + print(f"Kernel running time: {latency * 1000000:.3f} us") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser(description="Benchmark the layernorm kernel.") + parser.add_argument("--num-tokens", type=int, default=4096) + parser.add_argument("--hidden-size", type=int, default=8192) + parser.add_argument("--add-residual", action="store_true") + parser.add_argument( + "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half" + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--profile", action="store_true") + parser.add_argument("--num-warmup-iters", type=int, default=5) + parser.add_argument( + "--num-iters", + type=int, + default=100, + help="Number of benchmark iterations. " + "If --profile is set, this number is ignored", + ) + + args = parser.parse_args() + print(args) + + main( + num_tokens=args.num_tokens, + hidden_size=args.hidden_size, + add_residual=args.add_residual, + dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype], + seed=args.seed, + do_profile=args.profile, + num_warmup_iters=args.num_warmup_iters, + num_iters=args.num_iters, + ) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_lora.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..3d38d4b3534e8993abc8aff85712230e7262a740 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_lora.py @@ -0,0 +1,1065 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import json +import pickle +import time +from dataclasses import dataclass +from enum import Enum, auto +from itertools import product +from pathlib import Path +from typing import Any, Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from utils import ArgPool, Bench, CudaGraphBenchParams +from weight_shapes import WEIGHT_SHAPES + +from vllm.triton_utils import HAS_TRITON + +if HAS_TRITON: + from vllm.lora.ops.triton_ops import LoRAKernelMeta, lora_expand, lora_shrink + from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT + +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) +DEFAULT_TP_SIZES = [1] +DEFAULT_BATCH_SIZES = [ + 1, + 16, + 32, + 64, + 128, + 192, + 256, + 320, + 384, + 448, + 512, + 640, + 768, + 896, + 1024, + 2048, + 3072, + 4096, + 5120, + 6144, + 7168, + 8192, +] +DEFAULT_HIDDEN_SIZES = [1024, 2048, 4096, 8192, 16384] +DEFAULT_LORA_RANKS = [16] +DEFAULT_NUM_LORAS = [1, 2, 3, 4] +DEFAULT_SORT_BY_LORA_IDS = [False, True] +DEFAULT_SEQ_LENGTHS = [1] +DEFAULT_EXPAND_FN_ADD_INPUTS = [True, False] + + +# Utilities +def dtype_to_str(dtype: torch.dtype): + if dtype == torch.float16: + return "f16" + if dtype == torch.bfloat16: + return "bf16" + if dtype == torch.float32: + return "f32" + raise ValueError(f"Unsupported dtype {dtype}") + + +def make_rand_lora_weight_tensor( + k: int, n: int, num_loras: int, dtype: torch.dtype, device: str = "cuda" +) -> torch.Tensor: + # LoRA weights column major + return torch.rand((num_loras, n, k), dtype=dtype).to(device) + + +def make_rand_tensors( + a_shape: tuple[int], + b_shape: tuple[int], + c_shape: tuple[int], + a_dtype: torch.dtype, + b_dtype: torch.dtype, + c_dtype: torch.dtype, + num_slices: int, + device: str = "cuda", +) -> tuple[torch.Tensor, list[torch.Tensor], torch.Tensor]: + """ + Make LoRA input/output matrices. + """ + A = torch.rand(a_shape, dtype=a_dtype).to(device) + + # LoRA weights column major + Bs = [torch.rand(b_shape, dtype=b_dtype).to(device) for _ in range(num_slices)] + + C = torch.zeros(c_shape, dtype=c_dtype).to(device) + return A, Bs, C + + +def make_prompt_lora_mapping( + num_prompts: int, num_active_loras: int, sort_by_lora_id: bool, device: str +) -> torch.Tensor: + """ + All prompts are mapped to a LoRA ID in range [0, num_active_loras). + where 0 refers to first lora, 1 refers to second lora and so on. + """ + assert num_active_loras > 0 + + if not sort_by_lora_id: + return torch.randint(0, num_active_loras, (num_prompts,), dtype=torch.long) + + # Divide LoRAs equally and in order. + part_size = num_prompts // num_active_loras + part_size = max(part_size, 1) + + lora_id = 0 + prompt_lora_mapping = [] + while len(prompt_lora_mapping) < num_prompts: + prompt_lora_mapping.extend([lora_id] * part_size) + lora_id = lora_id + 1 if lora_id + 1 < num_active_loras else lora_id + return torch.tensor( + prompt_lora_mapping[:num_prompts], dtype=torch.long, device=device + ) + + +def make_token_lora_mapping( + num_tokens: int, + num_prompts: int, + prompt_lora_mapping: torch.Tensor, + seq_len_tensor: torch.Tensor, + device: str, +): + """ + Make token_lora_mapping from prompt_lora_mapping and seq_lens_tensor + """ + assert prompt_lora_mapping.shape[0] == num_prompts + + # token to lora index mapping + token_lora_mapping = [0] * num_tokens + current_offset = 0 + for b_id in range(num_prompts): + lora_index = prompt_lora_mapping[b_id].item() + s = current_offset + e = s + seq_len_tensor[b_id].item() + token_lora_mapping[s:e] = [lora_index] * (e - s) + current_offset += seq_len_tensor[b_id].item() + + return torch.tensor(token_lora_mapping, dtype=torch.long, device=device) + + +def ref_group_gemm( + ref_out: torch.Tensor, + input: torch.Tensor, + lora_weights: list[torch.Tensor], + seq_lens_cpu: torch.Tensor, + prompt_lora_mapping_cpu: torch.Tensor, + scaling: float, + add_inputs: Optional[bool], +): + """ + Torch group gemm reference implementation to test correctness of + benchmarking operations. + """ + batches = seq_lens_cpu.size(0) + out_list = [] + current_offset = 0 + for lora_index, b_length in zip(range(batches), seq_lens_cpu): + x = input[current_offset : b_length + current_offset, :] + current_offset += b_length + w = lora_weights[prompt_lora_mapping_cpu[lora_index]] + result = torch.nn.functional.linear(x, w) + result *= scaling + out_list.append(result) + + cat_result = torch.cat(out_list, dim=0) + + if add_inputs: + ref_out += cat_result + else: + ref_out.copy_(cat_result) + + +class OpType(Enum): + """ + LoRA Ops to benchmark and its properties. + """ + + LORA_SHRINK = auto() + LORA_EXPAND = auto() + + @staticmethod + def from_str(s: str) -> "OpType": + if s.lower() == "lora_shrink": + return OpType.LORA_SHRINK + if s.lower() == "lora_expand": + return OpType.LORA_EXPAND + raise ValueError(f"Unrecognized str {s} to convert to OpType") + + def is_shrink_fn(self) -> bool: + return self in [OpType.LORA_SHRINK] + + def is_expand_fn(self) -> bool: + return self in [OpType.LORA_EXPAND] + + def num_slices(self) -> list[int]: + return [1, 2, 3] + + def mkn( + self, batch_size: int, seq_length: int, hidden_size: int, lora_rank: int + ) -> tuple[int, int, int]: + num_tokens = batch_size * seq_length + if self.is_shrink_fn(): + m = num_tokens + k = hidden_size + n = lora_rank + else: + assert self.is_expand_fn() + m = num_tokens + k = lora_rank + n = hidden_size + return m, k, n + + def matmul_dtypes( + self, op_dtype: torch.dtype + ) -> tuple[torch.dtype, torch.dtype, torch.dtype]: + """ + return a type, b type and c type for A x B = C + """ + if self.is_shrink_fn(): + return op_dtype, op_dtype, torch.float32 + else: + assert self.is_expand_fn() + return torch.float32, op_dtype, op_dtype + + def matmul_shapes( + self, + batch_size: int, + seq_length: int, + hidden_size: int, + lora_rank: int, + num_loras: int, + num_slices: int, + ) -> tuple[tuple[int], tuple[int], tuple[int]]: + """ + Given num_slices, return the shapes of the A, B, and C matrices + in A x B = C, for the op_type + """ + m, k, n = self.mkn(batch_size, seq_length, hidden_size, lora_rank) + + b_shape = (num_loras, n, k) # col-major + if self in [OpType.LORA_SHRINK]: + # LoRA shrink kernels support num_slices inherently in the kernel. + return ((m, k), b_shape, (num_slices, m, n)) + if self in [OpType.LORA_EXPAND]: + # LoRA expand kernels support num_slices inherently in the kernel + return ((num_slices, m, k), b_shape, (m, n * num_slices)) + raise ValueError(f"Unrecognized op_type {self}") + + def bench_fn(self) -> Callable: + if self == OpType.LORA_SHRINK: + return lora_shrink + if self == OpType.LORA_EXPAND: + return lora_expand + + raise ValueError(f"Unrecognized optype {self}") + + def run_ref_group_gemm( + self, + output: torch.Tensor, + input: torch.Tensor, + lora_weights: list[torch.Tensor], + **kwargs, + ) -> Callable: + """Each benchmark operation expects the input, lora_weights and outputs + in a slightly different format. Refer to self.matmul_shapes(). + run_ref_group_gemm accounts for those differences in executing a + reference group gemm for correctness testing. + """ + w_dtype = lora_weights[0].dtype + num_slices = len(lora_weights) + if self in [OpType.LORA_SHRINK]: + for slice_idx in range(num_slices): + ref_group_gemm( + ref_out=output[slice_idx, :], + input=input, + lora_weights=lora_weights[slice_idx], + **kwargs, + ) + elif self in [OpType.LORA_EXPAND]: + hidden_size = lora_weights[0].shape[1] + for slice_idx in range(num_slices): + slice_offset = slice_idx * hidden_size + ref_group_gemm( + ref_out=output[:, slice_offset : slice_offset + hidden_size], + input=input[slice_idx].clone().to(dtype=w_dtype), + lora_weights=lora_weights[slice_idx], + **kwargs, + ) + else: + raise ValueError(f"Unrecognized optype {self}") + + +@dataclass +class BenchmarkContext: + """ + LoRA benchmark context + """ + + batch_size: int + hidden_size: int + num_loras: int + num_active_loras: int + lora_rank: int + sort_by_lora_id: bool + dtype: torch.dtype + seq_length: Optional[int] = None + num_slices: Optional[int] = None # num_slices for slice based ops + + def with_seq_length(self, seq_length: int) -> "BenchmarkContext": + ctx = copy.copy(self) + ctx.seq_length = seq_length + return ctx + + def with_num_slices(self, num_slices: int) -> "BenchmarkContext": + ctx = copy.copy(self) + ctx.num_slices = num_slices + return ctx + + def bench_label(self) -> str: + return f"lora-{self.dtype}" + + def bench_sublabel(self, op_type: OpType) -> str: + m, k, n = op_type.mkn( + self.batch_size, self.seq_length, self.hidden_size, self.lora_rank + ) + desc = { + "bs": self.batch_size, + "sl": self.seq_length, + "m": m, + "k": k, + "n": n, + "num_loras": self.num_loras, + "sort_by_lora": self.sort_by_lora_id, + "num_slices": self.num_slices, + } + return json.dumps(desc) + + +@dataclass +class BenchmarkTensors: + """ + Input/Output tensors used for benchmarks + """ + + # matmul tensors + input: torch.Tensor + lora_weights_lst: list[torch.Tensor] + output: torch.Tensor + # LoRA kernel metadata + lora_kernel_meta: LoRAKernelMeta + # Metadata tensors used in testing correctness + seq_lens: torch.Tensor + prompt_lora_mapping: torch.Tensor + + def io_types(self) -> str: + return ( + f"{dtype_to_str(self.input.dtype)}x" + f"{dtype_to_str(self.lora_weights_lst[0].dtype)}=>" + f"{dtype_to_str(self.output.dtype)}" + ) + + @staticmethod + def make( + ctx: BenchmarkContext, op_type: OpType, device: str = "cuda" + ) -> "BenchmarkTensors": + # Make input / output matmul tensors. + a_shape, b_shape, c_shape = op_type.matmul_shapes( + ctx.batch_size, + ctx.seq_length, + ctx.hidden_size, + ctx.lora_rank, + ctx.num_loras, + ctx.num_slices, + ) + a_type, b_type, c_type = op_type.matmul_dtypes(ctx.dtype) + input_tensor, lora_weights, output_tensor = make_rand_tensors( + a_shape, b_shape, c_shape, a_type, b_type, c_type, num_slices=ctx.num_slices + ) + + # Make metadata tensors. + # Keep the metadata tensors in the CPU for further processing if needed. + # The tensors get moved to the GPU before benchmarking. + assert ctx.num_active_loras <= ctx.num_loras + total_tokens = ctx.batch_size * ctx.seq_length + + # Make metadata tensors involved in correctness testing. + # Prepare seq lens tensor + seq_len_tensor = torch.randint( + ctx.seq_length, ctx.seq_length + 1, (ctx.batch_size,) + ) + assert total_tokens == seq_len_tensor.sum() + # Prepare prompt lora indices tensor + prompt_lora_indices_tensor = make_prompt_lora_mapping( + ctx.batch_size, ctx.num_active_loras, ctx.sort_by_lora_id, "cpu" + ) + + # Make LoRAKernelMeta + token_lora_indices_tensor = make_token_lora_mapping( + total_tokens, + ctx.batch_size, + prompt_lora_indices_tensor, + seq_len_tensor, + "cpu", + ) + lora_kernel_meta = LoRAKernelMeta.make( + max_loras=ctx.num_loras, + max_num_tokens=token_lora_indices_tensor.size(0), + device="cpu", + ) + lora_kernel_meta.prepare_tensors(token_lora_mapping=token_lora_indices_tensor) + + return BenchmarkTensors( + input_tensor, + lora_weights, + output_tensor, + lora_kernel_meta, + seq_len_tensor, + prompt_lora_indices_tensor, + ) + + def sanity_check(self) -> None: + """ + Fails asserts when non-conformality is detected. + """ + num_tokens = self.input.shape[-2] + # check metadata tensors + assert torch.sum(self.seq_lens) == num_tokens + num_seqs = self.seq_lens.shape[0] + # assert self.seq_start_loc.shape[0] == num_seqs + assert self.prompt_lora_mapping.shape[0] == num_seqs + assert self.lora_kernel_meta.token_lora_mapping.shape[0] == num_tokens + + def to_device(self, device: str): + """ + Transfer tensors to device if the tensors aren't already on the device + """ + + def to_device(tensor: torch.Tensor): + if tensor.device != device: + tensor = tensor.to(device=device) + return tensor + + self.input = to_device(self.input) + self.output = to_device(self.output) + self.seq_lens = to_device(self.seq_lens) + self.prompt_lora_mapping = to_device(self.prompt_lora_mapping) + for i in range(len(self.lora_weights_lst)): + self.lora_weights_lst[i] = to_device(self.lora_weights_lst[i]) + + # LoRA meta + for field_name in LoRAKernelMeta.__dataclass_fields__: + field = getattr(self.lora_kernel_meta, field_name) + assert isinstance(field, torch.Tensor) + setattr(self.lora_kernel_meta, field_name, to_device(field)) + + def metadata(self) -> tuple[int, int, int]: + """ + Return num_seqs, num_tokens and max_seq_len + """ + num_seqs = self.seq_lens.shape[0] + num_tokens = self.lora_kernel_meta.token_lora_mapping.shape[0] + max_seq_len = torch.max(self.seq_lens).item() + num_slices = len(self.lora_weights_lst) + return num_seqs, num_tokens, max_seq_len, num_slices + + def as_lora_shrink_kwargs(self) -> dict[str, Any]: + self.sanity_check() + self.to_device(self.input.device) + + _, num_tokens, _, num_slices = self.metadata() + + # Sanity check matrix shapes. + i_shape, lw_shape, o_shape = ( + self.input.shape, + self.lora_weights_lst[0].shape, + self.output.shape, + ) + # Expected input shape [num_tokens, hidden_size] + assert len(i_shape) == 2 + assert i_shape[0] == num_tokens + hidden_size = i_shape[1] + # Expected lora weight shape [num_loras, lora_rank, hidden_size] + assert len(lw_shape) == 3 + assert lw_shape[2] == hidden_size + lora_rank = lw_shape[1] + # Expected output shape [num_slices, num_tokens, lora_rank] + assert len(o_shape) == 3 + assert o_shape == (num_slices, num_tokens, lora_rank) + + return { + "inputs": self.input, + "lora_a_weights": self.lora_weights_lst, + "output_tensor": self.output, + "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, + "token_indices_sorted_by_lora_ids": ( + self.lora_kernel_meta.token_indices_sorted_by_lora_ids + ), + "num_tokens_per_lora": self.lora_kernel_meta.num_tokens_per_lora, + "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc, + "lora_ids": self.lora_kernel_meta.active_lora_ids, + "scaling": 1.0, + } + + def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]: + self.sanity_check() + self.to_device(self.input.device) + + _, num_tokens, _, num_slices = self.metadata() + + # Sanity check matrix shapes. + i_shape, lw_shape, o_shape = ( + self.input.shape, + self.lora_weights_lst[0].shape, + self.output.shape, + ) + # Expected input shape : [num_slices, num_tokens, lora_rank] + assert len(i_shape) == 3 + assert i_shape[0] == num_slices + assert i_shape[1] == num_tokens + lora_rank = i_shape[2] + # Expected lora weight shape : [num_lora, hidden_size, lora_rank] + assert len(lw_shape) == 3 + assert lw_shape[2] == lora_rank + hidden_size = lw_shape[1] + # Expected output shape : [num_tokens, hidden_size * num_slices] + assert len(o_shape) == 2 + assert o_shape == (num_tokens, hidden_size * num_slices) + + return { + "inputs": self.input, + "lora_b_weights": self.lora_weights_lst, + "output_tensor": self.output, + "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, + "token_indices_sorted_by_lora_ids": ( + self.lora_kernel_meta.token_indices_sorted_by_lora_ids + ), + "num_tokens_per_lora": self.lora_kernel_meta.num_tokens_per_lora, + "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc, + "lora_ids": self.lora_kernel_meta.active_lora_ids, + "offset_start": 0, + "add_inputs": add_inputs, + } + + def bench_fn_kwargs( + self, op_type: OpType, add_inputs: Optional[bool] = None + ) -> dict[str, Any]: + if op_type.is_shrink_fn(): + assert add_inputs is None + else: + assert add_inputs is not None + + if op_type == OpType.LORA_SHRINK: + return self.as_lora_shrink_kwargs() + if op_type == OpType.LORA_EXPAND: + return self.as_lora_expand_kwargs(add_inputs) + raise ValueError(f"Unrecognized optype {self}") + + def test_correctness( + self, op_type: OpType, expand_fn_add_inputs: Optional[bool] + ) -> bool: + """ + Test correctness of op_type implementation against a grouped gemm + reference implementation. + """ + seq_lens_cpu = self.seq_lens.to(device="cpu") + prompt_lora_mapping_cpu = self.prompt_lora_mapping.to(device="cpu") + ref_output = self.output.clone() + + self.output.zero_() + op_type.bench_fn()(**self.bench_fn_kwargs(op_type, expand_fn_add_inputs)) + + op_type.run_ref_group_gemm( + ref_output, + self.input, + self.lora_weights_lst, + seq_lens_cpu=seq_lens_cpu, + prompt_lora_mapping_cpu=prompt_lora_mapping_cpu, + scaling=1.0, + add_inputs=expand_fn_add_inputs, + ) + + rtol, atol = { + torch.float16: (6e-2, 6e-2), + torch.bfloat16: (6e-2, 6e-2), + torch.float32: (1e-2, 1e-2), + }[self.output.dtype] + + return torch.allclose(ref_output, self.output, rtol=rtol, atol=atol) + + +def bench_optype( + ctx: BenchmarkContext, + arg_pool_size: int, + op_type: OpType, + cuda_graph_nops: Optional[int] = None, + expand_fn_add_inputs: Optional[bool] = None, + test_correctness: bool = False, +) -> TMeasurement: + assert arg_pool_size >= 1 + if op_type.is_shrink_fn(): + assert expand_fn_add_inputs is None + else: + assert expand_fn_add_inputs is not None + + # BenchmarkContext -> BenchmarkTensors + bench_tensors: list[BenchmarkTensors] = [ + BenchmarkTensors.make(ctx, op_type) for _ in range(arg_pool_size) + ] + for bt in bench_tensors: + bt.sanity_check() + + # Test correctness of our implementation. + if test_correctness: + assert all( + [bt.test_correctness(op_type, expand_fn_add_inputs) for bt in bench_tensors] + ) + + # BenchmarkTensors -> dict (kwargs) + kwargs_list = [ + bt.bench_fn_kwargs(op_type, add_inputs=expand_fn_add_inputs) + for bt in bench_tensors + ] + + # Clear LoRA optimization hash-maps. + _LORA_A_PTR_DICT.clear() + _LORA_B_PTR_DICT.clear() + # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are setup + for kwargs in kwargs_list: + op_type.bench_fn()(**kwargs) + torch.cuda.synchronize() + + # Merge into a single kwargs and qualify arguments as ArgPool + kwargs = {k: ArgPool([]) for k in kwargs_list[0]} + for _kwargs in kwargs_list: + for k, v in _kwargs.items(): + kwargs[k].values.append(v) + + describe_args = ( + f"add_inputs={expand_fn_add_inputs}" if expand_fn_add_inputs is not None else "" + ) + description = f"{op_type.name}({describe_args}) ({bench_tensors[0].io_types()})" + + cuda_graph_params = None + if cuda_graph_nops: + cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) + timer = None + with Bench( + cuda_graph_params, + ctx.bench_label(), + ctx.bench_sublabel(op_type), + description, + op_type.bench_fn(), + **kwargs, + ) as bench: + timer = bench.run() + return timer + + +def bench_torch_mm( + ctx: BenchmarkContext, + arg_pool_size: int, + op_type: OpType, + cuda_graph_nops: Optional[int] = None, +) -> TMeasurement: + """ + Benchmark basic torch.mm as a roofline. + + When all the input tokens have the same LoRA ID, the LoRA kernels are just + a matmul. This torch.mm benchmark serves as a roofline for that case. + + input op_type is used in determining the m, k, n dimensions for the matmul. + """ + + batch_size, hidden_size, lora_rank, seq_length, dtype = ( + ctx.batch_size, + ctx.hidden_size, + ctx.lora_rank, + ctx.seq_length, + ctx.dtype, + ) + + m, k, n = op_type.mkn(batch_size, seq_length, hidden_size, lora_rank) + # For a fairer comparison. + n = n * ctx.num_slices + + # Get matmul input and output tensors for A x B = C + As, Bs, Cs = [], [], [] + for _ in range(arg_pool_size): + As.append(torch.rand((m, k), dtype=dtype).to("cuda")) + Bs.append(torch.rand((n, k), dtype=dtype).to("cuda").t()) + Cs.append(torch.rand((m, n), dtype=dtype).to("cuda")) + + # Make torch.mm kwargs + mm_kwargs = {"input": ArgPool(As), "mat2": ArgPool(Bs), "out": ArgPool(Cs)} + + description = ( + f"single-lora roofline using torch.mm ({dtype_to_str(dtype)}" + f"x{dtype_to_str(dtype)}" + f"=>{dtype_to_str(dtype)})" + ) + cuda_graph_params = None + if cuda_graph_nops: + cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) + with Bench( + cuda_graph_params, + ctx.bench_label(), + ctx.bench_sublabel(op_type), + description, + torch.mm, + **mm_kwargs, + ) as bench: + return bench.run() + + +# runner +def use_cuda_graph_recommendation() -> str: + return """ + Triton kernels have a significant launch overhead with + launched directly via python. This overhead is more noticeable + for small the problem sizes. For these cases, it is recommended + to use the script with `--cuda-graph-nops N` to benchmark N + consecutive invocations of the benchmarking operations from + inside a CUDA Graph. Note that the returned measurement is for N + invocations of the operation. + """ + + +def print_timers(timers: list[TMeasurement], args: Optional[argparse.Namespace] = None): + compare = TBenchmark.Compare(timers) + compare.print() + + if args and args.cuda_graph_nops: + print( + f"Note : The timings reported above is for {args.cuda_graph_nops} " + "consecutive invocations of the benchmarking functions. " + f"Please divide by {args.cuda_graph_nops} for single invocation " + "timings." + ) + + print( + "Note on Comparison with torch.mm : The torch.mm numbers are " + "benchmark numbers of a simple matmul emulating the single lora " + "case. It is provided as a roofline for comparing our LoRA Kernel " + "implementations. It is expected that the LoRA kernels will be " + "slower than torch.mm in cases where num_loras is big. But for " + "small num_loras the goal should be to match the torch.mm numbers." + ) + + +def run(args: argparse.Namespace, bench_ctxs: list[BenchmarkContext]): + if args.cuda_graph_nops is not None: + assert args.cuda_graph_nops > 0 + print(f"Benchmarking {args.cuda_graph_nops} invocations inside a CUDA Graph") + else: + print(f"CUDA Graphs not enabled.\n{use_cuda_graph_recommendation()}") + + timers = [] + for bench_ctx in bench_ctxs: + for seq_len in args.seq_lengths: + bench_ops: list[OpType] = args.op_types + seq_len_timers = [] + for bench_op in bench_ops: + for num_slices in bench_op.num_slices(): + _ctx = bench_ctx.with_seq_length(seq_len).with_num_slices( + num_slices + ) + # Benchmark torch.mm as a roofline + seq_len_timers.append( + bench_torch_mm( + _ctx, args.arg_pool_size, bench_op, args.cuda_graph_nops + ) + ) + + # Benchmark bench_op + expand_fn_add_inputs = ( + [None] if bench_op.is_shrink_fn() else args.expand_fn_add_inputs + ) + for add_input_arg in expand_fn_add_inputs: + seq_len_timers.append( + bench_optype( + _ctx, + args.arg_pool_size, + bench_op, + args.cuda_graph_nops, + add_input_arg, + args.test_correctness, + ) + ) + + print_timers(seq_len_timers) + timers.extend(seq_len_timers) + + # Result stdout dump + print("== All Results ====") + print_timers(timers, args) + + if args.output_directory: + # Result file dump + od = Path(args.output_directory) + if not od.exists(): + od.mkdir() + + timestamp = int(time.time()) + pkl_file = od / f"lora_bench-{timestamp}.pkl" + print(f"Writing benchmarks to {pkl_file}") + with open(pkl_file, "wb") as f: + pickle.dump(timers, f) + + +def as_benchmark_contexts( + hidden_sizes: list[int], lora_ranks: list[int], args: argparse.Namespace +) -> list[BenchmarkContext]: + ctxs: list[BenchmarkContext] = [] + for batch_size, hidden_size, lora_rank, num_loras, sort_by_lora_id in product( # noqa + args.batch_sizes, + list(hidden_sizes), + lora_ranks, + args.num_loras, + args.sort_by_lora_id, + ): + ctxs.append( + BenchmarkContext( + batch_size=batch_size, + hidden_size=hidden_size, + lora_rank=lora_rank, + num_loras=num_loras, + num_active_loras=args.num_active_loras + if args.num_active_loras + else num_loras, + # To be filled based on the OpType to benchmark + seq_length=None, + sort_by_lora_id=sort_by_lora_id, + dtype=args.dtype, + # To be filled based on the OpType to benchmark + num_slices=None, + ) + ) + + return ctxs + + +def run_list_bench(args: argparse.Namespace): + print(args) + + print( + "List bench :\n" + f" Hidden Sizes {args.hidden_sizes}" + f" LoRA Ranks {args.lora_ranks}" + ) + + # Get all benchmarking contexts + bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=args.hidden_sizes, lora_ranks=args.lora_ranks, args=args + ) + + run(args, bench_contexts) + + +def run_range_bench(args: argparse.Namespace): + print(args) + + hidden_sizes = list( + range( + args.hidden_sizes_start, + args.hidden_sizes_end + 1, + args.hidden_sizes_increment, + ) + ) + lora_ranks = list( + range(args.lora_ranks_start, args.lora_ranks_end + 1, args.lora_ranks_increment) + ) + + print(f"Range bench :\n Hidden Sizes {hidden_sizes} LoRA Ranks {lora_ranks}") + + # Get all benchmarking contexts + bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=hidden_sizes, lora_ranks=lora_ranks, args=args + ) + + run(args, bench_contexts) + + +def run_model_bench(args: argparse.Namespace): + print(args) + + def hidden_sizes_from_model(model: str, tp_size: int) -> set[int]: + hidden_sizes = set() + for KN, tp_split_dim in WEIGHT_SHAPES[model]: + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + hidden_sizes.add(KN[1]) + return hidden_sizes + + # Get all hidden sizes + hidden_sizes: set[int] = set() + for model_name, tp_size in product(args.models, args.tp_sizes): + hidden_sizes = hidden_sizes.union(hidden_sizes_from_model(model_name, tp_size)) + + print(f"Model bench :\n Hidden Sizes {hidden_sizes} LoRA Ranks {args.lora_ranks}") + + # Get all benchmarking contexts + bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( + hidden_sizes=hidden_sizes, lora_ranks=args.lora_ranks, args=args + ) + + run(args, bench_contexts) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "torch.float16": + return torch.float16 + if dt == "torch.bfloat16": + return torch.bfloat16 + raise ValueError("unsupported dtype") + + def get_bool(s: str) -> bool: + return s.lower() in ["true", "1"] + + def add_common_command_args(p: argparse.ArgumentParser): + p.add_argument( + "--dtype", + type=to_torch_dtype, + required=True, + help="Available options are ['torch.float16', 'torch.bfloat16']", + ) + + p.add_argument( + "--arg-pool-size", + type=int, + default=32, + help="Run profiles with a pool of input/output/meta tensors instead" + "of simply reusing the same tensors for all runs. A bigger arg-pool" + "mitigates hardware caching effects during benchmarking.", + ) + + p.add_argument( + "--cuda-graph-nops", + type=int, + help=( + "when set profiling is done using cudagraph, " + "with the given number of operations in a graph." + "Note that the measurement returned is the time " + "taken for N consecutive executions of the benchmarking " + "functions, where N is the value of this argument." + ), + ) + p.add_argument("--num-loras", nargs="+", type=int, default=DEFAULT_NUM_LORAS) + p.add_argument( + "--num-active-loras", + type=int, + default=None, + help="Active LoRAs. When None, all LoRAs are active", + ) + p.add_argument( + "--sort-by-lora-id", + nargs="+", + type=get_bool, + default=DEFAULT_SORT_BY_LORA_IDS, + ) + p.add_argument( + "--op-types", nargs="+", type=OpType.from_str, default=list(OpType) + ) + p.add_argument( + "--seq-lengths", nargs="+", type=int, default=DEFAULT_SEQ_LENGTHS + ) + p.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + p.add_argument( + "--expand-fn-add-inputs", + nargs="+", + type=get_bool, + default=DEFAULT_EXPAND_FN_ADD_INPUTS, + ) + p.add_argument( + "-o", + "--output-directory", + type=str, + help=( + "Output directory to store a the list of benchmarking" + "TMeasurement objects as a pickle file" + ), + ) + + p.add_argument( + "--test-correctness", + action="store_true", + help=( + "When enabled, the benchmarking functions are tested" + "for correctness before the actual benchmarking" + ), + ) + + parser = FlexibleArgumentParser( + description=f""" +Benchmark LoRA kernels: + {use_cuda_graph_recommendation()} + + list_bench example: + python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 + + model_bench example: + python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 + + range_bench example: + python3 benchmarks/kernels/benchmark_lora.py range_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + + subparsers = parser.add_subparsers(dest="cmd", required=True) + + list_parser = subparsers.add_parser("list_bench") + list_parser.add_argument( + "--hidden-sizes", nargs="+", type=int, default=DEFAULT_HIDDEN_SIZES + ) + list_parser.add_argument( + "--lora-ranks", nargs="+", type=int, default=DEFAULT_LORA_RANKS + ) + add_common_command_args(list_parser) + list_parser.set_defaults(func=run_list_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument("--hidden-sizes-start", type=int, required=True) + range_parser.add_argument("--hidden-sizes-end", type=int, required=True) + range_parser.add_argument("--hidden-sizes-increment", type=int, required=True) + range_parser.add_argument("--lora-ranks-start", type=int, required=True) + range_parser.add_argument("--lora-ranks-end", type=int, required=True) + range_parser.add_argument("--lora-ranks-increment", type=int, required=True) + add_common_command_args(range_parser) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--lora-ranks", nargs="+", type=int, default=DEFAULT_LORA_RANKS + ) + add_common_command_args(model_parser) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + args.func(args) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_machete.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_machete.py new file mode 100644 index 0000000000000000000000000000000000000000..f73d0511e01fc3851847278f8604a2baed8c8f33 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_machete.py @@ -0,0 +1,732 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import copy +import itertools +import math +import os +import pickle as pkl +import time +from collections.abc import Iterable +from dataclasses import dataclass +from itertools import product +from typing import Callable, Optional + +import pandas as pd +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement +from weight_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.marlin_utils import ( + GPTQ_MARLIN_MAX_PARALLEL, + GPTQ_MARLIN_MIN_THREAD_N, + marlin_permute_scales, + marlin_zero_points, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_test import ( + MarlinWorkspace, +) +from vllm.model_executor.layers.quantization.utils.quant_utils import ( + pack_rows, + quantize_weights, +) +from vllm.scalar_type import ScalarType, scalar_types +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = ["meta-llama/Llama-3-8b", "meta-llama/Llama-2-70b-hf"] +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024] +DEFAULT_TP_SIZES = [1] + +NVTX_PROFILE = os.environ.get("NVTX_PROFILE", False) + +if NVTX_PROFILE: + import nvtx + + +def terse_type_name(dt): + return { + torch.bfloat16: "bf16", + torch.float16: "fp16", + torch.int8: "int8", + torch.float8_e4m3fn: "fp8", + torch.float: "float", + torch.int: "int", + }[dt] + + +@dataclass +class BenchmarkTensors: + w_ref: torch.Tensor + a: torch.Tensor + + w_q: torch.Tensor + group_size: Optional[int] + wtype: ScalarType + w_g_s: torch.Tensor + w_g_zp: Optional[torch.Tensor] + w_ch_s: Optional[torch.Tensor] + w_tok_s: Optional[torch.Tensor] + + +@dataclass +class TypeConfig: + act_type: torch.dtype + weight_type: ScalarType + output_type: Optional[torch.dtype] + group_scale_type: Optional[torch.dtype] + group_zero_type: Optional[torch.dtype] + channel_scale_type: Optional[torch.dtype] + token_scale_type: Optional[torch.dtype] + + +def rand_data(shape, dtype=torch.float16, scale=1): + if dtype.is_floating_point: + return (scale * torch.rand(shape, device="cuda") - 0.3).to(dtype) + else: + return torch.randint(-15, 15, shape, dtype=dtype, device="cuda") + + +def quantize_and_pack( + atype: torch.dtype, + w: torch.Tensor, + wtype: ScalarType, + stype: Optional[torch.dtype], + group_size: Optional[int], + zero_points: bool = False, +): + assert wtype.is_integer(), "TODO: support floating point weights" + + w_ref, w_q, w_s, w_zp = quantize_weights( + w, + wtype, + group_size=group_size, + zero_points=zero_points, + # to match how the kernel applies zps + ref_zero_points_after_scales=True, + ) + + w_q = pack_rows(w_q, wtype.size_bits, *w_q.shape) + return w_ref, w_q, w_s, w_zp + + +def create_bench_tensors( + shape: tuple[int, int, int], types: TypeConfig, group_size: Optional[int] +) -> list[BenchmarkTensors]: + m, n, k = shape + + # we want to make sure that weights don't fit into L2 cache between runs so + # we construct enough weights to exceed L2 cache, which is 50mb on a H100 + # so we target total weight size > 2*50mb + num_weights = math.ceil( + 2 * 50 * 1024**2 * 8 / (k * n * types.weight_type.size_bits) + ) + + a = rand_data((m, k), types.act_type, scale=5) + + benchmark_tensors: list[BenchmarkTensors] = [] + for _ in range(num_weights): + w = rand_data((k, n), types.act_type, scale=5) + + if types.group_scale_type is not None: + w = w.to(types.group_scale_type) + if w.dtype.itemsize == 1: + w = w.to(torch.float16) + + w_ref, w_q_packed, w_s, w_zp = quantize_and_pack( + a.dtype, + w, + types.weight_type, + types.group_scale_type, + group_size, + types.group_zero_type is not None, + ) + + if not a.dtype.is_floating_point: + aiinfo = torch.iinfo(a.dtype) + w_ref = w_ref.round().clamp(aiinfo.min, aiinfo.max) + + w_ref = w_ref.to(torch.float32) + + w_ch_s = ( + None + if types.channel_scale_type is None + else rand_data((n,), types.channel_scale_type) + ) + w_tok_s = ( + None + if types.token_scale_type is None + else rand_data((m,), types.token_scale_type) + ) + + benchmark_tensors.append( + BenchmarkTensors( + w_ref=w_ref, + a=a, + w_q=w_q_packed, + wtype=types.weight_type, + w_g_s=w_s, + w_g_zp=w_zp, + group_size=group_size, + w_ch_s=w_ch_s, + w_tok_s=w_tok_s, + ) + ) + + return benchmark_tensors + + +def torch_matmul_f16_create_bench_fn(bt: BenchmarkTensors) -> Callable: + a = bt.a + w = bt.w_ref.to(bt.a.dtype) # use float reference tensor + if a.dtype not in [torch.float16, torch.bfloat16]: + a = a.to(torch.float16) + w = w.to(torch.float16) + return lambda: torch.matmul(a, w) + + +def cutlass_scaled_mm_create_bench_fn(bt: BenchmarkTensors) -> Callable: + if bt.w_ch_s is not None and bt.w_tok_s is not None: + scale_a = bt.w_tok_s.to(torch.float32) + scale_b = bt.w_ch_s.to(torch.float32) + else: + scale_a = torch.tensor(1.0, dtype=torch.float32, device=bt.a.device) + scale_b = torch.tensor(1.0, dtype=torch.float32, device=bt.a.device) + w_col_major = bt.w_ref.to(bt.a.dtype).t().contiguous().t() + return lambda: ops.cutlass_scaled_mm( + bt.a, w_col_major, scale_a, scale_b, out_dtype=torch.float16 + ) + + +def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable: + device = bt.a.device + + workspace = MarlinWorkspace( + bt.w_ref.shape[1], GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL + ) + + if bt.w_g_zp is None: + w_zp = torch.empty(0, dtype=torch.int, device=device) + else: + w_zp = marlin_zero_points( + bt.w_g_zp, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits + ) + + if bt.group_size is None: + w_s = torch.tensor([], device="cuda", dtype=torch.half) + else: + w_s = marlin_permute_scales( + bt.w_g_s, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.group_size + ) + + sort_indices = torch.empty(0, dtype=torch.int, device=device) + g_idx = torch.empty(0, dtype=torch.int, device=device) + w_q = ops.gptq_marlin_repack( + bt.w_q, sort_indices, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits + ) + + if bt.a.dtype.is_floating_point: + assert bt.w_ch_s is None + assert bt.w_tok_s is None + assert bt.group_size is not None + + fn = lambda: ops.gptq_marlin_gemm( + a=bt.a, + c=None, + b_q_weight=w_q, + b_scales=w_s, + global_scale=None, + b_zeros=w_zp, + g_idx=g_idx, + perm=sort_indices, + workspace=workspace.scratch, + b_q_type=bt.wtype, + size_m=bt.a.shape[0], + size_n=bt.w_ref.shape[1], + size_k=bt.w_ref.shape[0], + is_k_full=True, + is_zp_float=False, + ) + else: + assert bt.a.dtype == torch.int8 + assert bt.wtype == scalar_types.uint4b8 + + if bt.w_ch_s is not None: + s_ch = bt.w_ch_s.to(torch.float32) + else: + s_ch = torch.ones(bt.w_ref.shape[1], dtype=torch.float32, device=device) + + if bt.w_tok_s is not None: + s_tok = bt.w_tok_s.to(torch.float32) + else: + s_tok = torch.ones(bt.a.shape[0], dtype=torch.float32, device=device) + + fn = lambda: ops.marlin_qqq_gemm( + a=bt.a, + b_q_weight=w_q, + s_group=w_s, + s_tok=s_tok, + s_ch=s_ch, + workspace=workspace.scratch, + size_m=bt.a.shape[0], + size_n=bt.w_ref.shape[1], + size_k=bt.w_ref.shape[0], + ) + + return fn + + +def machete_create_bench_fn( + bt: BenchmarkTensors, out_type=torch.dtype, schedule=None +) -> Callable: + w_q = bt.w_q.t().contiguous().t() # make col major + w_q = ops.machete_prepack_B( + w_q, bt.a.dtype, bt.wtype, None if bt.w_g_s is None else bt.w_g_s.dtype + ) + + w_g_zp = bt.w_g_zp + if w_g_zp is not None: + w_g_zp = -1 * bt.w_g_s * (w_g_zp.to(bt.w_g_s.dtype)) + + return lambda: ops.machete_mm( + a=bt.a, + b_q=w_q, + b_type=bt.wtype, + b_group_scales=bt.w_g_s, + b_group_zeros=w_g_zp, + b_group_size=bt.group_size, + b_channel_scales=bt.w_ch_s, + a_token_scales=bt.w_tok_s, + out_type=out_type, + schedule=schedule, + ) + + +# impl + +# bench + + +def bench_fns(label: str, sub_label: str, description: str, fns: list[Callable]): + min_run_time = 1 if not NVTX_PROFILE else 0.1 + res = TBenchmark.Timer( + stmt=""" + for fn in fns: + fn() + """, + globals={"fns": fns}, + label=label, + sub_label=sub_label, + description=description, + ).blocked_autorange(min_run_time=min_run_time) + + if NVTX_PROFILE: + with ( + nvtx.annotate("mm-bench"), + nvtx.annotate(f"{label}|{sub_label}|{description}"), + ): + fns[0]() + + return res + + +_SWEEP_SCHEDULES_RESULTS: Optional[pd.DataFrame] = None +_SWEEP_SCHEDULES_RESULTS_CSV: Optional[str] = None + + +def bench( + types: TypeConfig, + group_size: int, + m: int, + k: int, + n: int, + label: str, + sub_label: str, + sweep_schedules: bool = True, +) -> list[TMeasurement]: + benchmark_tensors = create_bench_tensors((m, n, k), types, group_size) + sub_label += f", L={len(benchmark_tensors)}" + + name_type_string = f"W{types.weight_type}" + f"-A{terse_type_name(types.act_type)}" + if types.group_scale_type is not None: + name_type_string += f"-GS{terse_type_name(types.group_scale_type)}" + if types.group_zero_type is not None: + name_type_string += f"-GZ{terse_type_name(types.group_zero_type)}" + if group_size is not None: + name_type_string += f"-G{group_size}" + if types.channel_scale_type is not None: + name_type_string += f"-CS{terse_type_name(types.channel_scale_type)}" + if types.token_scale_type is not None: + name_type_string += f"-TS{terse_type_name(types.token_scale_type)}" + + timers = [] + # pytorch impl + timers.append( + bench_fns( + label, + sub_label, + "torch.matmul (fp16)", + [torch_matmul_f16_create_bench_fn(bt) for bt in benchmark_tensors], + ) + ) + + if types.act_type == torch.int8 or types.act_type == torch.float8_e4m3fn: + timers.append( + bench_fns( + label, + sub_label, + f"cutlass_scaled_mm ({terse_type_name(types.act_type)})", + [cutlass_scaled_mm_create_bench_fn(bt) for bt in benchmark_tensors], + ) + ) + + if types.act_type != torch.float8_e4m3fn: + timers.append( + bench_fns( + label, + sub_label, + f"marlin ({name_type_string})", + [marlin_create_bench_fn(bt) for bt in benchmark_tensors], + ) + ) + + # machete + timers.append( + bench_fns( + label, + sub_label, + f"machete ({name_type_string})", + [ + machete_create_bench_fn(bt, out_type=types.output_type) + for bt in benchmark_tensors + ], + ) + ) + + if sweep_schedules: + global _SWEEP_SCHEDULES_RESULTS + + print("Finding best schedule for machete") + best = None + best_schedule = None + schedules = ops.machete_supported_schedules( + a_type=types.act_type, + b_type=types.weight_type, + group_scales_type=types.group_scale_type, + group_zeros_type=types.group_zero_type, + token_scales_type=types.token_scale_type, + channel_scales_type=types.channel_scale_type, + out_type=types.output_type, + ) + + if schedules is None or len(schedules) == 0: + raise ValueError("No schedules found to sweep") + + for schedule in reversed(schedules): + schedule_M = int(schedule.split("_")[0].split("x")[1]) + + # Prune known bad schedules + if schedule_M >= 2 * max(m, 16) or schedule_M < m // 4: + continue + + res = bench_fns( + label, + sub_label, + "machete_best", + [ + machete_create_bench_fn( + bt, out_type=types.output_type, schedule=schedule + ) + for bt in benchmark_tensors + ], + ) + + results_row = { + "M": m, + "K": k, + "N": n, + "group_size": group_size, + "schedule": schedule, + "median": res.median, + } + if _SWEEP_SCHEDULES_RESULTS is None: + _SWEEP_SCHEDULES_RESULTS = pd.DataFrame(columns=results_row.keys()) + _SWEEP_SCHEDULES_RESULTS.loc[len(_SWEEP_SCHEDULES_RESULTS)] = results_row + + print(f" {res.median:5.5} ", schedule) + if not best or res.median < best.median: + best = res + best_schedule = schedule + print("Best schedule:", best_schedule) + timers.append(best) + + return timers + + +# runner +def print_timers(timers: list[TMeasurement]): + compare = TBenchmark.Compare(timers) + compare.print() + + +def run(args, MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]: + types = TypeConfig( + act_type=args.act_type, + weight_type=scalar_types.uint4b8 + if args.group_zero_type is None + else scalar_types.uint4, + output_type=args.out_type, + group_scale_type=args.group_scale_type, + group_zero_type=args.group_zero_type, + channel_scale_type=args.channel_scale_type, + token_scale_type=args.token_scale_type, + ) + + results: list[TMeasurement] = [] + for m, k, n in MKNs: + timers = bench( + types, + args.group_size, + m, + k, + n, + f"{args.act_type}-gemm", + f"MKN=({m}x{k}x{n})", + sweep_schedules=args.sweep_schedules, + ) + print_timers(timers) + results.extend(timers) + + return results + + +# output makers +def make_output( + data: list[TMeasurement], + MKNs: Iterable[tuple[int, int, int]], + base_description: str, + timestamp=None, +): + print(f"== All Results {base_description} ====") + print_timers(data) + + # pickle all the results + timestamp = int(time.time()) if timestamp is None else timestamp + with open(f"{base_description}-{timestamp}.pkl", "wb") as f: + pkl.dump(data, f) + + +# argparse runners + + +def run_square_bench(args): + dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment)) + MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes)) + data = run(args.dtype, args.sweep_schedules, MKNs) + + make_output(data, MKNs, f"square_bench-{args.dtype}") + + +def run_range_bench(args): + m_start, k_start, n_start = (int(x) for x in args.dim_start.split(",")) + m_end, k_end, n_end = (int(x) for x in args.dim_end.split(",")) + m_increment, k_increment, n_increment = ( + int(x) for x in args.dim_increment.split(",") + ) + Ms = list(range(m_start, m_end + 1, m_increment)) + Ks = list(range(k_start, k_end + 1, k_increment)) + Ns = list(range(n_start, n_end + 1, n_increment)) + MKNs = list(product(Ms, Ks, Ns)) + + data = run(args.dtype, args.sweep_schedules, MKNs) + + make_output(data, MKNs, f"range_bench-{args.dtype}") + + +def run_model_bench(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + + def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]: + KNs = [] + for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]): + KN[tp_split_dim] = KN[tp_split_dim] // tp_size + KNs.append(KN) + return KNs + + model_bench_data = [] + models_tps = list(itertools.product(args.models, args.tp_sizes)) + for model, tp_size in models_tps: + Ms = args.batch_sizes + KNs = model_shapes(model, tp_size) + MKNs = [] + for m in Ms: + for k, n in KNs: + MKNs.append((m, k, n)) + + data = run(args, MKNs) + model_bench_data.append(data) + + type_string = f"{args.act_type}" + + # Print all results + for data, model_tp in zip(model_bench_data, models_tps): + model, tp_size = model_tp + print(f"== Results {type_string} {model}-TP{tp_size} ====") + print_timers(data) + + timestr = time.strftime("%Y%m%d-%H%M%S") + + all_results = [] + for d in model_bench_data: + all_results.extend(d) + + # pickle all data + with open(f"model_bench-{type_string}-{timestr}.pkl", "wb") as f: + args_dict = vars(args) + args_dict.pop("func") + pkl.dump( + { + "args": args_dict, + "results": all_results, + }, + f, + ) + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + return { + "bfloat16": torch.bfloat16, + "float16": torch.float16, + "int8": torch.int8, + "float8_e4m3fn": torch.float8_e4m3fn, + "int": torch.int, + "float": torch.float, + }[dt] + + class ToTorchDtype(argparse.Action): + def __call__(self, parser, namespace, values, option_string=None): + setattr(namespace, self.dest, to_torch_dtype(values)) + + parser = FlexibleArgumentParser( + description=""" +Benchmark Machete GEMM. + + To run square GEMMs: + python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 square_bench --dim-start 128 --dim-end 512 --dim-increment 64 + + To run constant N and K and sweep M: + python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384 + + To run dimensions from a model: + python3 ./benchmarks/kernels/benchmark_machete.py --dtype float16 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1 + + Output: + - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs. + """, # noqa: E501 + formatter_class=argparse.RawTextHelpFormatter, + ) + parser.add_argument( + "--act-type", + action=ToTorchDtype, + required=True, + choices=["bfloat16", "float16", "int8", "float8_e4m3fn"], + ) + parser.add_argument( + "--group-scale-type", + action=ToTorchDtype, + choices=["bfloat16", "float16"], + ) + parser.add_argument( + "--group-zero-type", + type=to_torch_dtype, + choices=["bfloat16", "float16"], + ) + parser.add_argument( + "--channel-scale-type", + action=ToTorchDtype, + choices=["float"], + ) + parser.add_argument( + "--token-scale-type", + action=ToTorchDtype, + choices=["float"], + ) + parser.add_argument( + "--out-type", + action=ToTorchDtype, + choices=["bfloat16", "float16"], + ) + parser.add_argument( + "--group-size", + type=int, + help="Available options are ['None', '-1', '128'], default=128", + default=128, + ) + parser.add_argument( + "--sweep-schedules", + action="store_true", + help="Run a sweep over all supported schedules", + ) + parser.add_argument( + "--sweep-csv-out", + help="CSV to store sweep results", + default="sch_sweep_results.csv", + ) + subparsers = parser.add_subparsers(dest="cmd", required=True) + + square_parser = subparsers.add_parser("square_bench") + square_parser.add_argument("--dim-start", type=int, required=True) + square_parser.add_argument("--dim-end", type=int, required=True) + square_parser.add_argument("--dim-increment", type=int, required=True) + square_parser.set_defaults(func=run_square_bench) + + range_parser = subparsers.add_parser("range_bench") + range_parser.add_argument( + "--dim-start", + type=str, + required=True, + help="Start value for M,K,N as common separated list", + ) + range_parser.add_argument( + "--dim-end", + type=str, + required=True, + help="End value (inclusive) for M,K,N as common separated list", + ) + range_parser.add_argument( + "--dim-increment", + type=str, + required=True, + help="Increment value for M,K,N as common separated list", + ) + range_parser.set_defaults(func=run_range_bench) + + model_parser = subparsers.add_parser("model_bench") + model_parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + model_parser.add_argument( + "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES + ) + model_parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + model_parser.set_defaults(func=run_model_bench) + + args = parser.parse_args() + + _SWEEP_SCHEDULES_RESULTS_CSV = args.sweep_csv_out + args.func(args) + + if _SWEEP_SCHEDULES_RESULTS is not None: + _SWEEP_SCHEDULES_RESULTS.to_csv(_SWEEP_SCHEDULES_RESULTS_CSV) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_marlin.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_marlin.py new file mode 100644 index 0000000000000000000000000000000000000000..34cc45e94d76d930013c9d60ca8c5f78dd0b7254 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_marlin.py @@ -0,0 +1,413 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import torch +import torch.utils.benchmark as benchmark +from benchmark_shapes import WEIGHT_SHAPES + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.gptq_marlin_24 import ( + GPTQ_MARLIN_24_MAX_PARALLEL, + GPTQ_MARLIN_24_MIN_THREAD_N, + GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, + GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES, +) +from vllm.model_executor.layers.quantization.utils.allspark_utils import ( + ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD, + ALLSPARK_SUPPORTED_QUANT_TYPES, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils import ( + GPTQ_MARLIN_MAX_PARALLEL, + GPTQ_MARLIN_MIN_THREAD_N, + MARLIN_SUPPORTED_GROUP_SIZES, + query_marlin_supported_quant_types, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import ( + FP4_MARLIN_SUPPORTED_GROUP_SIZES, + rand_marlin_weight_fp4_like, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import ( + marlin_quant_fp8_torch, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_test import ( + MarlinWorkspace, + awq_marlin_quantize, + marlin_quantize, +) +from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import ( + marlin_24_quantize, +) +from vllm.model_executor.layers.quantization.utils.quant_utils import ( + gptq_pack, + gptq_quantize_weights, + quantize_weights, + sort_weights, +) +from vllm.scalar_type import ScalarType, scalar_types +from vllm.utils import FlexibleArgumentParser + +DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"] +DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192] + +ACT_ORDER_OPTS = [False, True] +K_FULL_OPTS = [False, True] + + +def bench_run( + results: list[benchmark.Measurement], + model: str, + act_order: bool, + is_k_full: bool, + quant_type: ScalarType, + group_size: int, + size_m: int, + size_k: int, + size_n: int, +): + label = "Quant Matmul" + sub_label = "{}, act={} k_full={}, q={}, g={}, MKN=({}x{}x{})".format( + model, act_order, is_k_full, str(quant_type), group_size, size_m, size_k, size_n + ) + print(f"Testing: {sub_label}") + + a = torch.randn(size_m, size_k).to(torch.half).cuda() + b = torch.rand(size_k, size_n).to(torch.half).cuda() + has_zp = quant_type in [scalar_types.uint4, scalar_types.uint8] + if act_order and (group_size == -1 or group_size == size_k or has_zp): + return + if size_k % group_size != 0: + return + + marlin_24_supported = ( + quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES + and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES + ) + repack_supported = ( + quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES + and group_size in MARLIN_SUPPORTED_GROUP_SIZES + ) + allspark_supported = ( + quant_type in ALLSPARK_SUPPORTED_QUANT_TYPES + and group_size == -1 + and not act_order + and is_k_full + ) + + def gen_marlin_params(): + # Marlin quant + marlin_g_idx = marlin_sort_indices = marlin_zp = marlin_s2 = None + if quant_type == scalar_types.float4_e2m1f: + if group_size != 16 or act_order: + return + marlin_w_ref, marlin_q_w, marlin_s, marlin_s2 = rand_marlin_weight_fp4_like( + b.T, group_size + ) + elif quant_type == scalar_types.float8_e4m3fn: + if group_size not in [-1, 128] or act_order: + return + marlin_w_ref, marlin_q_w, marlin_s = marlin_quant_fp8_torch(b.T, group_size) + elif group_size == 16: + return + elif has_zp: + marlin_w_ref, marlin_q_w, marlin_s, marlin_zp = awq_marlin_quantize( + b, quant_type, group_size + ) + else: + marlin_w_ref, marlin_q_w, marlin_s, marlin_g_idx, marlin_sort_indices, _ = ( + marlin_quantize(b, quant_type, group_size, act_order) + ) + return ( + marlin_w_ref, + marlin_q_w, + marlin_s, + marlin_s2, + marlin_zp, + marlin_g_idx, + marlin_sort_indices, + ) + + def gen_marlin_24_params(): + marlin_24_w_ref = marlin_24_q_w_comp = marlin_24_meta = marlin_24_s = None + if marlin_24_supported: + (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) = ( + marlin_24_quantize(b, quant_type, group_size) + ) + return (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) + + def gen_repack_params(): + q_w_gptq = None + repack_sort_indices = None + if repack_supported: + (w_ref, q_w, s, g_idx, rand_perm) = gptq_quantize_weights( + b, quant_type, group_size, act_order + ) + q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n) + + # For act_order, sort the "weights" and "g_idx" + # so that group ids are increasing + repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device) + if act_order: + (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx) + return q_w_gptq, repack_sort_indices + + def gen_allspark_params(): + qw_reorder = s_reorder = zp_reorder = sm_count = sm_version = ( + CUBLAS_M_THRESHOLD + ) = None + nonlocal allspark_supported + if allspark_supported: + properties = torch.cuda.get_device_properties(b.device.index) + sm_count = properties.multi_processor_count + sm_version = properties.major * 10 + properties.minor + + supported_arch = sm_version >= 80 and sm_version < 90 + allspark_supported = allspark_supported and supported_arch + if supported_arch: + w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size, has_zp) + qw = qw.to(torch.uint8) + + qw_reorder, s_reorder, zp_reorder = ops.allspark_repack_weight( + qw, s, zp, has_zp + ) + CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD + return ( + qw_reorder, + s_reorder, + zp_reorder, + sm_count, + sm_version, + CUBLAS_M_THRESHOLD, + ) + + ( + marlin_w_ref, + marlin_q_w, + marlin_s, + marlin_s2, + marlin_zp, + marlin_g_idx, + marlin_sort_indices, + ) = gen_marlin_params() + marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s = ( + gen_marlin_24_params() + ) + q_w_gptq, repack_sort_indices = gen_repack_params() + qw_reorder, s_reorder, zp_reorder, sm_count, sm_version, CUBLAS_M_THRESHOLD = ( + gen_allspark_params() + ) + + # Prepare + marlin_workspace = MarlinWorkspace( + size_n, GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL + ) + marlin_24_workspace = MarlinWorkspace( + size_n, GPTQ_MARLIN_24_MIN_THREAD_N, GPTQ_MARLIN_24_MAX_PARALLEL + ) + + globals = { + # Gen params + "quant_type": quant_type, + "group_size": group_size, + "size_m": size_m, + "size_n": size_n, + "size_k": size_k, + "a": a, + # Marlin params + "marlin_w_ref": marlin_w_ref, + "marlin_q_w": marlin_q_w, + "marlin_s": marlin_s, + "marlin_s2": marlin_s2, + "marlin_zp": marlin_zp, + "marlin_g_idx": marlin_g_idx, + "marlin_sort_indices": marlin_sort_indices, + "marlin_workspace": marlin_workspace, + "is_k_full": is_k_full, + # Marlin_24 params + "marlin_24_w_ref": marlin_24_w_ref, + "marlin_24_q_w_comp": marlin_24_q_w_comp, + "marlin_24_meta": marlin_24_meta, + "marlin_24_s": marlin_24_s, + "marlin_24_workspace": marlin_24_workspace, + # GPTQ params + "q_w_gptq": q_w_gptq, + "repack_sort_indices": repack_sort_indices, + # AllSpark W8A16 params + "qw_reorder": qw_reorder, + "s_reorder": s_reorder, + "zp_reorder": zp_reorder, + "sm_count": sm_count, + "sm_version": sm_version, + "CUBLAS_M_THRESHOLD": CUBLAS_M_THRESHOLD, + # Kernels + "gptq_marlin_gemm": ops.gptq_marlin_gemm, + "gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm, + "gptq_marlin_repack": ops.gptq_marlin_repack, + "allspark_w8a16_gemm": ops.allspark_w8a16_gemm, + } + + min_run_time = 1 + + # Warmup pytorch + for _ in range(5): + torch.matmul(a, marlin_w_ref) + + results.append( + benchmark.Timer( + stmt="torch.matmul(a, marlin_w_ref)", + globals=globals, + label=label, + sub_label=sub_label, + description="pytorch_gemm", + ).blocked_autorange(min_run_time=min_run_time) + ) + + results.append( + benchmark.Timer( + stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False, False)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_gemm", + ).blocked_autorange(min_run_time=min_run_time) + ) + + results.append( + benchmark.Timer( + stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True, False)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_gemm_fp32", + ).blocked_autorange(min_run_time=min_run_time) + ) + + if marlin_24_supported: + results.append( + benchmark.Timer( + stmt="output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, quant_type, size_m, size_n, size_k)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_24_gemm", + ).blocked_autorange(min_run_time=min_run_time) + ) + + if repack_supported: + results.append( + benchmark.Timer( + stmt="q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="gptq_marlin_repack", + ).blocked_autorange(min_run_time=min_run_time) + ) + + if allspark_supported: + results.append( + benchmark.Timer( + stmt="output = allspark_w8a16_gemm(a, qw_reorder, s_reorder, zp_reorder, size_n, group_size, sm_count, sm_version, CUBLAS_M_THRESHOLD, False, True)", # noqa: E501 + globals=globals, + label=label, + sub_label=sub_label, + description="allspark_w8a16_gemm_fp32", + ).blocked_autorange(min_run_time=min_run_time) + ) + + +def main(args): + print("Benchmarking models:") + for i, model in enumerate(args.models): + print(f"[{i}] {model}") + results: list[benchmark.Measurement] = [] + + for model in args.models: + for layer in WEIGHT_SHAPES[model]: + size_k = layer[0] + size_n = layer[1] + + if len(args.limit_k) > 0 and size_k not in args.limit_k: + continue + + if len(args.limit_n) > 0 and size_n not in args.limit_n: + continue + + for act_order in ACT_ORDER_OPTS: + if ( + len(args.limit_act_order) > 0 + and act_order not in args.limit_act_order + ): + continue + + for is_k_full in K_FULL_OPTS: + if ( + len(args.limit_k_full) > 0 + and is_k_full not in args.limit_k_full + ): + continue + + for quant_type in query_marlin_supported_quant_types(): + if ( + len(args.limit_num_bits) > 0 + and quant_type.size_bits not in args.limit_num_bits + ): + continue + + for group_size in ( + MARLIN_SUPPORTED_GROUP_SIZES + + FP4_MARLIN_SUPPORTED_GROUP_SIZES + ): + if ( + len(args.limit_group_size) > 0 + and group_size not in args.limit_group_size + ): + continue + + # For act_order, the group_size must be less than + # size_k + if act_order and (group_size == size_k or group_size == -1): + continue + + for size_m in args.batch_sizes: + bench_run( + results, + model, + act_order, + is_k_full, + quant_type, + group_size, + size_m, + size_k, + size_n, + ) + + compare = benchmark.Compare(results) + compare.print() + + +# For quick benchmarking use: +# python benchmark_marlin.py --batch-sizes 1 16 32 --limit-k 4096 --limit-n 4096 --limit-group-size 128 --limit-num-bits 4 --limit-act-order 0 --limit-k-full 1 # noqa E501 +# +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark Marlin across specified models/shapes/batches" + ) + parser.add_argument( + "--models", + nargs="+", + type=str, + default=DEFAULT_MODELS, + choices=WEIGHT_SHAPES.keys(), + ) + parser.add_argument( + "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES + ) + parser.add_argument("--limit-k", nargs="+", type=int, default=[]) + parser.add_argument("--limit-n", nargs="+", type=int, default=[]) + parser.add_argument("--limit-group-size", nargs="+", type=int, default=[]) + parser.add_argument("--limit-num-bits", nargs="+", type=int, default=[]) + parser.add_argument("--limit-act-order", nargs="+", type=int, default=[]) + parser.add_argument("--limit-k-full", nargs="+", type=int, default=[]) + + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_moe.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..07af58d81c68331597f1e63c0c62dc814b41025c --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_moe.py @@ -0,0 +1,737 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +import json +import time +from contextlib import nullcontext +from datetime import datetime +from itertools import product +from typing import Any, TypedDict + +import ray +import torch +from ray.experimental.tqdm_ray import tqdm + +from vllm.model_executor.layers.fused_moe.fused_moe import * +from vllm.platforms import current_platform +from vllm.transformers_utils.config import get_config +from vllm.triton_utils import triton +from vllm.utils import FlexibleArgumentParser + +FP8_DTYPE = current_platform.fp8_dtype() + + +class BenchmarkConfig(TypedDict): + BLOCK_SIZE_M: int + BLOCK_SIZE_N: int + BLOCK_SIZE_K: int + GROUP_SIZE_M: int + num_warps: int + num_stages: int + + +def benchmark_config( + config: BenchmarkConfig, + num_tokens: int, + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + num_iters: int = 100, + block_quant_shape: list[int] = None, + use_deep_gemm: bool = False, +) -> float: + init_dtype = torch.float16 if use_fp8_w8a8 else dtype + x = torch.randn(num_tokens, hidden_size, dtype=dtype) + if use_int8_w8a16: + w1 = torch.randint( + -127, + 127, + ( + num_experts, + shard_intermediate_size, + hidden_size, + ), + dtype=torch.int8, + ) + w2 = torch.randint( + -127, + 127, + ( + num_experts, + hidden_size, + shard_intermediate_size // 2, + ), + dtype=torch.int8, + ) + else: + w1 = torch.randn( + num_experts, shard_intermediate_size, hidden_size, dtype=init_dtype + ) + w2 = torch.randn( + num_experts, hidden_size, shard_intermediate_size // 2, dtype=init_dtype + ) + gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32) + + w1_scale = None + w2_scale = None + a1_scale = None + a2_scale = None + if use_int8_w8a16: + w1_scale = torch.randn( + (num_experts, 2 * shard_intermediate_size), dtype=torch.float32 + ) + w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32) + if use_fp8_w8a8: + if block_quant_shape: + block_n, block_k = block_quant_shape[0], block_quant_shape[1] + E = num_experts + N = shard_intermediate_size // 2 + K = hidden_size + factor_for_scale = 1e-2 + n_tiles_w1 = (2 * N + block_n - 1) // block_n + n_tiles_w2 = (K + block_n - 1) // block_n + k_tiles_w1 = (K + block_k - 1) // block_k + k_tiles_w2 = (N + block_k - 1) // block_k + w1_scale = ( + torch.rand((E, n_tiles_w1, k_tiles_w1), dtype=torch.float32) + * factor_for_scale + ) + w2_scale = ( + torch.rand((E, n_tiles_w2, k_tiles_w2), dtype=torch.float32) + * factor_for_scale + ) + else: + w1_scale = torch.randn(num_experts, dtype=torch.float32) + w2_scale = torch.randn(num_experts, dtype=torch.float32) + + a1_scale = torch.randn(1, dtype=torch.float32) + a2_scale = torch.randn(1, dtype=torch.float32) + + w1 = w1.to(FP8_DTYPE) + w2 = w2.to(FP8_DTYPE) + + input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32) + + def prepare(i: int): + input_gating.copy_(gating_output[i]) + + def run(): + from vllm.model_executor.layers.fused_moe import override_config + + with override_config(config): + if use_deep_gemm: + topk_weights, topk_ids, token_expert_indices = fused_topk( + x, input_gating, topk, False + ) + return fused_experts( + x, + w1, + w2, + topk_weights, + topk_ids, + inplace=True, + use_fp8_w8a8=use_fp8_w8a8, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_quant_shape, + allow_deep_gemm=True, + ) + else: + fused_moe( + x, + w1, + w2, + input_gating, + topk, + renormalize=True, + inplace=True, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + w1_scale=w1_scale, + w2_scale=w2_scale, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_quant_shape, + ) + + # JIT compilation & warmup + run() + torch.cuda.synchronize() + + # Capture 10 invocations with CUDA graph + graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(graph): + for _ in range(10): + run() + torch.cuda.synchronize() + + # Warmup + for _ in range(5): + graph.replay() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + prepare(i) + torch.cuda.synchronize() + + start_event.record() + graph.replay() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + graph.reset() + return avg + + +def get_rocm_tuning_space(use_fp16): + block_mn_range = [16, 32, 64, 128, 256] + block_k_range = [16, 32, 64, 128, 256] + if not use_fp16: + block_k_range.remove(16) # BLOCK_K=16 not supported for fp8 + num_warps_range = [1, 2, 4, 8] + group_m_range = [1, 4, 8, 16, 32] + num_stage_range = [2] + waves_per_eu_range = [0] + matrix_instr_nonkdim_range = [16, 32] if use_fp16 else [] + kpack_range = [1, 2] if use_fp16 else [] + + param_ranges = { + "BLOCK_SIZE_M": block_mn_range, + "BLOCK_SIZE_N": block_mn_range, + "BLOCK_SIZE_K": block_k_range, + "GROUP_SIZE_M": group_m_range, + "num_warps": num_warps_range, + "num_stages": num_stage_range, + "waves_per_eu": waves_per_eu_range, + } + if use_fp16: + param_ranges["matrix_instr_nonkdim"] = matrix_instr_nonkdim_range + param_ranges["kpack"] = kpack_range + + return param_ranges + + +def get_configs_compute_bound(use_fp16, block_quant_shape) -> list[dict[str, int]]: + configs: list[BenchmarkConfig] = [] + + if current_platform.is_rocm(): + param_ranges = get_rocm_tuning_space(use_fp16) + else: + # Reduced search space for faster tuning. + # TODO(woosuk): Increase the search space and use a performance model to + # prune the search space. + block_m_range = [16, 32, 64, 128, 256] + block_n_range = [32, 64, 128, 256] + block_k_range = [64, 128, 256] + num_warps_range = [4, 8] + group_m_range = [1, 16, 32, 64] + num_stage_range = [2, 3, 4, 5] + + param_ranges = { + "BLOCK_SIZE_M": block_m_range, + "BLOCK_SIZE_N": block_n_range, + "BLOCK_SIZE_K": block_k_range, + "GROUP_SIZE_M": group_m_range, + "num_warps": num_warps_range, + "num_stages": num_stage_range, + } + + keys, values = zip(*param_ranges.items()) + for config_values in product(*values): + config = dict(zip(keys, config_values)) + configs.append(config) + + # Remove configs that are not compatible with fp8 block quantization + # BLOCK_SIZE_K must be a multiple of block_k + # BLOCK_SIZE_N must be a multiple of block_n + if block_quant_shape is not None and not use_fp16: + block_n, block_k = block_quant_shape[0], block_quant_shape[1] + for config in configs[:]: + if ( + config["BLOCK_SIZE_K"] % block_k != 0 + or config["BLOCK_SIZE_N"] % block_n != 0 + ): + configs.remove(config) + return configs + + +def prune_rocm_search_space( + num_tokens, shard_intermediate_size, hidden_size, search_space, is_fp16, topk +): + N1, K1 = shard_intermediate_size, hidden_size + N2, K2 = hidden_size, shard_intermediate_size // 2 + pruned_space_1 = prune_rocm_configs( + num_tokens * topk, N1, K1, search_space, is_fp16 + ) + pruned_space_2 = prune_rocm_configs( + num_tokens * topk, N2, K2, search_space, is_fp16 + ) + search_space = merge_unique_dicts(pruned_space_1, pruned_space_2) + return search_space + + +# The following code is inspired by ROCm/Triton GEMM tuning script: +# https://github.com/ROCm/triton/blob/triton-mlir/scripts/amd/gemm/tune_gemm.py#L89 +def prune_rocm_configs(M, N, K, configs, is_fp16=True): + pruned_configs = [] + elemBytes_a = 2 if is_fp16 else 1 + elemBytes_b = 2 if is_fp16 else 1 + + mfma = 16 if M < 32 or N < 32 else 32 + + # TODO (zhanglx): figure out the boundary between large and small gemms + large_gemm = False + if M >= 2048 and N >= 2048: + large_gemm = True + + for config in configs: + BLOCK_SIZE_M = config.get("BLOCK_SIZE_M") + BLOCK_SIZE_N = config.get("BLOCK_SIZE_N") + BLOCK_SIZE_K = config.get("BLOCK_SIZE_K") + num_warps = config.get("num_warps") + + if is_fp16: + matrix_instr_nonkdim = config.get("matrix_instr_nonkdim") + if matrix_instr_nonkdim > mfma: + continue + if mfma == 4 and BLOCK_SIZE_K < 64: + continue + # some layouts could not work properly in case + # number elements per thread is less 1 + if BLOCK_SIZE_M * BLOCK_SIZE_N < 64: + continue + SPLIT_K = config.get("SPLIT_K", 1) + GROUP_M = config.get("GROUP_SIZE_M") + if is_fp16: + if ( + matrix_instr_nonkdim > BLOCK_SIZE_M + or matrix_instr_nonkdim > BLOCK_SIZE_N + ): + continue + if matrix_instr_nonkdim >= M and matrix_instr_nonkdim != BLOCK_SIZE_M: + continue + if matrix_instr_nonkdim >= N and matrix_instr_nonkdim != BLOCK_SIZE_N: + continue + # Skip BLOCK_SIZE that is too large compare to M/N + # unless BLOCK_SIZE is already small enough + if M * 2 < BLOCK_SIZE_M and BLOCK_SIZE_M != 16: + continue + if N * 2 < BLOCK_SIZE_N and BLOCK_SIZE_N != 16: + continue + # skip large split_k when not necessary + if SPLIT_K != 1 and not need_split_k(M, N, K): + continue + # skip split_k that leads to EVEN_K = false + leap = SPLIT_K * BLOCK_SIZE_K + modv = K % leap + if modv != 0: + continue + # skip large GROUP_M + if GROUP_M * BLOCK_SIZE_M > M and GROUP_M != 1: + continue + # out of shared memory resource + # TODO (zhanglx): This does not consider the LDS usage in the epilogue + LDS = ( + BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a + + BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b + ) + if LDS > 65536: + continue + # Skip small block sizes and num_warps for large gemm + # For fp16 and f8, we want to only use BLOCK_SIZE >= 64 + if large_gemm: + if BLOCK_SIZE_M < 64 or BLOCK_SIZE_N < 64: + continue + if BLOCK_SIZE_K < 64: + continue + if num_warps < 4: + continue + + pruned_configs.append(config) + + return pruned_configs + + +def need_split_k(SIZE_M, SIZE_N, SIZE_K): + return (SIZE_M < 64 or SIZE_N < 64) and SIZE_K > 1024 + + +def merge_unique_dicts(list1, list2): + result = [] + combined_list = list1.copy() + combined_list.extend(list2) + for dictionary in combined_list: + if dictionary not in result: + result.append(dictionary) + return result + + +@ray.remote(num_gpus=1) +class BenchmarkWorker: + def __init__(self, seed: int) -> None: + torch.set_default_device("cuda") + current_platform.seed_everything(seed) + self.seed = seed + # Get the device ID to allocate tensors and kernels + # on the respective GPU. This is required for Ray to work + # correctly with multi-GPU tuning on the ROCm platform. + self.device_id = int(ray.get_gpu_ids()[0]) + + def benchmark( + self, + num_tokens: int, + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + block_quant_shape: list[int] = None, + use_deep_gemm: bool = False, + ) -> tuple[dict[str, int], float]: + current_platform.seed_everything(self.seed) + dtype_str = get_config_dtype_str( + dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8 + ) + # NOTE(woosuk): The current naming convention uses w2.shape[2], which + # is the intermediate size after silu_and_mul. + op_config = get_moe_configs( + num_experts, shard_intermediate_size // 2, dtype_str + ) + if op_config is None: + config = get_default_config( + num_tokens, + num_experts, + shard_intermediate_size, + hidden_size, + topk, + dtype_str, + is_marlin=False, + ) + else: + config = op_config[min(op_config.keys(), key=lambda x: abs(x - num_tokens))] + kernel_time = benchmark_config( + config, + num_tokens, + num_experts, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=100, + block_quant_shape=block_quant_shape, + use_deep_gemm=use_deep_gemm, + ) + return config, kernel_time + + def tune( + self, + num_tokens: int, + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + search_space: list[dict[str, int]], + block_quant_shape: list[int], + use_deep_gemm: bool, + ) -> dict[str, int]: + best_config = None + best_time = float("inf") + if current_platform.is_rocm(): + is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16) + search_space = prune_rocm_search_space( + num_tokens, + shard_intermediate_size, + hidden_size, + search_space, + is_fp16, + topk, + ) + + need_device_guard = False + if current_platform.is_rocm(): + visible_device = os.environ.get("ROCR_VISIBLE_DEVICES", None) + if visible_device != f"{self.device_id}": + need_device_guard = True + + with torch.cuda.device(self.device_id) if need_device_guard else nullcontext(): + for config in tqdm(search_space): + try: + kernel_time = benchmark_config( + config, + num_tokens, + num_experts, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=20, + block_quant_shape=block_quant_shape, + use_deep_gemm=use_deep_gemm, + ) + except triton.runtime.autotuner.OutOfResources: + # Some configurations may be invalid and fail to compile. + continue + + if kernel_time < best_time: + best_time = kernel_time + best_config = config + now = datetime.now() + print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}") + assert best_config is not None + return best_config + + +def sort_config(config: BenchmarkConfig) -> BenchmarkConfig: + return { + "BLOCK_SIZE_M": config["BLOCK_SIZE_M"], + "BLOCK_SIZE_N": config["BLOCK_SIZE_N"], + "BLOCK_SIZE_K": config["BLOCK_SIZE_K"], + "GROUP_SIZE_M": config["GROUP_SIZE_M"], + "num_warps": config["num_warps"], + "num_stages": config["num_stages"], + **( + {"waves_per_eu": config["waves_per_eu"]} if "waves_per_eu" in config else {} + ), + **( + {"matrix_instr_nonkdim": config["matrix_instr_nonkdim"]} + if "matrix_instr_nonkdim" in config + else {} + ), + **({"kpack": config["kpack"]} if "kpack" in config else {}), + } + + +def save_configs( + configs: dict[int, BenchmarkConfig], + num_experts: int, + shard_intermediate_size: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + block_quant_shape: list[int], +) -> None: + dtype_str = get_config_dtype_str( + dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8 + ) + + # NOTE(woosuk): The current naming convention uses w2.shape[2], which + # is the intermediate size after silu_and_mul. + filename = get_config_file_name( + num_experts, shard_intermediate_size // 2, dtype_str, block_quant_shape + ) + + print(f"Writing best config to {filename}...") + with open(filename, "w") as f: + json.dump(configs, f, indent=4) + f.write("\n") + + +def get_weight_block_size_safety(config, default_value=None): + quantization_config = getattr(config, "quantization_config", {}) + if isinstance(quantization_config, dict): + return quantization_config.get("weight_block_size", default_value) + return default_value + + +def main(args: argparse.Namespace): + print(args) + + config = get_config(model=args.model, trust_remote_code=args.trust_remote_code) + if args.model_prefix: + config = getattr(config, args.model_prefix) + + if config.architectures[0] == "DbrxForCausalLM": + E = config.ffn_config.moe_num_experts + topk = config.ffn_config.moe_top_k + intermediate_size = config.ffn_config.ffn_hidden_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + elif config.architectures[0] == "JambaForCausalLM": + E = config.num_experts + topk = config.num_experts_per_tok + intermediate_size = config.intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + elif config.architectures[0] in ("DeepseekV3ForCausalLM", "DeepseekV2ForCausalLM"): + E = config.n_routed_experts + topk = config.num_experts_per_tok + intermediate_size = config.moe_intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + elif config.architectures[0] in ("Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"): + E = config.num_experts + topk = config.num_experts_per_tok + intermediate_size = config.moe_intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + else: + # Support for llama4 + config = config.get_text_config() + # Default: Mixtral. + E = config.num_local_experts + topk = config.num_experts_per_tok + intermediate_size = config.intermediate_size + shard_intermediate_size = 2 * intermediate_size // args.tp_size + + hidden_size = config.hidden_size + dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype + use_fp8_w8a8 = args.dtype == "fp8_w8a8" + use_int8_w8a16 = args.dtype == "int8_w8a16" + block_quant_shape = get_weight_block_size_safety(config) + + if args.batch_size is None: + batch_sizes = [ + 1, + 2, + 4, + 8, + 16, + 24, + 32, + 48, + 64, + 96, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ] + else: + batch_sizes = args.batch_size + + use_deep_gemm = bool(args.use_deep_gemm) + + if current_platform.is_rocm() and "HIP_VISIBLE_DEVICES" in os.environ: + # Ray will set ROCR_VISIBLE_DEVICES for device visibility + logger.warning( + "Ray uses ROCR_VISIBLE_DEVICES to control device accessibility." + "Replacing HIP_VISIBLE_DEVICES with ROCR_VISIBLE_DEVICES." + ) + val = os.environ["HIP_VISIBLE_DEVICES"] + os.environ["ROCR_VISIBLE_DEVICES"] = val + del os.environ["HIP_VISIBLE_DEVICES"] + + ray.init() + num_gpus = int(ray.available_resources()["GPU"]) + workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)] + + def _distribute(method: str, inputs: list[Any]) -> list[Any]: + outputs = [] + worker_idx = 0 + for input_args in inputs: + worker = workers[worker_idx] + worker_method = getattr(worker, method) + output = worker_method.remote(*input_args) + outputs.append(output) + worker_idx = (worker_idx + 1) % num_gpus + return ray.get(outputs) + + if args.tune: + is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16) + search_space = get_configs_compute_bound(is_fp16, block_quant_shape) + print(f"Start tuning over {len(search_space)} configurations...") + + start = time.time() + configs = _distribute( + "tune", + [ + ( + batch_size, + E, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + search_space, + block_quant_shape, + use_deep_gemm, + ) + for batch_size in batch_sizes + ], + ) + best_configs = { + M: sort_config(config) for M, config in zip(batch_sizes, configs) + } + save_configs( + best_configs, + E, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + block_quant_shape, + ) + end = time.time() + print(f"Tuning took {end - start:.2f} seconds") + else: + outputs = _distribute( + "benchmark", + [ + ( + batch_size, + E, + shard_intermediate_size, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + block_quant_shape, + use_deep_gemm, + ) + for batch_size in batch_sizes + ], + ) + + for batch_size, (config, kernel_time) in zip(batch_sizes, outputs): + print(f"Batch size: {batch_size}, config: {config}") + print(f"Kernel time: {kernel_time:.2f} us") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser() + parser.add_argument( + "--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1" + ) + parser.add_argument( + "--tp-size", "-tp", "--tensor-parallel-size", type=int, default=2 + ) + parser.add_argument( + "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto" + ) + parser.add_argument("--use-deep-gemm", action="store_true") + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--batch-size", type=int, nargs="+", required=False) + parser.add_argument("--tune", action="store_true") + parser.add_argument("--trust-remote-code", action="store_true") + parser.add_argument("--model-prefix", type=str, required=False) + args = parser.parse_args() + + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_moe_align_block_size.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_moe_align_block_size.py new file mode 100644 index 0000000000000000000000000000000000000000..5170ac09dc42af8fac0f15e48dc5d348b728f04a --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_moe_align_block_size.py @@ -0,0 +1,159 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import argparse +import itertools + +import torch + +from vllm import _custom_ops as ops +from vllm.model_executor.layers.fused_moe.moe_align_block_size import ( + moe_align_block_size_triton, +) +from vllm.triton_utils import triton + + +def get_topk_ids(num_tokens: int, num_experts: int, topk: int) -> torch.Tensor: + return torch.stack( + [ + torch.randperm(num_experts, dtype=torch.int32, device="cuda")[:topk] + for _ in range(num_tokens) + ] + ) + + +def check_correctness(num_tokens, num_experts=256, block_size=256, topk=8): + """ + Verifies vllm vs. Triton + """ + topk_ids = get_topk_ids(num_tokens, num_experts, topk) + + # 1. malloc space for triton and vllm + # malloc enough space (max_num_tokens_padded) for the sorted ids + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + sorted_ids_triton = torch.empty( + (max_num_tokens_padded,), dtype=torch.int32, device="cuda" + ) + sorted_ids_triton.fill_(topk_ids.numel()) # fill with sentinel value + expert_ids_triton = torch.zeros( + (max_num_tokens_padded // block_size,), dtype=torch.int32, device="cuda" + ) + num_tokens_post_pad_triton = torch.empty((1,), dtype=torch.int32, device="cuda") + + sorted_ids_vllm = torch.empty_like(sorted_ids_triton) + sorted_ids_vllm.fill_(topk_ids.numel()) + expert_ids_vllm = torch.zeros_like(expert_ids_triton) + num_tokens_post_pad_vllm = torch.empty_like(num_tokens_post_pad_triton) + + # 2. run implementations + moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids_triton, + expert_ids_triton, + num_tokens_post_pad_triton, + ) + + ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids_vllm, + expert_ids_vllm, + num_tokens_post_pad_vllm, + ) + print(f"✅ VLLM implementation works with {num_experts} experts!") + + # 3. compare results + if torch.allclose(expert_ids_triton, expert_ids_vllm) and torch.allclose( + num_tokens_post_pad_triton, num_tokens_post_pad_vllm + ): + print("✅ Triton and VLLM implementations match.") + else: + print("❌ Triton and VLLM implementations DO NOT match.") + print("Triton expert_ids:", expert_ids_triton) + print("VLLM expert_ids:", expert_ids_vllm) + print("Triton num_tokens_post_pad:", num_tokens_post_pad_triton) + print("VLLM num_tokens_post_pad:", num_tokens_post_pad_vllm) + + +# test configurations +num_tokens_range = [1, 16, 256, 4096] +num_experts_range = [16, 64, 224, 256, 280, 512] +topk_range = [1, 2, 8] +configs = list(itertools.product(num_tokens_range, num_experts_range, topk_range)) + + +@triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["num_tokens", "num_experts", "topk"], + x_vals=configs, + line_arg="provider", + line_vals=["vllm", "triton"], # "triton" + line_names=["VLLM", "Triton"], # "Triton" + plot_name="moe-align-block-size-performance", + args={}, + ) +) +def benchmark(num_tokens, num_experts, topk, provider): + """Benchmark function for Triton.""" + block_size = 256 + topk_ids = get_topk_ids(num_tokens, num_experts, topk) + + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + sorted_ids = torch.empty((max_num_tokens_padded,), dtype=torch.int32, device="cuda") + sorted_ids.fill_(topk_ids.numel()) + max_num_m_blocks = max_num_tokens_padded // block_size + expert_ids = torch.empty((max_num_m_blocks,), dtype=torch.int32, device="cuda") + num_tokens_post_pad = torch.empty((1,), dtype=torch.int32, device="cuda") + + quantiles = [0.5, 0.2, 0.8] + + if provider == "vllm": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids.clone(), + expert_ids.clone(), + num_tokens_post_pad.clone(), + ), + quantiles=quantiles, + ) + elif provider == "triton": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids.clone(), + expert_ids.clone(), + num_tokens_post_pad.clone(), + ), + quantiles=quantiles, + ) + + return 1000 * ms, 1000 * max_ms, 1000 * min_ms + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--num_experts", + type=int, + default=64, + choices=[8, 16, 32, 64, 128, 256], + ) + parser.add_argument( + "--topk", + type=int, + default=8, + choices=[2, 4, 8], + help="Top-k value for correctness check.", + ) + args = parser.parse_args() + + print("Running correctness check...") + check_correctness(num_tokens=1024, num_experts=args.num_experts, topk=args.topk) + benchmark.run(print_data=True, show_plots=True) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_moe_permute_unpermute.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_moe_permute_unpermute.py new file mode 100644 index 0000000000000000000000000000000000000000..dba1f3943b96c370acf1262273695dbdb4097711 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_moe_permute_unpermute.py @@ -0,0 +1,417 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import argparse +from typing import Any, TypedDict + +import ray +import torch +from transformers import AutoConfig + +from vllm.model_executor.layers.fused_moe.deep_gemm_moe import ( + _moe_permute, + _moe_unpermute_and_reduce, +) +from vllm.model_executor.layers.fused_moe.fused_moe import * +from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import * +from vllm.model_executor.layers.fused_moe.utils import _fp8_quantize +from vllm.platforms import current_platform +from vllm.utils import FlexibleArgumentParser + +FP8_DTYPE = current_platform.fp8_dtype() + + +class BenchmarkConfig(TypedDict): + BLOCK_SIZE_M: int + BLOCK_SIZE_N: int + BLOCK_SIZE_K: int + GROUP_SIZE_M: int + num_warps: int + num_stages: int + + +def benchmark_permute( + num_tokens: int, + num_experts: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + num_iters: int = 100, + use_customized_permute: bool = False, +) -> float: + # init_dtype = torch.float16 if use_fp8_w8a8 else dtype + hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype) + # output_hidden_states = torch.empty_like(hidden_states) + if use_fp8_w8a8: + align_block_size = 128 # deepgemm needs 128 m aligned block + qhidden_states, scale = _fp8_quantize(hidden_states, None, None) + else: + align_block_size = None + qhidden_states = hidden_states + + gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32) + + input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32) + topk_weights, topk_ids, token_expert_indices = fused_topk( + qhidden_states, input_gating, topk, False + ) + + def prepare(i: int): + input_gating.copy_(gating_output[i]) + + def run(): + if use_customized_permute: + (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = ( + moe_permute( + qhidden_states, + topk_weights=topk_weights, + topk_ids=topk_ids, + token_expert_indices=token_expert_indices, + topk=topk, + n_expert=num_experts, + n_local_expert=num_experts, + expert_map=None, + align_block_size=align_block_size, + ) + ) + else: + ( + permuted_hidden_states, + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) = _moe_permute( + qhidden_states, None, topk_ids, num_experts, None, align_block_size + ) + + # JIT compilation & warmup + run() + torch.cuda.synchronize() + + # Capture 10 invocations with CUDA graph + graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(graph): + for _ in range(10): + run() + torch.cuda.synchronize() + + # Warmup + for _ in range(5): + graph.replay() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + prepare(i) + torch.cuda.synchronize() + + start_event.record() + graph.replay() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + graph.reset() + return avg + + +def benchmark_unpermute( + num_tokens: int, + num_experts: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + num_iters: int = 100, + use_customized_permute: bool = False, +) -> float: + # init_dtype = torch.float16 if use_fp8_w8a8 else dtype + hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype) + output_hidden_states = torch.empty_like(hidden_states) + if use_fp8_w8a8: + align_block_size = 128 # deepgemm needs 128 m aligned block + qhidden_states, scale = _fp8_quantize(hidden_states, None, None) + else: + align_block_size = None + qhidden_states = hidden_states + + input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32) + + topk_weights, topk_ids, token_expert_indices = fused_topk( + qhidden_states, input_gating, topk, False + ) + + def prepare(): + if use_customized_permute: + (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = ( + moe_permute( + qhidden_states, + topk_weights=topk_weights, + topk_ids=topk_ids, + token_expert_indices=token_expert_indices, + topk=topk, + n_expert=num_experts, + n_local_expert=num_experts, + expert_map=None, + align_block_size=align_block_size, + ) + ) + # convert to fp16/bf16 as gemm output + return ( + permuted_hidden_states.to(dtype), + first_token_off, + inv_perm_idx, + m_indices, + ) + else: + ( + permuted_qhidden_states, + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) = _moe_permute( + qhidden_states, None, topk_ids, num_experts, None, align_block_size + ) + # convert to fp16/bf16 as gemm output + return ( + permuted_qhidden_states.to(dtype), + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) + + def run(input: tuple): + if use_customized_permute: + (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = input + moe_unpermute( + permuted_hidden_states, + topk_weights, + topk_ids, + inv_perm_idx, + first_token_off, + topk, + num_experts, + num_experts, + ) + else: + ( + permuted_hidden_states, + a1q_scale, + sorted_token_ids, + expert_ids, + inv_perm, + ) = input + _moe_unpermute_and_reduce( + output_hidden_states, permuted_hidden_states, inv_perm, topk_weights + ) + + # JIT compilation & warmup + input = prepare() + run(input) + torch.cuda.synchronize() + + # Capture 10 invocations with CUDA graph + graph = torch.cuda.CUDAGraph() + with torch.cuda.graph(graph): + for _ in range(10): + run(input) + torch.cuda.synchronize() + + # Warmup + for _ in range(5): + graph.replay() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + torch.cuda.synchronize() + start_event.record() + graph.replay() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + graph.reset() + return avg + + +@ray.remote(num_gpus=1) +class BenchmarkWorker: + def __init__(self, seed: int) -> None: + torch.set_default_device("cuda") + current_platform.seed_everything(seed) + self.seed = seed + # Get the device ID to allocate tensors and kernels + # on the respective GPU. This is required for Ray to work + # correctly with multi-GPU tuning on the ROCm platform. + self.device_id = int(ray.get_gpu_ids()[0]) + + def benchmark( + self, + num_tokens: int, + num_experts: int, + hidden_size: int, + topk: int, + dtype: torch.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + use_customized_permute: bool = False, + ) -> tuple[dict[str, int], float]: + current_platform.seed_everything(self.seed) + + permute_time = benchmark_permute( + num_tokens, + num_experts, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=100, + use_customized_permute=use_customized_permute, + ) + unpermute_time = benchmark_unpermute( + num_tokens, + num_experts, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + num_iters=100, + use_customized_permute=use_customized_permute, + ) + return permute_time, unpermute_time + + +def get_weight_block_size_safety(config, default_value=None): + quantization_config = getattr(config, "quantization_config", {}) + if isinstance(quantization_config, dict): + return quantization_config.get("weight_block_size", default_value) + return default_value + + +def main(args: argparse.Namespace): + print(args) + + config = AutoConfig.from_pretrained( + args.model, trust_remote_code=args.trust_remote_code + ) + if config.architectures[0] == "DbrxForCausalLM": + E = config.ffn_config.moe_num_experts + topk = config.ffn_config.moe_top_k + elif config.architectures[0] == "JambaForCausalLM": + E = config.num_experts + topk = config.num_experts_per_tok + elif ( + config.architectures[0] == "DeepseekV3ForCausalLM" + or config.architectures[0] == "DeepseekV2ForCausalLM" + ): + E = config.n_routed_experts + topk = config.num_experts_per_tok + elif config.architectures[0] in ["Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"]: + E = config.num_experts + topk = config.num_experts_per_tok + + else: + # Support for llama4 + config = config.get_text_config() + # Default: Mixtral. + E = config.num_local_experts + topk = config.num_experts_per_tok + + hidden_size = config.hidden_size + dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype + use_fp8_w8a8 = args.dtype == "fp8_w8a8" + use_int8_w8a16 = args.dtype == "int8_w8a16" + use_customized_permute = args.use_customized_permute + + if args.batch_size is None: + batch_sizes = [ + 1, + 2, + 4, + 8, + 16, + 24, + 32, + 48, + 64, + 96, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ] + else: + batch_sizes = [args.batch_size] + + ray.init() + num_gpus = int(ray.available_resources()["GPU"]) + workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)] + + def _distribute(method: str, inputs: list[Any]) -> list[Any]: + outputs = [] + worker_idx = 0 + for input_args in inputs: + worker = workers[worker_idx] + worker_method = getattr(worker, method) + output = worker_method.remote(*input_args) + outputs.append(output) + worker_idx = (worker_idx + 1) % num_gpus + return ray.get(outputs) + + outputs = _distribute( + "benchmark", + [ + ( + batch_size, + E, + hidden_size, + topk, + dtype, + use_fp8_w8a8, + use_int8_w8a16, + use_customized_permute, + ) + for batch_size in batch_sizes + ], + ) + + for batch_size, (permute, unpermute) in zip(batch_sizes, outputs): + print(f"Batch size: {batch_size}") + print(f"Permute time: {permute:.2f} us") + print(f"Unpermute time: {unpermute:.2f} us") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser() + parser.add_argument( + "--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1" + ) + parser.add_argument( + "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto" + ) + parser.add_argument("--use-customized-permute", action="store_true") + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--batch-size", type=int, required=False) + parser.add_argument("--trust-remote-code", action="store_true") + args = parser.parse_args() + + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_paged_attention.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_paged_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..7e0376c18ecc79690ca49e3995258c5e749bbac7 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_paged_attention.py @@ -0,0 +1,251 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import random +import time +from typing import Optional + +import torch + +from vllm import _custom_ops as ops +from vllm.logger import init_logger +from vllm.platforms import current_platform +from vllm.utils import ( + STR_DTYPE_TO_TORCH_DTYPE, + FlexibleArgumentParser, + create_kv_caches_with_random, +) + +logger = init_logger(__name__) + +NUM_BLOCKS = 128 * 1024 +PARTITION_SIZE = 512 +PARTITION_SIZE_ROCM = 256 + + +@torch.inference_mode() +def main( + version: str, + num_seqs: int, + seq_len: int, + num_query_heads: int, + num_kv_heads: int, + head_size: int, + use_alibi: bool, + block_size: int, + dtype: torch.dtype, + seed: int, + do_profile: bool, + device: str = "cuda", + kv_cache_dtype: Optional[str] = None, +) -> None: + current_platform.seed_everything(seed) + + scale = float(1.0 / (head_size**0.5)) + query = torch.empty( + num_seqs, num_query_heads, head_size, dtype=dtype, device=device + ) + query.uniform_(-scale, scale) + + assert num_query_heads % num_kv_heads == 0 + alibi_slopes = None + if use_alibi: + alibi_slopes = torch.randn(num_query_heads, dtype=torch.float, device=device) + + seq_lens = [seq_len for _ in range(num_seqs)] + max_seq_len = max(seq_lens) + seq_lens = torch.tensor(seq_lens, dtype=torch.int, device=device) + + # Create the block tables. + max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size + block_tables_lst: list[list[int]] = [] + for _ in range(num_seqs): + block_table = [ + random.randint(0, NUM_BLOCKS - 1) for _ in range(max_num_blocks_per_seq) + ] + block_tables_lst.append(block_table) + + block_tables = torch.tensor(block_tables_lst, dtype=torch.int, device=device) + + # Create the KV cache. + key_caches, value_caches = create_kv_caches_with_random( + NUM_BLOCKS, + block_size, + 1, + num_kv_heads, + head_size, + kv_cache_dtype, + dtype, + device=device, + ) + key_cache, value_cache = key_caches[0], value_caches[0] + + # Prepare for the paged attention kernel. + output = torch.empty_like(query) + if version == "v2": + if current_platform.is_rocm(): + global PARTITION_SIZE + if not args.custom_paged_attn and not current_platform.is_navi(): + PARTITION_SIZE = 1024 + else: + PARTITION_SIZE = PARTITION_SIZE_ROCM + num_partitions = (max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE + tmp_output = torch.empty( + size=(num_seqs, num_query_heads, num_partitions, head_size), + dtype=output.dtype, + device=output.device, + ) + exp_sums = torch.empty( + size=(num_seqs, num_query_heads, num_partitions), + dtype=torch.float32, + device=output.device, + ) + max_logits = torch.empty_like(exp_sums) + + def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: + torch.cuda.synchronize() + if profile: + torch.cuda.cudart().cudaProfilerStart() + start_time = time.perf_counter() + + # Using default kv_scale + k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device) + + for _ in range(num_iters): + if version == "v1": + ops.paged_attention_v1( + output, + query, + key_cache, + value_cache, + num_kv_heads, + scale, + block_tables, + seq_lens, + block_size, + max_seq_len, + alibi_slopes, + kv_cache_dtype, + k_scale, + v_scale, + ) + elif version == "v2": + if not args.custom_paged_attn: + ops.paged_attention_v2( + output, + exp_sums, + max_logits, + tmp_output, + query, + key_cache, + value_cache, + num_kv_heads, + scale, + block_tables, + seq_lens, + block_size, + max_seq_len, + alibi_slopes, + kv_cache_dtype, + k_scale, + v_scale, + ) + else: + ops.paged_attention_rocm( + output, + exp_sums, + max_logits, + tmp_output, + query, + key_cache, + value_cache, + num_kv_heads, + scale, + block_tables, + seq_lens, + None, + block_size, + max_seq_len, + alibi_slopes, + kv_cache_dtype, + k_scale, + v_scale, + ) + else: + raise ValueError(f"Invalid version: {version}") + torch.cuda.synchronize() + + end_time = time.perf_counter() + if profile: + torch.cuda.cudart().cudaProfilerStop() + return (end_time - start_time) / num_iters + + # Warmup. + print("Warming up...") + run_benchmark = run_cuda_benchmark + run_benchmark(num_iters=3, profile=False) + + # Benchmark. + if do_profile: + latency = run_benchmark(num_iters=1, profile=True) + else: + latency = run_benchmark(num_iters=100, profile=False) + print(f"Kernel running time: {latency * 1000000:.3f} us") + + +if __name__ == "__main__": + logger.warning( + "This script benchmarks the paged attention kernel. " + "By default this is no longer used in vLLM inference." + ) + + parser = FlexibleArgumentParser(description="Benchmark the paged attention kernel.") + parser.add_argument("--version", type=str, choices=["v1", "v2"], default="v2") + parser.add_argument("--batch-size", type=int, default=8) + parser.add_argument("--seq-len", type=int, default=4096) + parser.add_argument("--num-query-heads", type=int, default=64) + parser.add_argument("--num-kv-heads", type=int, default=8) + parser.add_argument( + "--head-size", + type=int, + choices=[64, 80, 96, 112, 120, 128, 192, 256], + default=128, + ) + parser.add_argument("--block-size", type=int, choices=[16, 32], default=16) + parser.add_argument("--use-alibi", action="store_true") + parser.add_argument( + "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half" + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--profile", action="store_true") + parser.add_argument( + "--kv-cache-dtype", + type=str, + choices=["auto", "fp8", "fp8_e5m2", "fp8_e4m3"], + default="auto", + help="Data type for kv cache storage. If 'auto', will use model " + "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. " + "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)", + ) + parser.add_argument( + "--custom-paged-attn", action="store_true", help="Use custom paged attention" + ) + args = parser.parse_args() + print(args) + + if args.num_query_heads % args.num_kv_heads != 0: + raise ValueError("num_query_heads must be divisible by num_kv_heads") + main( + version=args.version, + num_seqs=args.batch_size, + seq_len=args.seq_len, + num_query_heads=args.num_query_heads, + num_kv_heads=args.num_kv_heads, + head_size=args.head_size, + block_size=args.block_size, + use_alibi=args.use_alibi, + dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype], + seed=args.seed, + do_profile=args.profile, + kv_cache_dtype=args.kv_cache_dtype, + ) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_quant.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_quant.py new file mode 100644 index 0000000000000000000000000000000000000000..6ab26f5f1adf73ccdc68c0d5f20d622d0c09b7ba --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_quant.py @@ -0,0 +1,108 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import time + +import torch + +from vllm import _custom_ops as ops +from vllm.platforms import current_platform +from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser + + +@torch.inference_mode() +def main( + num_tokens: int, + hidden_size: int, + static_scale: bool, + quant_dtype: torch.dtype, + dtype: torch.dtype, + seed: int = 0, + do_profile: bool = False, + num_warmup_iters: int = 5, + num_iters: int = 100, +) -> None: + current_platform.seed_everything(seed) + torch.set_default_device("cuda") + + x = torch.randn(num_tokens, hidden_size, dtype=dtype) + scale = torch.randn(1, 1, dtype=torch.float32) if static_scale else None + + def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float: + torch.cuda.synchronize() + if profile: + torch.cuda.cudart().cudaProfilerStart() + start_time = time.perf_counter() + + for _ in range(num_iters): + if quant_dtype == torch.int8: + ops.scaled_int8_quant(x, scale) + else: + ops.scaled_fp8_quant(x, scale) + torch.cuda.synchronize() + + end_time = time.perf_counter() + if profile: + torch.cuda.cudart().cudaProfilerStop() + return (end_time - start_time) / num_iters + + # Warmup. + print("Warming up...") + run_benchmark = run_cuda_benchmark + run_benchmark(num_iters=num_warmup_iters, profile=False) + + # Benchmark. + if do_profile: + latency = run_benchmark(num_iters=1, profile=True) + else: + latency = run_benchmark(num_iters=num_iters, profile=False) + print(f"Kernel running time: {latency * 1000000:.3f} us") + + +if __name__ == "__main__": + + def to_torch_dtype(dt): + if dt == "int8": + return torch.int8 + if dt == "fp8": + return torch.float8_e4m3fn + raise ValueError(f"Unsupported dtype: {dt}") + + parser = FlexibleArgumentParser( + description="Benchmark the quantization (fp8 or int8) kernel." + ) + parser.add_argument("--num-tokens", type=int, default=4096) + parser.add_argument("--hidden-size", type=int, default=8192) + parser.add_argument("--static-scale", action="store_true") + parser.add_argument( + "--quant-dtype", type=str, choices=["fp8", "int8"], default="int8" + ) + parser.add_argument( + "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half" + ) + + parser.add_argument("--seed", type=int, default=0) + parser.add_argument("--profile", action="store_true") + parser.add_argument("--num-warmup-iters", type=int, default=5) + parser.add_argument( + "--num-iters", + type=int, + default=100, + help="Number of benchmark iterations. " + "If --profile is set, this number is ignored", + ) + + args = parser.parse_args() + print(args) + + main( + num_tokens=args.num_tokens, + hidden_size=args.hidden_size, + static_scale=args.static_scale, + quant_dtype=to_torch_dtype(args.quant_dtype), + dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype], + seed=args.seed, + do_profile=args.profile, + num_warmup_iters=args.num_warmup_iters, + num_iters=args.num_iters, + ) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_rmsnorm.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_rmsnorm.py new file mode 100644 index 0000000000000000000000000000000000000000..4cf633a81358d80202fa383d7312dac4a818ea62 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_rmsnorm.py @@ -0,0 +1,256 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import itertools +from typing import Optional, Union + +import torch +from flashinfer.norm import fused_add_rmsnorm, rmsnorm +from torch import nn + +from vllm import _custom_ops as vllm_ops +from vllm.triton_utils import triton + + +class HuggingFaceRMSNorm(nn.Module): + def __init__(self, hidden_size: int, eps: float = 1e-6) -> None: + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward( + self, + x: torch.Tensor, + residual: Optional[torch.Tensor] = None, + ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]: + orig_dtype = x.dtype + x = x.to(torch.float32) + if residual is not None: + x = x + residual.to(torch.float32) + residual = x.to(orig_dtype) + + variance = x.pow(2).mean(dim=-1, keepdim=True) + x = x * torch.rsqrt(variance + self.variance_epsilon) + x = x.to(orig_dtype) * self.weight + if residual is None: + return x + else: + return x, residual + + +def rmsnorm_naive( + x: torch.Tensor, + weight: torch.Tensor, + residual: Optional[torch.Tensor] = None, + eps: float = 1e-6, +): + naive_norm = HuggingFaceRMSNorm(x.shape[-1], eps=eps) + naive_norm.weight = nn.Parameter(weight) + naive_norm = naive_norm.to(x.device) + + orig_shape = x.shape + x = x.view(-1, x.shape[-1]) + if residual is not None: + residual = residual.view(-1, residual.shape[-1]) + + output = naive_norm(x, residual) + + if isinstance(output, tuple): + output = (output[0].view(orig_shape), output[1].view(orig_shape)) + else: + output = output.view(orig_shape) + return output + + +def rmsnorm_flashinfer( + x: torch.Tensor, + weight: torch.Tensor, + residual: Optional[torch.Tensor] = None, + eps: float = 1e-6, +): + orig_shape = x.shape + x = x.view(-1, x.shape[-1]) + if residual is not None: + residual = residual.view(-1, residual.shape[-1]) + + if residual is not None: + fused_add_rmsnorm(x, residual, weight, eps) + output = (x, residual) + else: + output = rmsnorm(x, weight, eps) + + if isinstance(output, tuple): + output = (output[0].view(orig_shape), output[1].view(orig_shape)) + else: + output = output.view(orig_shape) + return output + + +def rmsnorm_vllm( + x: torch.Tensor, + weight: torch.Tensor, + residual: Optional[torch.Tensor] = None, + eps: float = 1e-6, +): + orig_shape = x.shape + x = x.view(-1, x.shape[-1]) + if residual is not None: + residual = residual.view(-1, residual.shape[-1]) + + if residual is not None: + vllm_ops.fused_add_rms_norm(x, residual, weight, eps) + output = (x, residual) + else: + out = torch.empty_like(x) + vllm_ops.rms_norm(out, x, weight, eps) + output = out + + if isinstance(output, tuple): + output = (output[0].view(orig_shape), output[1].view(orig_shape)) + else: + output = output.view(orig_shape) + return output + + +def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True): + dtype = torch.bfloat16 + x = torch.randn(batch_size, seq_len, hidden_size, dtype=dtype, device="cuda") + weight = torch.ones(hidden_size, dtype=dtype, device="cuda") + residual = torch.randn_like(x) if use_residual else None + + output_naive = rmsnorm_naive( + x.clone(), weight, residual.clone() if residual is not None else None + ) + output_flashinfer = rmsnorm_flashinfer( + x.clone(), weight, residual.clone() if residual is not None else None + ) + output_vllm = rmsnorm_vllm( + x.clone(), weight, residual.clone() if residual is not None else None + ) + + if use_residual: + output_naive = output_naive[0] + output_flashinfer = output_flashinfer[0] + output_vllm = output_vllm[0] + + print(f"Naive output={output_naive}") + print(f"FlashInfer output={output_flashinfer}") + print(f"vLLM output={output_vllm}") + + if torch.allclose( + output_naive, output_flashinfer, atol=1e-2, rtol=1e-2 + ) and torch.allclose(output_naive, output_vllm, atol=1e-2, rtol=1e-2): + print("✅ All implementations match") + else: + print("❌ Implementations differ") + + +batch_size_range = [2**i for i in range(0, 7, 2)] +seq_length_range = [2**i for i in range(6, 11, 1)] +head_num_range = [32, 48] +configs = list(itertools.product(head_num_range, batch_size_range, seq_length_range)) + + +def get_benchmark(use_residual): + @triton.testing.perf_report( + triton.testing.Benchmark( + x_names=["head_num", "batch_size", "seq_len"], + x_vals=[list(_) for _ in configs], + line_arg="provider", + line_vals=["huggingface", "flashinfer", "vllm"], + line_names=["HuggingFace", "FlashInfer", "vLLM"], + styles=[("blue", "-"), ("green", "-"), ("red", "-")], + ylabel="us", + plot_name=f"rmsnorm-perf-{'with' if use_residual else 'without'}-residual", + args={}, + ) + ) + def benchmark(head_num, batch_size, seq_len, provider): + dtype = torch.bfloat16 + hidden_size = head_num * 128 # assuming head_dim = 128 + + x = torch.randn(batch_size, seq_len, hidden_size, dtype=dtype, device="cuda") + weight = torch.ones(hidden_size, dtype=dtype, device="cuda") + residual = torch.randn_like(x) if use_residual else None + + quantiles = [0.5, 0.2, 0.8] + + if provider == "huggingface": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: rmsnorm_naive( + x.clone(), + weight, + residual.clone() if residual is not None else None, + ), + quantiles=quantiles, + ) + elif provider == "flashinfer": + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: rmsnorm_flashinfer( + x.clone(), + weight, + residual.clone() if residual is not None else None, + ), + quantiles=quantiles, + ) + else: + ms, min_ms, max_ms = triton.testing.do_bench( + lambda: rmsnorm_vllm( + x.clone(), + weight, + residual.clone() if residual is not None else None, + ), + quantiles=quantiles, + ) + + return 1000 * ms, 1000 * max_ms, 1000 * min_ms + + return benchmark + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument( + "--batch-size", + type=int, + default=4, + help="Batch size", + ) + parser.add_argument( + "--seq-len", + type=int, + default=128, + help="Sequence length", + ) + parser.add_argument( + "--hidden-size", + type=int, + default=4096, + help="Hidden size (2nd dimension) of the sequence", + ) + parser.add_argument( + "--use-residual", action="store_true", help="Whether to use residual connection" + ) + parser.add_argument( + "--save-path", + type=str, + default="./configs/rmsnorm/", + help="Path to save rmsnorm benchmark results", + ) + + args = parser.parse_args() + + # Run correctness test + calculate_diff( + batch_size=args.batch_size, + seq_len=args.seq_len, + hidden_size=args.hidden_size, + use_residual=args.use_residual, + ) + + # Get the benchmark function with proper use_residual setting + benchmark = get_benchmark(args.use_residual) + # Run performance benchmark + benchmark.run(print_data=True, save_path=args.save_path) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_rope.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_rope.py new file mode 100644 index 0000000000000000000000000000000000000000..b81baf17a8c674edd72fe9a6397a0b027a8d33b9 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_rope.py @@ -0,0 +1,133 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +from itertools import accumulate +from typing import Optional + +import nvtx +import torch + +from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope +from vllm.platforms import current_platform +from vllm.utils import FlexibleArgumentParser + + +def benchmark_rope_kernels_multi_lora( + is_neox_style: bool, + batch_size: int, + seq_len: int, + num_heads: int, + head_size: int, + rotary_dim: Optional[int], + dtype: torch.dtype, + seed: int, + device: str, + max_position: int = 8192, + base: float = 10000, +) -> None: + current_platform.seed_everything(seed) + torch.set_default_device(device) + if rotary_dim is None: + rotary_dim = head_size + # silulating serving 4 LoRAs + scaling_factors = [1, 2, 4, 8] + # batched RoPE can take multiple scaling factors + batched_rope = get_rope( + head_size, + rotary_dim, + max_position, + base, + is_neox_style, + {"rope_type": "linear", "factor": tuple(scaling_factors)}, + ) + # non-batched RoPE takes only one scaling factor, we create multiple + # instances to simulate the same behavior + non_batched_ropes: list[RotaryEmbedding] = [] + for scaling_factor in scaling_factors: + non_batched_ropes.append( + get_rope( + head_size, + rotary_dim, + max_position, + base, + is_neox_style, + {"rope_type": "linear", "factor": (scaling_factor,)}, + ) + ) + + positions = torch.randint(0, max_position, (batch_size, seq_len)) + query = torch.randn(batch_size, seq_len, num_heads * head_size, dtype=dtype) + key = torch.randn_like(query) + + # create query offsets for batched RoPE, we concat multiple kv cache + # together and each query needs to find the right kv cache of its type + offset_map = torch.tensor( + list( + accumulate( + [0] + + [ + max_position * scaling_factor * 2 + for scaling_factor in scaling_factors[:-1] + ] + ) + ) + ) + query_types = torch.randint( + 0, len(scaling_factors), (batch_size, seq_len), device=device + ) + # map query types to offsets + query_offsets = offset_map[query_types] + # the kernel takes flattened offsets + flatten_offsets = query_offsets.flatten() + + # batched queries of the same type together for non-batched RoPE + queries = [query[query_types == i] for i in range(len(scaling_factors))] + keys = [key[query_types == i] for i in range(len(scaling_factors))] + packed_qkr = zip(queries, keys, non_batched_ropes) + # synchronize before start timing + torch.cuda.synchronize() + with nvtx.annotate("non-batched", color="yellow"): + for q, k, r in packed_qkr: + r.forward(positions, q, k) + torch.cuda.synchronize() + with nvtx.annotate("batched", color="green"): + batched_rope.forward(positions, query, key, flatten_offsets) + torch.cuda.synchronize() + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark the rotary embedding kernels." + ) + parser.add_argument("--is-neox-style", type=bool, default=True) + parser.add_argument("--batch-size", type=int, default=16) + parser.add_argument("--seq-len", type=int, default=512) + parser.add_argument("--num-heads", type=int, default=8) + parser.add_argument( + "--head-size", + type=int, + choices=[64, 80, 96, 112, 120, 128, 192, 256], + default=128, + ) + parser.add_argument("--rotary-dim", type=int, choices=[16, 32], default=32) + parser.add_argument( + "--dtype", type=str, choices=["bfloat16", "float"], default="float" + ) + parser.add_argument("--seed", type=int, default=0) + parser.add_argument( + "--device", type=str, choices=["cuda:0", "cuda:1"], default="cuda:0" + ) + args = parser.parse_args() + print(args) + + benchmark_rope_kernels_multi_lora( + is_neox_style=args.is_neox_style, + batch_size=args.batch_size, + seq_len=args.seq_len, + num_heads=args.num_heads, + head_size=args.head_size, + rotary_dim=args.rotary_dim, + dtype=getattr(torch, args.dtype), + seed=args.seed, + device=args.device, + ) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_shapes.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_shapes.py new file mode 100644 index 0000000000000000000000000000000000000000..18c459c31d3f84d1ca7b7e31d00d50b521177e96 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_shapes.py @@ -0,0 +1,94 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +WEIGHT_SHAPES = { + "ideal": [[4 * 256 * 32, 256 * 32]], + "mistralai/Mistral-7B-v0.1/TP1": [ + [4096, 6144], + [4096, 4096], + [4096, 28672], + [14336, 4096], + ], + "mistralai/Mistral-7B-v0.1/TP2": [ + [4096, 3072], + [2048, 4096], + [4096, 14336], + [7168, 4096], + ], + "mistralai/Mistral-7B-v0.1/TP4": [ + [4096, 1536], + [1024, 4096], + [4096, 7168], + [3584, 4096], + ], + "meta-llama/Llama-2-7b-hf/TP1": [ + [4096, 12288], + [4096, 4096], + [4096, 22016], + [11008, 4096], + ], + "meta-llama/Llama-2-7b-hf/TP2": [ + [4096, 6144], + [2048, 4096], + [4096, 11008], + [5504, 4096], + ], + "meta-llama/Llama-2-7b-hf/TP4": [ + [4096, 3072], + [1024, 4096], + [4096, 5504], + [2752, 4096], + ], + "meta-llama/Llama-2-13b-hf/TP1": [ + [5120, 15360], + [5120, 5120], + [5120, 27648], + [13824, 5120], + ], + "meta-llama/Llama-2-13b-hf/TP2": [ + [5120, 7680], + [2560, 5120], + [5120, 13824], + [6912, 5120], + ], + "meta-llama/Llama-2-13b-hf/TP4": [ + [5120, 3840], + [1280, 5120], + [5120, 6912], + [3456, 5120], + ], + "meta-llama/Llama-2-70b-hf/TP1": [ + [8192, 10240], + [8192, 8192], + [8192, 57344], + [28672, 8192], + ], + "meta-llama/Llama-2-70b-hf/TP2": [ + [8192, 5120], + [4096, 8192], + [8192, 28672], + [14336, 8192], + ], + "meta-llama/Llama-2-70b-hf/TP4": [ + [8192, 2560], + [2048, 8192], + [8192, 14336], + [7168, 8192], + ], +} + +WEIGHT_SHAPES_MOE = { + "nm-testing/Mixtral-8x7B-Instruct-v0.1": [ + [8, 2, 4096, 28672], + [8, 2, 14336, 4096], + ], + "nm-testing/deepseekv2-lite": [ + [64, 6, 2048, 1408], + ], + "ibm-granite/granite-3.0-1b-a400m": [ + [32, 8, 1024, 1024], + ], + "ibm-granite/granite-3.0-3b-a800m": [ + [40, 8, 1024, 1536], + ], +} diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_w8a8_block_fp8.py b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_w8a8_block_fp8.py new file mode 100644 index 0000000000000000000000000000000000000000..4fcdbadd65ecd366f51ce5d9053cc700854dcb71 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/benchmark_w8a8_block_fp8.py @@ -0,0 +1,414 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +# Adapted from sglang quantization/tuning_block_wise_kernel.py + +import argparse +import json +import multiprocessing as mp +import os +import time +from datetime import datetime +from typing import Any + +import torch +import tqdm +import triton + +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + _w8a8_block_fp8_matmul, +) +from vllm.platforms import current_platform +from vllm.utils import FlexibleArgumentParser + +mp.set_start_method("spawn", force=True) + +assert current_platform.is_cuda(), ( + "Only support tune w8a8 block fp8 kernel on CUDA device." +) + +DTYPE_MAP = { + "float32": torch.float32, + "float16": torch.float16, + "half": torch.half, + "bfloat16": torch.bfloat16, +} + + +def w8a8_block_matmul( + A: torch.Tensor, + B: torch.Tensor, + As: torch.Tensor, + Bs: torch.Tensor, + block_size: list[int], + config: dict[str, Any], + output_dtype: torch.dtype = torch.float16, +) -> torch.Tensor: + """This function performs matrix multiplication with + block-wise quantization. + + It takes two input tensors `A` and `B` with scales `As` and `Bs`. + The output is returned in the specified `output_dtype`. + + Args: + A: The input tensor, e.g., activation. + B: The input tensor, e.g., weight. + As: The per-token-group quantization scale for `A`. + Bs: The per-block quantization scale for `B`. + block_size: The block size for per-block quantization. + It should be 2-dim, e.g., [128, 128]. + output_dytpe: The dtype of the returned tensor. + + Returns: + torch.Tensor: The result of matmul. + """ + assert len(block_size) == 2 + block_n, block_k = block_size[0], block_size[1] + + assert A.shape[-1] == B.shape[-1] + assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous() + assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1] + M = A.numel() // A.shape[-1] + + assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2 + N, K = B.shape + assert triton.cdiv(N, block_n) == Bs.shape[0] + assert triton.cdiv(K, block_k) == Bs.shape[1] + + C_shape = A.shape[:-1] + (N,) + C = A.new_empty(C_shape, dtype=output_dtype) + + def grid(META): + return ( + triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]), + ) + + if A.dtype == torch.float8_e4m3fn: + kernel = _w8a8_block_fp8_matmul + else: + raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.") + + kernel[grid]( + A, + B, + C, + As, + Bs, + M, + N, + K, + block_n, + block_k, + A.stride(-2), + A.stride(-1), + B.stride(1), + B.stride(0), + C.stride(-2), + C.stride(-1), + As.stride(-2), + As.stride(-1), + Bs.stride(1), + Bs.stride(0), + **config, + ) + + return C + + +def get_configs_compute_bound(): + configs = [] + for num_stages in [2, 3, 4, 5]: + for block_m in [16, 32, 64, 128, 256]: + for block_k in [64, 128]: + for block_n in [32, 64, 128, 256]: + for num_warps in [4, 8]: + for group_size in [1, 16, 32, 64]: + configs.append( + { + "BLOCK_SIZE_M": block_m, + "BLOCK_SIZE_N": block_n, + "BLOCK_SIZE_K": block_k, + "GROUP_SIZE_M": group_size, + "num_warps": num_warps, + "num_stages": num_stages, + } + ) + return configs + + +def get_weight_shapes(tp_size): + # NOTE(HandH1998): The weight shapes only works for DeepSeek-V3. + # Modify them, if you tune for another different model. + # cannot TP + total = [ + (512 + 64, 7168), + ((128 + 64) * 128, 7168), + (128 * (128 + 128), 512), + (7168, 16384), + (7168, 18432), + ] + # N can TP + n_tp = [ + (18432 * 2, 7168), + ((128 + 64) * 128, 7168), + (128 * (128 + 128), 512), + (24576, 1536), + (12288, 7168), + (4096, 7168), + ] + # K can TP + k_tp = [(7168, 18432), (7168, 16384), (7168, 2048)] + + weight_shapes = [] + for t in total: + weight_shapes.append(t) + for n_t in n_tp: + new_t = (n_t[0] // tp_size, n_t[1]) + weight_shapes.append(new_t) + for k_t in k_tp: + new_t = (k_t[0], k_t[1] // tp_size) + weight_shapes.append(new_t) + return weight_shapes + + +def benchmark_config( + A, B, As, Bs, block_size, config, out_dtype=torch.float16, num_iters=10 +): + def run(): + w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype) + + torch.cuda.synchronize() + # JIT complication & warmup + for _ in range(5): + run() + torch.cuda.synchronize() + + start_event = torch.cuda.Event(enable_timing=True) + end_event = torch.cuda.Event(enable_timing=True) + + latencies: list[float] = [] + for i in range(num_iters): + torch.cuda.synchronize() + start_event.record() + run() + end_event.record() + end_event.synchronize() + latencies.append(start_event.elapsed_time(end_event)) + avg = sum(latencies) / (num_iters * 10) * 1000 # us + return avg + + +def tune(M, N, K, block_size, out_dtype, search_space, input_type): + factor_for_scale = 1e-2 + + if input_type == "fp8": + fp8_info = torch.finfo(torch.float8_e4m3fn) + fp8_max, fp8_min = fp8_info.max, fp8_info.min + + A_fp32 = ( + (torch.rand(M, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max + ) + A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) + + B_fp32 = ( + (torch.rand(N, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max + ) + B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) + else: + raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.") + + block_n, block_k = block_size[0], block_size[1] + n_tiles = (N + block_n - 1) // block_n + k_tiles = (K + block_k - 1) // block_k + + As = torch.rand(M, k_tiles, dtype=torch.float32, device="cuda") * factor_for_scale + Bs = ( + torch.rand(n_tiles, k_tiles, dtype=torch.float32, device="cuda") + * factor_for_scale + ) + + best_config = None + best_time = float("inf") + for config in tqdm(search_space): + try: + kernel_time = benchmark_config( + A, + B, + As, + Bs, + block_size, + config, + out_dtype, + num_iters=10, + ) + except triton.runtime.autotuner.OutOfResources: + # Some configurations may be invalid and fail to compile. + continue + + if kernel_time < best_time: + best_time = kernel_time + best_config = config + now = datetime.now() + print(f"{now.ctime()}] Completed tuning for batch_size={M}") + assert best_config is not None + return best_config + + +def save_configs( + N, + K, + block_n, + block_k, + configs, + save_path, + input_type="fp8", +) -> None: + os.makedirs(save_path, exist_ok=True) + device_name = current_platform.get_device_name().replace(" ", "_") + json_file_name = ( + f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8," + f"block_shape=[{block_n},{block_k}].json" + ) + + config_file_path = os.path.join(save_path, json_file_name) + print(f"Writing best config to {config_file_path}...") + + with open(config_file_path, "w") as f: + json.dump(configs, f, indent=4) + f.write("\n") + + +def tune_on_gpu(args_dict): + """Run tuning on a specific GPU.""" + gpu_id = args_dict["gpu_id"] + batch_sizes = args_dict["batch_sizes"] + weight_shapes = args_dict["weight_shapes"] + args = args_dict["args"] + + torch.cuda.set_device(gpu_id) + print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}") + + block_n = args.block_n + block_k = args.block_k + out_dtype = DTYPE_MAP[args.out_dtype] + save_path = args.save_path + input_type = args.input_type + + search_space = get_configs_compute_bound() + search_space = [ + config for config in search_space if block_k % config["BLOCK_SIZE_K"] == 0 + ] + + start = time.time() + for shape in tqdm(weight_shapes, desc=f"GPU {gpu_id} - Shapes"): + N, K = shape[0], shape[1] + print(f"[GPU {gpu_id}] Tune for weight shape of `N: {N}, K: {K}`") + benchmark_results = [ + tune( + batch_size, + N, + K, + [block_n, block_k], + out_dtype, + search_space, + input_type, + ) + for batch_size in tqdm(batch_sizes, desc=f"GPU {gpu_id} - Batch sizes") + ] + best_configs = {M: config for M, config in zip(batch_sizes, benchmark_results)} + save_configs(N, K, block_n, block_k, best_configs, save_path, input_type) + + end = time.time() + print(f"Tuning on GPU {gpu_id} took {end - start:.2f} seconds") + + +def distribute_batch_sizes(batch_sizes, num_gpus): + """Distribute batch sizes across available GPUs.""" + batches_per_gpu = [] + for i in range(num_gpus): + start_idx = i * len(batch_sizes) // num_gpus + end_idx = (i + 1) * len(batch_sizes) // num_gpus + batches_per_gpu.append(batch_sizes[start_idx:end_idx]) + return batches_per_gpu + + +def main(args): + print(args) + num_gpus = torch.cuda.device_count() + if num_gpus == 0: + raise RuntimeError("No GPU available for tuning") + print(f"Found {num_gpus} GPUs for parallel tuning") + + torch.cuda.init() + + if args.batch_size is None: + batch_sizes = [ + 1, + 2, + 4, + 8, + 16, + 24, + 32, + 48, + 64, + 96, + 128, + 256, + 512, + 1024, + 1536, + 2048, + 3072, + 4096, + ] + else: + batch_sizes = [args.batch_size] + num_gpus = 1 # If only one batch size, use only one GPU + + weight_shapes = get_weight_shapes(args.tp_size) + + batches_per_gpu = distribute_batch_sizes(batch_sizes, num_gpus) + + process_args = [] + for gpu_id in range(num_gpus): + process_args.append( + { + "gpu_id": gpu_id, + "batch_sizes": batches_per_gpu[gpu_id], + "weight_shapes": weight_shapes, # Each GPU processes all weight shapes + "args": args, + } + ) + + ctx = mp.get_context("spawn") + with ctx.Pool(num_gpus) as pool: + pool.map(tune_on_gpu, process_args) + + print("Multi-GPU tuning completed") + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description=""" +Tune triton w8a8 block fp8 for DeepSeek-V3/DeepSeek-R1: + python3 benchmark_w8a8_block_fp8.py --tp-size 8 --input-type fp8 +Then copy to model_executor/layers/quantization/utils/configs + """, + formatter_class=argparse.RawTextHelpFormatter, + ) + + parser.add_argument("--tp-size", "-tp", type=int, default=8) + parser.add_argument("--input-type", type=str, choices=["fp8"], default="fp8") + parser.add_argument( + "--out-dtype", + type=str, + choices=["float32", "float16", "bfloat16", "half"], + default="float16", + ) + parser.add_argument("--block-n", type=int, default=128) + parser.add_argument("--block-k", type=int, default=128) + parser.add_argument("--batch-size", type=int, required=False) + parser.add_argument("--save-path", type=str, default="./") + args = parser.parse_args() + + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/deepgemm/README.md b/online_apiserver_test_maxbs/benchmarks/kernels/deepgemm/README.md new file mode 100644 index 0000000000000000000000000000000000000000..917e814010f894cd812e42849feb5822ad332d04 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/deepgemm/README.md @@ -0,0 +1,129 @@ +# DeepSeek DeepGEMM Kernels Benchmark + +This directory includes benchmarks between DeepSeek's DeepGEMM block fp8 kernels against vLLM's existing triton and CUTLASS-based kernels. + +Currently this just includes dense GEMMs and only works on Hopper GPUs. + +## Setup + +You need to install vLLM in your usual fashion, then install DeepGEMM from source in its own directory: + +``` +git clone --recursive https://github.com/deepseek-ai/DeepGEMM +cd DeepGEMM +python setup.py install +uv pip install -e . +``` + +## Usage + +``` +python benchmark_fp8_block_dense_gemm.py +INFO 02-26 21:55:13 [__init__.py:207] Automatically detected platform cuda. +===== STARTING FP8 GEMM BENCHMARK ===== +PyTorch version: 2.5.1+cu124 +CUDA version: 12.4 +Triton version: 3.1.0 +Using device: NVIDIA H100 80GB HBM3 +WARNING 02-26 21:55:15 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +INFO 02-26 21:55:15 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel. +WARNING 02-26 21:55:16 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=18432,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +WARNING 02-26 21:55:17 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel. +INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel. + +===== PERFORMANCE COMPARISON ===== + +DeepGEMM Implementation: ++------+-------+-------+-----------+--------+--------+ +| m | n | k | Time (μs) | TFLOPS | GB/s | ++------+-------+-------+-----------+--------+--------+ +| 8 | 4096 | 7168 | 102.9 | 4.6 | 286.4 | +| 8 | 7168 | 18432 | 70.8 | 29.8 | 1868.8 | +| 8 | 18432 | 7168 | 69.3 | 30.5 | 1911.8 | +| 64 | 4096 | 7168 | 69.1 | 54.4 | 439.0 | +| 64 | 7168 | 18432 | 69.4 | 243.6 | 1933.6 | +| 64 | 18432 | 7168 | 70.4 | 240.3 | 1917.2 | +| 64 | 24576 | 1536 | 70.1 | 68.9 | 584.6 | +| 64 | 32768 | 512 | 68.4 | 31.4 | 307.1 | +| 64 | 7168 | 16384 | 69.5 | 216.3 | 1718.5 | +| 128 | 4096 | 7168 | 141.1 | 53.3 | 222.1 | +| 128 | 7168 | 18432 | 71.9 | 470.5 | 1896.1 | +| 128 | 18432 | 7168 | 69.3 | 488.2 | 1988.2 | +| 1024 | 4096 | 7168 | 89.7 | 670.1 | 502.5 | +| 1024 | 18432 | 7168 | 279.0 | 969.8 | 635.2 | +| 2048 | 4096 | 7168 | 175.1 | 687.0 | 347.4 | +| 4096 | 4096 | 7168 | 335.4 | 717.0 | 275.1 | ++------+-------+-------+-----------+--------+--------+ + +vLLM Triton Implementation: ++------+-------+-------+-----------+--------+--------+--------------+ +| m | n | k | Time (μs) | TFLOPS | GB/s | vs DeepGEMM | ++------+-------+-------+-----------+--------+--------+--------------+ +| 8 | 4096 | 7168 | 74.0 | 6.3 | 398.2 | 1.39x faster | +| 8 | 7168 | 18432 | 89.6 | 23.6 | 1478.1 | 0.79x slower | +| 8 | 18432 | 7168 | 113.2 | 18.7 | 1170.4 | 0.61x slower | +| 64 | 4096 | 7168 | 79.4 | 47.3 | 382.2 | 0.87x slower | +| 64 | 7168 | 18432 | 98.5 | 171.7 | 1363.0 | 0.70x slower | +| 64 | 18432 | 7168 | 119.5 | 141.5 | 1129.4 | 0.59x slower | +| 64 | 24576 | 1536 | 37.6 | 128.4 | 1089.7 | 1.86x faster | +| 64 | 32768 | 512 | 38.7 | 55.5 | 542.6 | 1.77x faster | +| 64 | 7168 | 16384 | 86.1 | 174.5 | 1386.4 | 0.81x slower | +| 128 | 4096 | 7168 | 90.7 | 82.9 | 345.4 | 1.56x faster | +| 128 | 7168 | 18432 | 144.0 | 234.9 | 946.9 | 0.50x slower | +| 128 | 18432 | 7168 | 229.5 | 147.4 | 600.1 | 0.30x slower | +| 1024 | 4096 | 7168 | 242.3 | 248.2 | 186.1 | 0.37x slower | +| 1024 | 18432 | 7168 | 897.8 | 301.4 | 197.4 | 0.31x slower | +| 2048 | 4096 | 7168 | 463.0 | 259.7 | 131.4 | 0.38x slower | +| 4096 | 4096 | 7168 | 901.8 | 266.7 | 102.3 | 0.37x slower | ++------+-------+-------+-----------+--------+--------+--------------+ + +vLLM CUTLASS Implementation: ++------+-------+-------+-----------+--------+--------+--------------+--------------+ +| m | n | k | Time (μs) | TFLOPS | GB/s | vs DeepGEMM | vs Triton | ++------+-------+-------+-----------+--------+--------+--------------+--------------+ +| 8 | 4096 | 7168 | 34.6 | 13.6 | 852.3 | 2.98x faster | 2.14x faster | +| 8 | 7168 | 18432 | 78.9 | 26.8 | 1677.3 | 0.90x slower | 1.13x faster | +| 8 | 18432 | 7168 | 81.2 | 26.0 | 1631.1 | 0.85x slower | 1.39x faster | +| 64 | 4096 | 7168 | 36.9 | 101.9 | 822.9 | 1.87x faster | 2.15x faster | +| 64 | 7168 | 18432 | 87.4 | 193.4 | 1535.2 | 0.79x slower | 1.13x faster | +| 64 | 18432 | 7168 | 85.0 | 199.0 | 1587.6 | 0.83x slower | 1.41x faster | +| 64 | 24576 | 1536 | 28.0 | 172.8 | 1465.8 | 2.51x faster | 1.35x faster | +| 64 | 32768 | 512 | 28.8 | 74.5 | 728.5 | 2.37x faster | 1.34x faster | +| 64 | 7168 | 16384 | 77.9 | 193.0 | 1532.8 | 0.89x slower | 1.11x faster | +| 128 | 4096 | 7168 | 39.1 | 192.4 | 802.0 | 3.61x faster | 2.32x faster | +| 128 | 7168 | 18432 | 93.7 | 360.8 | 1454.2 | 0.77x slower | 1.54x faster | +| 128 | 18432 | 7168 | 85.7 | 394.8 | 1608.0 | 0.81x slower | 2.68x faster | +| 1024 | 4096 | 7168 | 99.7 | 603.1 | 452.2 | 0.90x slower | 2.43x faster | +| 1024 | 18432 | 7168 | 331.3 | 816.7 | 534.9 | 0.84x slower | 2.71x faster | +| 2048 | 4096 | 7168 | 198.3 | 606.6 | 306.7 | 0.88x slower | 2.34x faster | +| 4096 | 4096 | 7168 | 392.2 | 613.2 | 235.3 | 0.86x slower | 2.30x faster | ++------+-------+-------+-----------+--------+--------+--------------+--------------+ + +===== AVERAGE PERFORMANCE ===== ++----------------+------------+----------+---------------+ +| Implementation | Avg TFLOPS | Avg GB/s | Avg Time (ms) | ++----------------+------------+----------+---------------+ +| DeepGEMM | 310.98 | 1052.10 | 0.11 | +| vLLM Triton | 144.30 | 715.60 | 0.23 | +| vLLM CUTLASS | 286.78 | 1076.67 | 0.11 | ++----------------+------------+----------+---------------+ + +===== AVERAGE SPEEDUPS ===== ++-----------------------------+--------------+ +| Comparison | Speedup | ++-----------------------------+--------------+ +| DeepGEMM vs vLLM Triton | 1.71x faster | +| DeepGEMM vs vLLM CUTLASS | 0.94x slower | +| vLLM CUTLASS vs vLLM Triton | 1.84x faster | ++-----------------------------+--------------+ + +===== ACCURACY COMPARISON ===== ++----------------+-----------------------+ +| Implementation | Avg Diff vs Reference | ++----------------+-----------------------+ +| DeepGEMM | 0.000684 | +| vLLM Triton | 0.000684 | +| vLLM CUTLASS | 0.000684 | ++----------------+-----------------------+ +``` diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py b/online_apiserver_test_maxbs/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py new file mode 100644 index 0000000000000000000000000000000000000000..43c54d56ca8c1be027bec0d5f54dd3e53a294771 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py @@ -0,0 +1,456 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +# fmt: off +# ruff: noqa: E501 +import time + +# Import DeepGEMM functions +import deep_gemm +import torch +from deep_gemm import calc_diff, ceil_div, get_col_major_tma_aligned_tensor + +# Import vLLM functions +from vllm import _custom_ops as ops +from vllm.model_executor.layers.quantization.utils.fp8_utils import ( + per_token_group_quant_fp8, + w8a8_block_fp8_matmul, +) +from vllm.triton_utils import triton + + +# Copied from +# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L9 +def per_token_cast_to_fp8( + x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """Convert tensor to FP8 format with per-token scaling.""" + assert x.dim() == 2 and x.size(1) % 128 == 0 + m, n = x.shape + x_view = x.view(m, -1, 128) + x_amax = x_view.abs().float().amax(dim=2).view(m, -1).clamp(1e-4) + return (x_view * (448.0 / x_amax.unsqueeze(2))).to( + torch.float8_e4m3fn).view(m, n), (x_amax / 448.0).view(m, -1) + + +# Copied from +# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L17 +def per_block_cast_to_fp8( + x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """Convert tensor to FP8 format with per-block scaling.""" + assert x.dim() == 2 + m, n = x.shape + x_padded = torch.zeros((ceil_div(m, 128) * 128, ceil_div(n, 128) * 128), + dtype=x.dtype, + device=x.device) + x_padded[:m, :n] = x + x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128) + x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4) + x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn) + return x_scaled.view_as(x_padded)[:m, :n].contiguous(), ( + x_amax / 448.0).view(x_view.size(0), x_view.size(2)) + + +def benchmark_shape(m: int, + n: int, + k: int, + warmup: int = 100, + repeat: int = 10000, + verbose: bool = False) -> dict: + """Benchmark all implementations for a specific (m, n, k) shape.""" + if verbose: + print(f"\n=== Benchmarking shape: m={m}, n={n}, k={k} ===") + + # Create test tensors + A = torch.randn((m, k), device='cuda', dtype=torch.bfloat16) + B = torch.randn((n, k), device='cuda', dtype=torch.bfloat16) + + # Reference result in BF16 + torch.cuda.synchronize() + C_ref = A @ B.t() + + # Pre-quantize B for all implementations + # (weights can be pre-quantized offline) + B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B) + B_vllm, B_scale_vllm = per_block_cast_to_fp8(B) + + # Block size configuration + block_size = [128, 128] + + # Pre-quantize A for all implementations + A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A) + A_scale_deepgemm = get_col_major_tma_aligned_tensor(A_scale_deepgemm) + C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16) + A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1]) + A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8( + A, block_size[1], column_major_scales=True) + + # === DeepGEMM Implementation === + def deepgemm_gemm(): + deep_gemm.gemm_fp8_fp8_bf16_nt((A_deepgemm, A_scale_deepgemm), + (B_deepgemm, B_scale_deepgemm), + C_deepgemm) + return C_deepgemm + + # === vLLM Triton Implementation === + def vllm_triton_gemm(): + return w8a8_block_fp8_matmul(A_vllm, + B_vllm, + A_scale_vllm, + B_scale_vllm, + block_size, + output_dtype=torch.bfloat16) + + # === vLLM CUTLASS Implementation === + def vllm_cutlass_gemm(): + return ops.cutlass_scaled_mm(A_vllm_cutlass, + B_vllm.T, + scale_a=A_scale_vllm_cutlass, + scale_b=B_scale_vllm.T, + out_dtype=torch.bfloat16) + + # Run correctness check first + if verbose: + print("Running correctness check...") + C_deepgemm = deepgemm_gemm() + C_vllm_triton = vllm_triton_gemm() + C_vllm_cutlass = vllm_cutlass_gemm() + + deepgemm_diff = calc_diff(C_deepgemm, C_ref) + vllm_triton_diff = calc_diff(C_vllm_triton, C_ref) + vllm_cutlass_diff = calc_diff(C_vllm_cutlass, C_ref) + + if verbose: + print(f"DeepGEMM vs Reference difference: {deepgemm_diff:.6f}") + print(f"vLLM Triton vs Reference difference: {vllm_triton_diff:.6f}") + print(f"vLLM CUTLASS vs Reference difference: {vllm_cutlass_diff:.6f}") + print("vLLM Triton vs DeepGEMM difference: " + f"{calc_diff(C_vllm_triton, C_deepgemm):.6f}") + print("vLLM CUTLASS vs DeepGEMM difference: " + f"{calc_diff(C_vllm_cutlass, C_deepgemm):.6f}") + + # Benchmark implementations + implementations = { + "DeepGEMM": deepgemm_gemm, + "vLLM Triton": vllm_triton_gemm, + "vLLM CUTLASS": vllm_cutlass_gemm + } + + benchmark_results = { + "shape": { + "m": m, + "n": n, + "k": k + }, + "implementations": {} + } + + for name, func in implementations.items(): + # Warmup + for _ in range(warmup): + func() + torch.cuda.synchronize() + + # Timing loop + torch.cuda.synchronize() + start = time.time() + for _ in range(repeat): + func() + torch.cuda.synchronize() + end = time.time() + + # Calculate timing and TFLOPS + avg_time_ms = (end - start) / repeat * 1000 + avg_time_us = avg_time_ms * 1000 + tflops = 2 * m * n * k / (avg_time_ms * 1e-3) / 1e12 + gb_s = (m * k + k * n + m * n * 2) / 1e9 / (avg_time_ms * 1e-3) + + benchmark_results["implementations"][name] = { + "time_ms": avg_time_ms, + "time_us": avg_time_us, + "tflops": tflops, + "gb_s": gb_s, + "diff": { + "DeepGEMM": + 0.0 if name == "DeepGEMM" else calc_diff(func(), C_deepgemm), + "Reference": + deepgemm_diff if name == "DeepGEMM" else + (vllm_triton_diff + if name == "vLLM Triton" else vllm_cutlass_diff) + } + } + + if verbose: + print( + f"{name}: {avg_time_ms:.3f} ms, {tflops:.2f} TFLOPS, {gb_s:.2f} GB/s" + ) + + # Calculate speedups + baseline = benchmark_results["implementations"]["DeepGEMM"]["time_ms"] + for name, data in benchmark_results["implementations"].items(): + if name != "DeepGEMM": + speedup = baseline / data["time_ms"] + benchmark_results["implementations"][name][ + "speedup_vs_deepgemm"] = speedup + if verbose: + print(f"DeepGEMM is {1/speedup:.2f}x " + f"{'faster' if 1/speedup > 1 else 'slower'} than {name}") + + vllm_triton_time = benchmark_results["implementations"]["vLLM Triton"][ + "time_ms"] + vllm_cutlass_time = benchmark_results["implementations"]["vLLM CUTLASS"][ + "time_ms"] + cutlass_vs_triton = vllm_triton_time / vllm_cutlass_time + benchmark_results["implementations"]["vLLM CUTLASS"][ + "speedup_vs_triton"] = cutlass_vs_triton + if verbose: + print( + f"vLLM CUTLASS is {cutlass_vs_triton:.2f}x " + f"{'faster' if cutlass_vs_triton > 1 else 'slower'} than vLLM Triton" + ) + + return benchmark_results + + +def format_table_row(values, widths): + """Format a row with specified column widths.""" + return "| " + " | ".join(f"{val:{w}}" + for val, w in zip(values, widths)) + " |" + + +def print_table(headers, rows, title=None): + """Print a table with headers and rows.""" + if title: + print(f"\n{title}") + + # Calculate column widths based on headers and data + widths = [ + max(len(str(h)), max(len(str(row[i])) for row in rows)) + for i, h in enumerate(headers) + ] + + # Create separator line + separator = "+-" + "-+-".join("-" * w for w in widths) + "-+" + + # Print table + print(separator) + print(format_table_row(headers, widths)) + print(separator) + for row in rows: + print(format_table_row(row, widths)) + print(separator) + + +def format_speedup(value): + """Format speedup value with indicator if it's faster or slower.""" + return f"{value:.2f}x {'faster' if value > 1.0 else 'slower'}" + + +def run_benchmarks(verbose: bool = False): + """Run benchmarks for a set of common shapes.""" + print("===== STARTING FP8 GEMM BENCHMARK =====") + + # Make sure we're using the GPU + if not torch.cuda.is_available(): + print("CUDA not available! Tests require GPU.") + return + + # Print system information + print(f"PyTorch version: {torch.__version__}") + print(f"CUDA version: {torch.version.cuda}") + print(f"Triton version: {triton.__version__}") + print(f"Using device: {torch.cuda.get_device_name()}") + + # Enable TF32 for better performance + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + + # Set seeds for reproducibility + torch.manual_seed(42) + torch.cuda.manual_seed(42) + + # Define benchmark shapes (m, n, k) + shapes = [ + (8, 4096, 7168), + (8, 7168, 18432), + (8, 18432, 7168), + (64, 4096, 7168), + (64, 7168, 18432), + (64, 18432, 7168), + (64, 24576, 1536), + (64, 32768, 512), + (64, 7168, 16384), + (128, 4096, 7168), + (128, 7168, 18432), + (128, 18432, 7168), + (1024, 4096, 7168), + (1024, 18432, 7168), + (2048, 4096, 7168), + (4096, 4096, 7168), + ] + shapes = [ + # (64, 2112, 7168), + (64, 24576, 1536), + (64, 32768, 512), + (64, 7168, 16384), + (64, 4096, 7168), + (64, 7168, 2048), + # (128, 2112, 7168), + (128, 24576, 1536), + (128, 32768, 512), + (128, 7168, 16384), + (128, 4096, 7168), + (128, 7168, 2048), + # (4096, 2112, 7168), + (4096, 24576, 1536), + (4096, 32768, 512), + (4096, 7168, 16384), + (4096, 4096, 7168), + (4096, 7168, 2048), + ] + + all_results = [] + for m, n, k in shapes: + result = benchmark_shape(m, n, k, verbose=verbose) + all_results.append(result) + + # Print results in a nicely formatted table + print("\n===== PERFORMANCE COMPARISON =====") + + # Print DeepGEMM table + deepgemm_headers = ["m", "n", "k", "Time (μs)", "TFLOPS", "GB/s"] + deepgemm_rows = [] + for result in all_results: + shape = result["shape"] + impl_data = result["implementations"]["DeepGEMM"] + deepgemm_rows.append([ + shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}", + f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}" + ]) + + print_table(deepgemm_headers, + deepgemm_rows, + title="DeepGEMM Implementation:") + + # Print vLLM Triton table + triton_headers = [ + "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM" + ] + triton_rows = [] + for result in all_results: + shape = result["shape"] + impl_data = result["implementations"]["vLLM Triton"] + speedup = impl_data.get("speedup_vs_deepgemm", 1.0) + triton_rows.append([ + shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}", + f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}", + format_speedup(speedup) + ]) + + print_table(triton_headers, + triton_rows, + title="vLLM Triton Implementation:") + + # Print vLLM CUTLASS table + cutlass_headers = [ + "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM", + "vs Triton" + ] + cutlass_rows = [] + for result in all_results: + shape = result["shape"] + impl_data = result["implementations"]["vLLM CUTLASS"] + vs_deepgemm = impl_data.get("speedup_vs_deepgemm", 1.0) + vs_triton = impl_data.get("speedup_vs_triton", 1.0) + cutlass_rows.append([ + shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}", + f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}", + format_speedup(vs_deepgemm), + format_speedup(vs_triton) + ]) + + print_table(cutlass_headers, + cutlass_rows, + title="vLLM CUTLASS Implementation:") + + # Calculate and print averages + print("\n===== AVERAGE PERFORMANCE =====") + + implementations = ["DeepGEMM", "vLLM Triton", "vLLM CUTLASS"] + avg_metrics = { + impl: { + "tflops": 0, + "gb_s": 0, + "time_ms": 0 + } + for impl in implementations + } + + for result in all_results: + for impl in implementations: + impl_data = result["implementations"][impl] + avg_metrics[impl]["tflops"] += impl_data["tflops"] + avg_metrics[impl]["gb_s"] += impl_data["gb_s"] + avg_metrics[impl]["time_ms"] += impl_data["time_ms"] + + num_shapes = len(all_results) + avg_headers = ["Implementation", "Avg TFLOPS", "Avg GB/s", "Avg Time (ms)"] + avg_rows = [] + + for impl in implementations: + avg_tflops = avg_metrics[impl]["tflops"] / num_shapes + avg_mem_bw = avg_metrics[impl]["gb_s"] / num_shapes + avg_time = avg_metrics[impl]["time_ms"] / num_shapes + avg_rows.append([ + impl, f"{avg_tflops:.2f}", f"{avg_mem_bw:.2f}", f"{avg_time:.2f}" + ]) + + print_table(avg_headers, avg_rows) + + # Calculate average speedups + avg_speedups = { + "DeepGEMM vs vLLM Triton": 0, + "DeepGEMM vs vLLM CUTLASS": 0, + "vLLM CUTLASS vs vLLM Triton": 0 + } + + for result in all_results: + deepgemm_time = result["implementations"]["DeepGEMM"]["time_ms"] + vllm_triton_time = result["implementations"]["vLLM Triton"]["time_ms"] + vllm_cutlass_time = result["implementations"]["vLLM CUTLASS"][ + "time_ms"] + + avg_speedups[ + "DeepGEMM vs vLLM Triton"] += vllm_triton_time / deepgemm_time + avg_speedups[ + "DeepGEMM vs vLLM CUTLASS"] += vllm_cutlass_time / deepgemm_time + avg_speedups[ + "vLLM CUTLASS vs vLLM Triton"] += vllm_triton_time / vllm_cutlass_time + + print("\n===== AVERAGE SPEEDUPS =====") + speedup_headers = ["Comparison", "Speedup"] + speedup_rows = [] + for comparison, total in avg_speedups.items(): + avg_speedup = total / num_shapes + status = "faster" if avg_speedup > 1 else "slower" + speedup_rows.append([comparison, f"{avg_speedup:.2f}x {status}"]) + + print_table(speedup_headers, speedup_rows) + + # Average accuracy comparison + print("\n===== ACCURACY COMPARISON =====") + avg_diff = {impl: 0 for impl in implementations} + + for result in all_results: + for impl in implementations: + avg_diff[impl] += result["implementations"][impl]["diff"][ + "Reference"] + + diff_headers = ["Implementation", "Avg Diff vs Reference"] + diff_rows = [] + for impl in implementations: + diff_rows.append([impl, f"{avg_diff[impl] / num_shapes:.6f}"]) + + print_table(diff_headers, diff_rows) + + +if __name__ == "__main__": + run_benchmarks(verbose=False) diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/graph_machete_bench.py b/online_apiserver_test_maxbs/benchmarks/kernels/graph_machete_bench.py new file mode 100644 index 0000000000000000000000000000000000000000..9a4da0ef5a85d0df2178d5e990ad177e3f76bf37 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/graph_machete_bench.py @@ -0,0 +1,64 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import math +import pickle +from collections import defaultdict + +import matplotlib.pyplot as plt +import pandas as pd +import regex as re +import seaborn as sns +from torch.utils.benchmark import Measurement as TMeasurement + +from vllm.utils import FlexibleArgumentParser + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark the latency of processing a single batch of " + "requests till completion." + ) + parser.add_argument("filename", type=str) + + args = parser.parse_args() + + with open(args.filename, "rb") as f: + data = pickle.load(f) + raw_results: list[TMeasurement] = data["results"] + + results = defaultdict(lambda: list()) + for v in raw_results: + result = re.search(r"MKN=\(\d+x(\d+x\d+)\)", v.task_spec.sub_label) + if result is not None: + KN = result.group(1) + else: + raise Exception("MKN not found") + result = re.search(r"MKN=\((\d+)x\d+x\d+\)", v.task_spec.sub_label) + if result is not None: + M = result.group(1) + else: + raise Exception("MKN not found") + + kernel = v.task_spec.description + results[KN].append({"kernel": kernel, "batch_size": M, "median": v.median}) + + rows = int(math.ceil(len(results) / 2)) + fig, axs = plt.subplots(rows, 2, figsize=(12, 5 * rows)) + axs = axs.flatten() + for axs_idx, (shape, data) in enumerate(results.items()): + plt.sca(axs[axs_idx]) + df = pd.DataFrame(data) + sns.lineplot( + data=df, + x="batch_size", + y="median", + hue="kernel", + style="kernel", + markers=True, + dashes=False, + palette="Dark2", + ) + plt.title(f"Shape: {shape}") + plt.ylabel("time (median, s)") + plt.tight_layout() + plt.savefig("graph_machete_bench.pdf") diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/requirements.txt b/online_apiserver_test_maxbs/benchmarks/kernels/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..1411a4a0b5ab886adfb744e685d150151ab10023 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/requirements.txt @@ -0,0 +1 @@ +pandas \ No newline at end of file diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/utils.py b/online_apiserver_test_maxbs/benchmarks/kernels/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..4bbb36bb4359259944310adb8e1047a02f126dea --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/utils.py @@ -0,0 +1,214 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import dataclasses +from collections.abc import Iterable +from typing import Any, Callable, Optional + +import torch +import torch.utils.benchmark as TBenchmark +from torch.utils.benchmark import Measurement as TMeasurement + + +@dataclasses.dataclass +class CudaGraphBenchParams: + num_ops_in_cuda_graph: int + + +@dataclasses.dataclass +class ArgPool: + """ + When some argument of the benchmarking function is annotated with this type, + the benchmarking class (BenchMM) will collapse the argument to a pick a + single value from the given list of values, during function invocation. + For every invocation during a benchmarking run, it will choose a + different value from the list. + """ + + values: Iterable[Any] + + def __getitem__(self, index): + return self.values[index] + + +class Bench: + class ArgsIterator: + def __init__(self, args_list, kwargs_list): + assert len(args_list) == len(kwargs_list) + self.args_list = args_list + self.kwargs_list = kwargs_list + self.n = len(self.args_list) + self.idx = 0 + + def __next__(self): + while True: + yield (self.args_list[self.idx], self.kwargs_list[self.idx]) + self.idx += 1 + self.idx = self.idx % self.n + + def reset(self): + self.idx = 0 + + @property + def n_args(self): + return self.n + + def __init__( + self, + cuda_graph_params: Optional[CudaGraphBenchParams], + label: str, + sub_label: str, + description: str, + fn: Callable, + *args, + **kwargs, + ): + self.cuda_graph_params = cuda_graph_params + self.use_cuda_graph = self.cuda_graph_params is not None + self.label = label + self.sub_label = sub_label + self.description = description + self.fn = fn + + # Process args + self._args = args + self._kwargs = kwargs + self.args_list, self.kwargs_list = self.collapse_argpool(*args, **kwargs) + self.args_iterator = self.ArgsIterator(self.args_list, self.kwargs_list) + + # Cudagraph runner + self.g = None + if self.use_cuda_graph: + self.g = self.get_cuda_graph_runner() + + # benchmark run params + self.min_run_time = 1 + + def collapse_argpool(self, *args, **kwargs): + argpool_args = [arg for arg in args if isinstance(arg, ArgPool)] + [ + arg for arg in kwargs.values() if isinstance(arg, ArgPool) + ] + if len(argpool_args) == 0: + return [args], [kwargs] + + # Make sure all argpools are of the same size + argpool_size = len(argpool_args[0].values) + assert all([argpool_size == len(arg.values) for arg in argpool_args]) + + # create copies of the args + args_list = [] + kwargs_list = [] + for _ in range(argpool_size): + args_list.append(args) + kwargs_list.append(kwargs.copy()) + + for i in range(argpool_size): + # collapse args; Just pick the ith value + args_list[i] = tuple( + [arg[i] if isinstance(arg, ArgPool) else arg for arg in args_list[i]] + ) + + # collapse kwargs + kwargs_i = kwargs_list[i] + arg_pool_keys = [k for k, v in kwargs_i.items() if isinstance(v, ArgPool)] + for k in arg_pool_keys: + # again just pick the ith value + kwargs_i[k] = kwargs_i[k][i] + kwargs_list[i] = kwargs_i + + return args_list, kwargs_list + + def get_cuda_graph_runner(self): + assert self.use_cuda_graph + assert self.args_iterator is not None + + num_graph_ops = self.cuda_graph_params.num_ops_in_cuda_graph + + # warmup + args_it = self.args_iterator.__next__() + for _ in range(2): + args, kwargs = next(args_it) + self.fn(*args, **kwargs) + + self.args_iterator.reset() + args_it = self.args_iterator.__next__() + stream = torch.cuda.Stream() + with torch.cuda.stream(stream): + g = torch.cuda.CUDAGraph() + with torch.cuda.graph(g): + for _ in range(num_graph_ops): + args, kwargs = next(args_it) + self.fn(*args, **kwargs) + return g + + def run_cudagrah(self) -> TMeasurement: + assert self.use_cuda_graph + globals = {"g": self.g} + + return TBenchmark.Timer( + stmt="g.replay()", + globals=globals, + label=( + f"{self.label}" + f" | cugraph {self.cuda_graph_params.num_ops_in_cuda_graph} ops" + ), + sub_label=self.sub_label, + description=self.description, + ).blocked_autorange(min_run_time=self.min_run_time) + + def run_eager(self) -> TMeasurement: + setup = None + stmt = None + globals = None + + has_arg_pool = self.args_iterator.n_args > 1 + if has_arg_pool: + setup = """ + args_iterator.reset() + args_it = args_iterator.__next__() + """ + stmt = """ + args, kwargs = next(args_it) + fn(*args, **kwargs) + """ + globals = {"fn": self.fn, "args_iterator": self.args_iterator} + else: + # no arg pool. Just use the args and kwargs directly + self.args_iterator.reset() + args_it = self.args_iterator.__next__() + args, kwargs = next(args_it) + + setup = "" + stmt = """ + fn(*args, **kwargs) + """ + globals = {"fn": self.fn, "args": args, "kwargs": kwargs} + + return TBenchmark.Timer( + stmt=stmt, + setup=setup, + globals=globals, + label=self.label, + sub_label=self.sub_label, + description=self.description, + ).blocked_autorange(min_run_time=self.min_run_time) + + def run(self) -> TMeasurement: + timer = None + if self.use_cuda_graph: # noqa SIM108 + timer = self.run_cudagrah() + else: + timer = self.run_eager() + if not timer.meets_confidence() or timer.has_warnings: + print("Doesn't meet confidence - re-running bench ...") + return self.run() + return timer + + def __enter__(self): + return self + + def __exit__(self, exc_type, exc_value, traceback): + if exc_type: + print(f"exc type {exc_type}") + print(f"exc value {exc_value}") + print(f"exc traceback {traceback}") diff --git a/online_apiserver_test_maxbs/benchmarks/kernels/weight_shapes.py b/online_apiserver_test_maxbs/benchmarks/kernels/weight_shapes.py new file mode 100644 index 0000000000000000000000000000000000000000..a27f02394afbdfa4f3a352a599db349a2601bea5 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/kernels/weight_shapes.py @@ -0,0 +1,98 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +# Weight Shapes are in the format +# ([K, N], TP_SPLIT_DIM) +# Example: +# A shape of ([14336, 4096], 0) indicates the following GEMM shape, +# - TP1 : K = 14336, N = 4096 +# - TP2 : K = 7168, N = 4096 +# A shape of ([4096, 6144], 1) indicates the following GEMM shape, +# - TP1 : K = 4096, N = 6144 +# - TP4 : K = 4096, N = 1536 + +# TP1 shapes +WEIGHT_SHAPES = { + "mistralai/Mistral-7B-v0.1": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-7b-hf": [ + ([4096, 12288], 1), + ([4096, 4096], 0), + ([4096, 22016], 1), + ([11008, 4096], 0), + ], + "meta-llama/Llama-3-8b": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-2-13b-hf": [ + ([5120, 15360], 1), + ([5120, 5120], 0), + ([5120, 27648], 1), + ([13824, 5120], 0), + ], + "meta-llama/Llama-2-70b-hf": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 57344], 1), + ([28672, 8192], 0), + ], + "meta-llama/Llama-3.1-405b-hf": [ + ([16384, 18432], 1), + ([16384, 16384], 0), + ([16384, 106496], 1), + ([53248, 16384], 0), + ], + "meta-llama/Llama-3.1-8B-Instruct": [ + ([4096, 6144], 1), + ([4096, 4096], 0), + ([4096, 28672], 1), + ([14336, 4096], 0), + ], + "meta-llama/Llama-3.3-70B-Instruct": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 57344], 1), + ([28672, 8192], 0), + ], + "mistralai/Mistral-Large-Instruct-2407": [ + ([12288, 14336], 1), + ([12288, 12288], 0), + ([12288, 57344], 1), + ([28672, 12288], 0), + ], + "Qwen/Qwen2.5-7B-Instruct": [ + ([3584, 4608], 1), + ([3584, 3584], 0), + ([3584, 37888], 1), + ([18944, 3584], 0), + ], + "Qwen/Qwen2.5-32B-Instruct": [ + ([5120, 7168], 1), + ([5120, 5120], 0), + ([5120, 55296], 1), + ([27648, 5120], 0), + ], + "Qwen/Qwen2.5-72B-Instruct": [ + ([8192, 10240], 1), + ([8192, 8192], 0), + ([8192, 59136], 1), + ([29568, 8192], 0), + ], + "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": [ + ([2048, 3072], 1), + ([2048, 4096], 1), + ([2048, 2048], 0), + ([2048, 576], 0), + ([2048, 21888], 1), + ([10944, 2048], 0), + ([2048, 2816], 1), + ([1408, 2048], 0), + ], +} diff --git a/online_apiserver_test_maxbs/benchmarks/overheads/benchmark_hashing.py b/online_apiserver_test_maxbs/benchmarks/overheads/benchmark_hashing.py new file mode 100644 index 0000000000000000000000000000000000000000..0957a9c65f06c912ae73588e017602af72ae9772 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/overheads/benchmark_hashing.py @@ -0,0 +1,64 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +import cProfile +import pstats + +from vllm import LLM, SamplingParams +from vllm.utils import FlexibleArgumentParser + +# A very long prompt, total number of tokens is about 15k. +LONG_PROMPT = ["You are an expert in large language models, aren't you?"] * 1000 +LONG_PROMPT = " ".join(LONG_PROMPT) + + +def main(args): + llm = LLM( + model=args.model, + enforce_eager=True, + enable_prefix_caching=True, + tensor_parallel_size=args.tensor_parallel_size, + ) + + sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len) + profiler = cProfile.Profile() + + print("------warm up------") + for i in range(3): + output = llm.generate(LONG_PROMPT, sampling_params) + print(output[0].outputs[0].text) + + print("------start generating------") + for i in range(3): + profiler.runctx( + "llm.generate(LONG_PROMPT, sampling_params)", globals(), locals() + ) + + # analyze the runtime of hashing function + stats = pstats.Stats(profiler) + stats.sort_stats("cumulative") + total_time = 0 + total_calls = 0 + for func in stats.stats: + if "hash_of_block" in func[2]: + total_time = stats.stats[func][3] + total_calls = stats.stats[func][0] + percentage = (total_time / stats.total_tt) * 100 + print( + f"Hashing took {total_time:.2f} seconds,{percentage:.2f}% of the total runtime." + ) + + +if __name__ == "__main__": + parser = FlexibleArgumentParser( + description="Benchmark the performance of hashing function in" + "automatic prefix caching." + ) + parser.add_argument("--model", type=str, default="lmsys/longchat-7b-16k") + parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1) + parser.add_argument("--output-len", type=int, default=10) + parser.add_argument( + "--enable-prefix-caching", action="store_true", help="enable prefix caching" + ) + args = parser.parse_args() + main(args) diff --git a/online_apiserver_test_maxbs/benchmarks/pyproject.toml b/online_apiserver_test_maxbs/benchmarks/pyproject.toml new file mode 100644 index 0000000000000000000000000000000000000000..65b1e09a247e2d74e31b616c38a8544de75ad525 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/pyproject.toml @@ -0,0 +1,49 @@ +# This local pyproject file is part of the migration from yapf to ruff format. +# It uses the same core rules as the main pyproject.toml file, but with the +# following differences: +# - ruff line length is overridden to 88 +# - deprecated typing ignores (UP006, UP035) have been removed + +[tool.ruff] +line-length = 88 + +[tool.ruff.lint.per-file-ignores] +"vllm/third_party/**" = ["ALL"] +"vllm/version.py" = ["F401"] +"vllm/_version.py" = ["ALL"] + +[tool.ruff.lint] +select = [ + # pycodestyle + "E", + # Pyflakes + "F", + # pyupgrade + "UP", + # flake8-bugbear + "B", + # flake8-simplify + "SIM", + # isort + "I", + # flake8-logging-format + "G", +] +ignore = [ + # star imports + "F405", "F403", + # lambda expression assignment + "E731", + # Loop control variable not used within loop body + "B007", + # f-string format + "UP032", + # Can remove once 3.10+ is the minimum Python version + "UP007", +] + +[tool.ruff.lint.isort] +known-first-party = ["vllm"] + +[tool.ruff.format] +docstring-code-format = true \ No newline at end of file diff --git a/online_apiserver_test_maxbs/benchmarks/run_structured_output_benchmark.sh b/online_apiserver_test_maxbs/benchmarks/run_structured_output_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..b043ab83e4608e7b27e2f94e6ec24a05f7c474aa --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/run_structured_output_benchmark.sh @@ -0,0 +1,129 @@ +#!/bin/bash + +# default values +MODEL=${MODEL:-"Qwen/Qwen2.5-7B-Instruct"} +BACKEND=${BACKEND:-"vllm"} +DATASET=${DATASET:-"xgrammar_bench"} +SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" +OUTPUT_DIR=${OUTPUT_DIR:-"$SCRIPT_DIR/structured_output_benchmark_results"} +PORT=${PORT:-8000} +STRUCTURED_OUTPUT_RATIO=${STRUCTURED_OUTPUT_RATIO:-1} +TOTAL_SECONDS=${TOTAL_SECONDS:-90} +MAX_NEW_TOKENS=${MAX_NEW_TOKENS:-300} +TOKENIZER_MODE=${TOKENIZER_MODE:-"auto"} + +usage() { + echo "Usage: $0 [options]" + echo "Options:" + echo " --model MODEL Model to benchmark (default: $MODEL)" + echo " --backend BACKEND Backend to use (default: $BACKEND)" + echo " --dataset DATASET Dataset to use (default: $DATASET)" + echo " --max-new-tokens N Maximum number of tokens to generate (default: $MAX_NEW_TOKENS)" + echo " --output-dir DIR Output directory for results (default: $OUTPUT_DIR)" + echo " --port PORT Port to use (default: $PORT)" + echo " --structured-output-ratio N Ratio of structured outputs (default: $STRUCTURED_OUTPUT_RATIO)" + echo " --tokenizer-mode MODE Tokenizer mode to use (default: $TOKENIZER_MODE)" + echo " --total-seconds N Total seconds to run the benchmark (default: $TOTAL_SECONDS)" + echo " -h, --help Show this help message and exit" + exit 0 +} + +# parse command line arguments +while [[ $# -gt 0 ]]; do + case $1 in + --model) + MODEL="$2" + shift 2 + ;; + --backend) + BACKEND="$2" + shift 2 + ;; + --dataset) + DATASET="$2" + shift 2 + ;; + --max-new-tokens) + MAX_NEW_TOKENS="$2" + shift 2 + ;; + --output-dir) + OUTPUT_DIR="$2" + shift 2 + ;; + --port) + PORT="$2" + shift 2 + ;; + --structured-output-ratio) + STRUCTURED_OUTPUT_RATIO="$2" + shift 2 + ;; + --tokenizer-mode) + TOKENIZER_MODE="$2" + shift 2 + ;; + --total-seconds) + TOTAL_SECONDS="$2" + shift 2 + ;; + -h|--help) + usage + ;; + *) + echo "Unknown argument: $1\n" + usage + ;; + esac +done + +# Create output directory if it doesn't exist +mkdir -p "$OUTPUT_DIR" + +# Define QPS values to test +QPS_VALUES=(25 20 15 10 5 1) + +# Common parameters +COMMON_PARAMS="--backend $BACKEND \ + --model $MODEL \ + --dataset $DATASET \ + --structured-output-ratio $STRUCTURED_OUTPUT_RATIO \ + --save-results \ + --result-dir $OUTPUT_DIR \ + --output-len $MAX_NEW_TOKENS \ + --port $PORT \ + --tokenizer-mode $TOKENIZER_MODE" + +echo "Starting structured output benchmark with model: $MODEL" +echo "Backend: $BACKEND" +echo "Dataset: $DATASET" +echo "Results will be saved to: $OUTPUT_DIR" +echo "----------------------------------------" + +# Run benchmarks with different QPS values +for qps in "${QPS_VALUES[@]}"; do + echo "Running benchmark with QPS: $qps" + + # Get git hash and branch for the filename + GIT_HASH=$(git rev-parse --short HEAD 2>/dev/null || echo "unknown") + GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "unknown") + + # Construct filename for this run + FILENAME="${BACKEND}_${qps}qps_$(basename $MODEL)_${DATASET}_${GIT_HASH}.json" + + NUM_PROMPTS=$(echo "$TOTAL_SECONDS * $qps" | bc) + NUM_PROMPTS=${NUM_PROMPTS%.*} # Remove fractional part + echo "Running benchmark with $NUM_PROMPTS prompts" + + # Run the benchmark + python "$SCRIPT_DIR/benchmark_serving_structured_output.py" $COMMON_PARAMS \ + --request-rate $qps \ + --result-filename "$FILENAME" \ + --num-prompts $NUM_PROMPTS + + echo "Completed benchmark with QPS: $qps" + echo "----------------------------------------" +done + +echo "All benchmarks completed!" +echo "Results saved to: $OUTPUT_DIR" diff --git a/online_apiserver_test_maxbs/benchmarks/sonnet.txt b/online_apiserver_test_maxbs/benchmarks/sonnet.txt new file mode 100644 index 0000000000000000000000000000000000000000..34c444e8ce8e2dc701ec80931401c57014ae0bd1 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/sonnet.txt @@ -0,0 +1,518 @@ +FROM fairest creatures we desire increase, +That thereby beauty's rose might never die, +But as the riper should by time decease, +His tender heir might bear his memory: +But thou, contracted to thine own bright eyes, +Feed'st thy light'st flame with self-substantial fuel, +Making a famine where abundance lies, +Thyself thy foe, to thy sweet self too cruel. +Thou that art now the world's fresh ornament +And only herald to the gaudy spring, +Within thine own bud buriest thy content +And, tender churl, makest waste in niggarding. +Pity the world, or else this glutton be, +To eat the world's due, by the grave and thee. +When forty winters shall beseige thy brow, +And dig deep trenches in thy beauty's field, +Thy youth's proud livery, so gazed on now, +Will be a tatter'd weed, of small worth held: +Then being ask'd where all thy beauty lies, +Where all the treasure of thy lusty days, +To say, within thine own deep-sunken eyes, +Were an all-eating shame and thriftless praise. +How much more praise deserved thy beauty's use, +If thou couldst answer 'This fair child of mine +Shall sum my count and make my old excuse,' +Proving his beauty by succession thine! +This were to be new made when thou art old, +And see thy blood warm when thou feel'st it cold. +Look in thy glass, and tell the face thou viewest +Now is the time that face should form another; +Whose fresh repair if now thou not renewest, +Thou dost beguile the world, unbless some mother. +For where is she so fair whose unear'd womb +Disdains the tillage of thy husbandry? +Or who is he so fond will be the tomb +Of his self-love, to stop posterity? +Thou art thy mother's glass, and she in thee +Calls back the lovely April of her prime: +So thou through windows of thine age shall see +Despite of wrinkles this thy golden time. +But if thou live, remember'd not to be, +Die single, and thine image dies with thee. +Unthrifty loveliness, why dost thou spend +Upon thyself thy beauty's legacy? +Nature's bequest gives nothing but doth lend, +And being frank she lends to those are free. +Then, beauteous niggard, why dost thou abuse +The bounteous largess given thee to give? +Profitless usurer, why dost thou use +So great a sum of sums, yet canst not live? +For having traffic with thyself alone, +Thou of thyself thy sweet self dost deceive. +Then how, when nature calls thee to be gone, +What acceptable audit canst thou leave? +Thy unused beauty must be tomb'd with thee, +Which, used, lives th' executor to be. +Those hours, that with gentle work did frame +The lovely gaze where every eye doth dwell, +Will play the tyrants to the very same +And that unfair which fairly doth excel: +For never-resting time leads summer on +To hideous winter and confounds him there; +Sap cheque'd with frost and lusty leaves quite gone, +Beauty o'ersnow'd and bareness every where: +Then, were not summer's distillation left, +A liquid prisoner pent in walls of glass, +Beauty's effect with beauty were bereft, +Nor it nor no remembrance what it was: +But flowers distill'd though they with winter meet, +Leese but their show; their substance still lives sweet. +Then let not winter's ragged hand deface +In thee thy summer, ere thou be distill'd: +Make sweet some vial; treasure thou some place +With beauty's treasure, ere it be self-kill'd. +That use is not forbidden usury, +Which happies those that pay the willing loan; +That's for thyself to breed another thee, +Or ten times happier, be it ten for one; +Ten times thyself were happier than thou art, +If ten of thine ten times refigured thee: +Then what could death do, if thou shouldst depart, +Leaving thee living in posterity? +Be not self-will'd, for thou art much too fair +To be death's conquest and make worms thine heir. +Lo! in the orient when the gracious light +Lifts up his burning head, each under eye +Doth homage to his new-appearing sight, +Serving with looks his sacred majesty; +And having climb'd the steep-up heavenly hill, +Resembling strong youth in his middle age, +yet mortal looks adore his beauty still, +Attending on his golden pilgrimage; +But when from highmost pitch, with weary car, +Like feeble age, he reeleth from the day, +The eyes, 'fore duteous, now converted are +From his low tract and look another way: +So thou, thyself out-going in thy noon, +Unlook'd on diest, unless thou get a son. +Music to hear, why hear'st thou music sadly? +Sweets with sweets war not, joy delights in joy. +Why lovest thou that which thou receivest not gladly, +Or else receivest with pleasure thine annoy? +If the true concord of well-tuned sounds, +By unions married, do offend thine ear, +They do but sweetly chide thee, who confounds +In singleness the parts that thou shouldst bear. +Mark how one string, sweet husband to another, +Strikes each in each by mutual ordering, +Resembling sire and child and happy mother +Who all in one, one pleasing note do sing: +Whose speechless song, being many, seeming one, +Sings this to thee: 'thou single wilt prove none.' +Is it for fear to wet a widow's eye +That thou consumest thyself in single life? +Ah! if thou issueless shalt hap to die. +The world will wail thee, like a makeless wife; +The world will be thy widow and still weep +That thou no form of thee hast left behind, +When every private widow well may keep +By children's eyes her husband's shape in mind. +Look, what an unthrift in the world doth spend +Shifts but his place, for still the world enjoys it; +But beauty's waste hath in the world an end, +And kept unused, the user so destroys it. +No love toward others in that bosom sits +That on himself such murderous shame commits. +For shame! deny that thou bear'st love to any, +Who for thyself art so unprovident. +Grant, if thou wilt, thou art beloved of many, +But that thou none lovest is most evident; +For thou art so possess'd with murderous hate +That 'gainst thyself thou stick'st not to conspire. +Seeking that beauteous roof to ruinate +Which to repair should be thy chief desire. +O, change thy thought, that I may change my mind! +Shall hate be fairer lodged than gentle love? +Be, as thy presence is, gracious and kind, +Or to thyself at least kind-hearted prove: +Make thee another self, for love of me, +That beauty still may live in thine or thee. +As fast as thou shalt wane, so fast thou growest +In one of thine, from that which thou departest; +And that fresh blood which youngly thou bestowest +Thou mayst call thine when thou from youth convertest. +Herein lives wisdom, beauty and increase: +Without this, folly, age and cold decay: +If all were minded so, the times should cease +And threescore year would make the world away. +Let those whom Nature hath not made for store, +Harsh featureless and rude, barrenly perish: +Look, whom she best endow'd she gave the more; +Which bounteous gift thou shouldst in bounty cherish: +She carved thee for her seal, and meant thereby +Thou shouldst print more, not let that copy die. +When I do count the clock that tells the time, +And see the brave day sunk in hideous night; +When I behold the violet past prime, +And sable curls all silver'd o'er with white; +When lofty trees I see barren of leaves +Which erst from heat did canopy the herd, +And summer's green all girded up in sheaves +Borne on the bier with white and bristly beard, +Then of thy beauty do I question make, +That thou among the wastes of time must go, +Since sweets and beauties do themselves forsake +And die as fast as they see others grow; +And nothing 'gainst Time's scythe can make defence +Save breed, to brave him when he takes thee hence. +O, that you were yourself! but, love, you are +No longer yours than you yourself here live: +Against this coming end you should prepare, +And your sweet semblance to some other give. +So should that beauty which you hold in lease +Find no determination: then you were +Yourself again after yourself's decease, +When your sweet issue your sweet form should bear. +Who lets so fair a house fall to decay, +Which husbandry in honour might uphold +Against the stormy gusts of winter's day +And barren rage of death's eternal cold? +O, none but unthrifts! Dear my love, you know +You had a father: let your son say so. +Not from the stars do I my judgment pluck; +And yet methinks I have astronomy, +But not to tell of good or evil luck, +Of plagues, of dearths, or seasons' quality; +Nor can I fortune to brief minutes tell, +Pointing to each his thunder, rain and wind, +Or say with princes if it shall go well, +By oft predict that I in heaven find: +But from thine eyes my knowledge I derive, +And, constant stars, in them I read such art +As truth and beauty shall together thrive, +If from thyself to store thou wouldst convert; +Or else of thee this I prognosticate: +Thy end is truth's and beauty's doom and date. +When I consider every thing that grows +Holds in perfection but a little moment, +That this huge stage presenteth nought but shows +Whereon the stars in secret influence comment; +When I perceive that men as plants increase, +Cheered and cheque'd even by the self-same sky, +Vaunt in their youthful sap, at height decrease, +And wear their brave state out of memory; +Then the conceit of this inconstant stay +Sets you most rich in youth before my sight, +Where wasteful Time debateth with Decay, +To change your day of youth to sullied night; +And all in war with Time for love of you, +As he takes from you, I engraft you new. +But wherefore do not you a mightier way +Make war upon this bloody tyrant, Time? +And fortify yourself in your decay +With means more blessed than my barren rhyme? +Now stand you on the top of happy hours, +And many maiden gardens yet unset +With virtuous wish would bear your living flowers, +Much liker than your painted counterfeit: +So should the lines of life that life repair, +Which this, Time's pencil, or my pupil pen, +Neither in inward worth nor outward fair, +Can make you live yourself in eyes of men. +To give away yourself keeps yourself still, +And you must live, drawn by your own sweet skill. +Who will believe my verse in time to come, +If it were fill'd with your most high deserts? +Though yet, heaven knows, it is but as a tomb +Which hides your life and shows not half your parts. +If I could write the beauty of your eyes +And in fresh numbers number all your graces, +The age to come would say 'This poet lies: +Such heavenly touches ne'er touch'd earthly faces.' +So should my papers yellow'd with their age +Be scorn'd like old men of less truth than tongue, +And your true rights be term'd a poet's rage +And stretched metre of an antique song: +But were some child of yours alive that time, +You should live twice; in it and in my rhyme. +Shall I compare thee to a summer's day? +Thou art more lovely and more temperate: +Rough winds do shake the darling buds of May, +And summer's lease hath all too short a date: +Sometime too hot the eye of heaven shines, +And often is his gold complexion dimm'd; +And every fair from fair sometime declines, +By chance or nature's changing course untrimm'd; +But thy eternal summer shall not fade +Nor lose possession of that fair thou owest; +Nor shall Death brag thou wander'st in his shade, +When in eternal lines to time thou growest: +So long as men can breathe or eyes can see, +So long lives this and this gives life to thee. +Devouring Time, blunt thou the lion's paws, +And make the earth devour her own sweet brood; +Pluck the keen teeth from the fierce tiger's jaws, +And burn the long-lived phoenix in her blood; +Make glad and sorry seasons as thou fleets, +And do whate'er thou wilt, swift-footed Time, +To the wide world and all her fading sweets; +But I forbid thee one most heinous crime: +O, carve not with thy hours my love's fair brow, +Nor draw no lines there with thine antique pen; +Him in thy course untainted do allow +For beauty's pattern to succeeding men. +Yet, do thy worst, old Time: despite thy wrong, +My love shall in my verse ever live young. +A woman's face with Nature's own hand painted +Hast thou, the master-mistress of my passion; +A woman's gentle heart, but not acquainted +With shifting change, as is false women's fashion; +An eye more bright than theirs, less false in rolling, +Gilding the object whereupon it gazeth; +A man in hue, all 'hues' in his controlling, +Much steals men's eyes and women's souls amazeth. +And for a woman wert thou first created; +Till Nature, as she wrought thee, fell a-doting, +And by addition me of thee defeated, +By adding one thing to my purpose nothing. +But since she prick'd thee out for women's pleasure, +Mine be thy love and thy love's use their treasure. +So is it not with me as with that Muse +Stirr'd by a painted beauty to his verse, +Who heaven itself for ornament doth use +And every fair with his fair doth rehearse +Making a couplement of proud compare, +With sun and moon, with earth and sea's rich gems, +With April's first-born flowers, and all things rare +That heaven's air in this huge rondure hems. +O' let me, true in love, but truly write, +And then believe me, my love is as fair +As any mother's child, though not so bright +As those gold candles fix'd in heaven's air: +Let them say more than like of hearsay well; +I will not praise that purpose not to sell. +My glass shall not persuade me I am old, +So long as youth and thou are of one date; +But when in thee time's furrows I behold, +Then look I death my days should expiate. +For all that beauty that doth cover thee +Is but the seemly raiment of my heart, +Which in thy breast doth live, as thine in me: +How can I then be elder than thou art? +O, therefore, love, be of thyself so wary +As I, not for myself, but for thee will; +Bearing thy heart, which I will keep so chary +As tender nurse her babe from faring ill. +Presume not on thy heart when mine is slain; +Thou gavest me thine, not to give back again. +As an unperfect actor on the stage +Who with his fear is put besides his part, +Or some fierce thing replete with too much rage, +Whose strength's abundance weakens his own heart. +So I, for fear of trust, forget to say +The perfect ceremony of love's rite, +And in mine own love's strength seem to decay, +O'ercharged with burden of mine own love's might. +O, let my books be then the eloquence +And dumb presagers of my speaking breast, +Who plead for love and look for recompense +More than that tongue that more hath more express'd. +O, learn to read what silent love hath writ: +To hear with eyes belongs to love's fine wit. +Mine eye hath play'd the painter and hath stell'd +Thy beauty's form in table of my heart; +My body is the frame wherein 'tis held, +And perspective it is the painter's art. +For through the painter must you see his skill, +To find where your true image pictured lies; +Which in my bosom's shop is hanging still, +That hath his windows glazed with thine eyes. +Now see what good turns eyes for eyes have done: +Mine eyes have drawn thy shape, and thine for me +Are windows to my breast, where-through the sun +Delights to peep, to gaze therein on thee; +Yet eyes this cunning want to grace their art; +They draw but what they see, know not the heart. +Let those who are in favour with their stars +Of public honour and proud titles boast, +Whilst I, whom fortune of such triumph bars, +Unlook'd for joy in that I honour most. +Great princes' favourites their fair leaves spread +But as the marigold at the sun's eye, +And in themselves their pride lies buried, +For at a frown they in their glory die. +The painful warrior famoused for fight, +After a thousand victories once foil'd, +Is from the book of honour razed quite, +And all the rest forgot for which he toil'd: +Then happy I, that love and am beloved +Where I may not remove nor be removed. +Lord of my love, to whom in vassalage +Thy merit hath my duty strongly knit, +To thee I send this written embassage, +To witness duty, not to show my wit: +Duty so great, which wit so poor as mine +May make seem bare, in wanting words to show it, +But that I hope some good conceit of thine +In thy soul's thought, all naked, will bestow it; +Till whatsoever star that guides my moving +Points on me graciously with fair aspect +And puts apparel on my tatter'd loving, +To show me worthy of thy sweet respect: +Then may I dare to boast how I do love thee; +Till then not show my head where thou mayst prove me. +Weary with toil, I haste me to my bed, +The dear repose for limbs with travel tired; +But then begins a journey in my head, +To work my mind, when body's work's expired: +For then my thoughts, from far where I abide, +Intend a zealous pilgrimage to thee, +And keep my drooping eyelids open wide, +Looking on darkness which the blind do see +Save that my soul's imaginary sight +Presents thy shadow to my sightless view, +Which, like a jewel hung in ghastly night, +Makes black night beauteous and her old face new. +Lo! thus, by day my limbs, by night my mind, +For thee and for myself no quiet find. +How can I then return in happy plight, +That am debarr'd the benefit of rest? +When day's oppression is not eased by night, +But day by night, and night by day, oppress'd? +And each, though enemies to either's reign, +Do in consent shake hands to torture me; +The one by toil, the other to complain +How far I toil, still farther off from thee. +I tell the day, to please them thou art bright +And dost him grace when clouds do blot the heaven: +So flatter I the swart-complexion'd night, +When sparkling stars twire not thou gild'st the even. +But day doth daily draw my sorrows longer +And night doth nightly make grief's strength seem stronger. +When, in disgrace with fortune and men's eyes, +I all alone beweep my outcast state +And trouble deal heaven with my bootless cries +And look upon myself and curse my fate, +Wishing me like to one more rich in hope, +Featured like him, like him with friends possess'd, +Desiring this man's art and that man's scope, +With what I most enjoy contented least; +Yet in these thoughts myself almost despising, +Haply I think on thee, and then my state, +Like to the lark at break of day arising +From sullen earth, sings hymns at heaven's gate; +For thy sweet love remember'd such wealth brings +That then I scorn to change my state with kings. +When to the sessions of sweet silent thought +I summon up remembrance of things past, +I sigh the lack of many a thing I sought, +And with old woes new wail my dear time's waste: +Then can I drown an eye, unused to flow, +For precious friends hid in death's dateless night, +And weep afresh love's long since cancell'd woe, +And moan the expense of many a vanish'd sight: +Then can I grieve at grievances foregone, +And heavily from woe to woe tell o'er +The sad account of fore-bemoaned moan, +Which I new pay as if not paid before. +But if the while I think on thee, dear friend, +All losses are restored and sorrows end. +Thy bosom is endeared with all hearts, +Which I by lacking have supposed dead, +And there reigns love and all love's loving parts, +And all those friends which I thought buried. +How many a holy and obsequious tear +Hath dear religious love stol'n from mine eye +As interest of the dead, which now appear +But things removed that hidden in thee lie! +Thou art the grave where buried love doth live, +Hung with the trophies of my lovers gone, +Who all their parts of me to thee did give; +That due of many now is thine alone: +Their images I loved I view in thee, +And thou, all they, hast all the all of me. +If thou survive my well-contented day, +When that churl Death my bones with dust shall cover, +And shalt by fortune once more re-survey +These poor rude lines of thy deceased lover, +Compare them with the bettering of the time, +And though they be outstripp'd by every pen, +Reserve them for my love, not for their rhyme, +Exceeded by the height of happier men. +O, then vouchsafe me but this loving thought: +'Had my friend's Muse grown with this growing age, +A dearer birth than this his love had brought, +To march in ranks of better equipage: +But since he died and poets better prove, +Theirs for their style I'll read, his for his love.' +Full many a glorious morning have I seen +Flatter the mountain-tops with sovereign eye, +Kissing with golden face the meadows green, +Gilding pale streams with heavenly alchemy; +Anon permit the basest clouds to ride +With ugly rack on his celestial face, +And from the forlorn world his visage hide, +Stealing unseen to west with this disgrace: +Even so my sun one early morn did shine +With all triumphant splendor on my brow; +But out, alack! he was but one hour mine; +The region cloud hath mask'd him from me now. +Yet him for this my love no whit disdaineth; +Suns of the world may stain when heaven's sun staineth. +Why didst thou promise such a beauteous day, +And make me travel forth without my cloak, +To let base clouds o'ertake me in my way, +Hiding thy bravery in their rotten smoke? +'Tis not enough that through the cloud thou break, +To dry the rain on my storm-beaten face, +For no man well of such a salve can speak +That heals the wound and cures not the disgrace: +Nor can thy shame give physic to my grief; +Though thou repent, yet I have still the loss: +The offender's sorrow lends but weak relief +To him that bears the strong offence's cross. +Ah! but those tears are pearl which thy love sheds, +And they are rich and ransom all ill deeds. +No more be grieved at that which thou hast done: +Roses have thorns, and silver fountains mud; +Clouds and eclipses stain both moon and sun, +And loathsome canker lives in sweetest bud. +All men make faults, and even I in this, +Authorizing thy trespass with compare, +Myself corrupting, salving thy amiss, +Excusing thy sins more than thy sins are; +For to thy sensual fault I bring in sense-- +Thy adverse party is thy advocate-- +And 'gainst myself a lawful plea commence: +Such civil war is in my love and hate +That I an accessary needs must be +To that sweet thief which sourly robs from me. +Let me confess that we two must be twain, +Although our undivided loves are one: +So shall those blots that do with me remain +Without thy help by me be borne alone. +In our two loves there is but one respect, +Though in our lives a separable spite, +Which though it alter not love's sole effect, +Yet doth it steal sweet hours from love's delight. +I may not evermore acknowledge thee, +Lest my bewailed guilt should do thee shame, +Nor thou with public kindness honour me, +Unless thou take that honour from thy name: +But do not so; I love thee in such sort +As, thou being mine, mine is thy good report. +As a decrepit father takes delight +To see his active child do deeds of youth, +So I, made lame by fortune's dearest spite, +Take all my comfort of thy worth and truth. +For whether beauty, birth, or wealth, or wit, +Or any of these all, or all, or more, +Entitled in thy parts do crowned sit, +I make my love engrafted to this store: +So then I am not lame, poor, nor despised, +Whilst that this shadow doth such substance give +That I in thy abundance am sufficed +And by a part of all thy glory live. +Look, what is best, that best I wish in thee: +This wish I have; then ten times happy me! \ No newline at end of file diff --git a/online_apiserver_test_maxbs/benchmarks/structured_schemas/structured_schema_1.json b/online_apiserver_test_maxbs/benchmarks/structured_schemas/structured_schema_1.json new file mode 100644 index 0000000000000000000000000000000000000000..13bd6b6d16c601e7e589952b49119205a407acd1 --- /dev/null +++ b/online_apiserver_test_maxbs/benchmarks/structured_schemas/structured_schema_1.json @@ -0,0 +1,19 @@ +{ + "type": "object", + "properties": { + "name": { "type": "string" }, + "email": { "type": "string" }, + "street": { "type": "string" }, + "city": { "type": "string" }, + "state": { "type": "string" }, + "zip": { "type": "string" }, + "phone": { "type": "string" }, + "website": { "type": "string" }, + "company": { "type": "string" }, + "age": { "type": "integer" } + }, + "required": [ + "name", + "email" + ] +} diff --git a/online_apiserver_test_maxbs/online_quick_check_maxbs.sh b/online_apiserver_test_maxbs/online_quick_check_maxbs.sh new file mode 100644 index 0000000000000000000000000000000000000000..bdc4a02fc45d8de452846c25ee043658b8f193e5 --- /dev/null +++ b/online_apiserver_test_maxbs/online_quick_check_maxbs.sh @@ -0,0 +1,138 @@ +#!/bin/bash + +HOST=127.0.0.1 +PORT=8081 +TIMEOUT=1200 # 监控超时设置 +INTERVAL=60 # 监控时间间隔 + +# 读取json配置文件 +json_data=$(cat auto_quick_check_config.json) + +DCU=$(echo $json_data | jq -r '.DCU') +vllm_version=$(echo $json_data | jq -r '.vllm_version') +pkg_version=$(echo $json_data | jq -r '.pkg_version') +dst_path=$(echo $json_data | jq -r '.dst_path') +items=$(echo $json_data | jq -c '.items[]') + +while read -r item; do + model_name=$(echo "$item" | jq -r '.model_name') + model_path=$(echo "$item" | jq -r '.model_path') + dtype=$(echo "$item" | jq -r '.dtype') + tensor_parallel=$(echo "$item" | jq -r '.tensor_parallel') + + seqlen_len=$(echo "$item" | jq -r '.seqlen_tuple | length') + ttft_len=$(echo "$item" | jq -r '.ttft_thres | length') + tpot_len=$(echo "$item" | jq -r '.tpot_thres | length') + + if ! [[ $seqlen_len -eq $tpot_len && $seqlen_len -eq $tpot_len ]]; then + echo "***********************************" + echo "测试项：模型 ${model_name} " + echo "输入输出序列、ttft阈值、tpot阈值数量存在不一致，无法测试最大并发量" + echo "跳过该测试项" + echo "***********************************" + continue + fi + + result_path=${dst_path}/${model_name}/ + if [ ! -f ${result_path} ]; then + mkdir ${result_path} -p + fi + + if [ -e "${result_path}output.csv" ] && [ -s "${result_path}output.csv" ]; then + : + else + echo "model_name,DCU,DCU nums,precision,input_len,output_len,bs,TTFT_mean(ms),TPOT_mean(ms),ITL_mean(ms),GenerateThroughput(tokens/s),TotalThroughput(tokens/s),Duration(s),version" > ${result_path}output.csv + fi + + echo $tensor_parallel | jq -c '.[]' | while read -r tp; do + # 运行服务端启动脚本 + nohup bash run_apiserver.sh $model_name $model_path $tp $dtype $HOST $PORT $result_path & + + start_time=$(date +%s) + while true; do + if nc -zv localhost $PORT; then # 检查端口是否打开 + for ((i=0; i&1 | tee ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-ttft_thre-${ttft_thre}-tpot_thre-${tpot_thre}-bs-${bs}.log + + output_throughput=`grep -a "^Output token throughput (tok/s):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-ttft_thre-${ttft_thre}-tpot_thre-${tpot_thre}-bs-${bs}.log | awk -F ' ' '{print $5}'` + total_throughput=`grep -a "^Total Token throughput (tok/s):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-ttft_thre-${ttft_thre}-tpot_thre-${tpot_thre}-bs-${bs}.log | awk -F ' ' '{print $5}'` + TTFT_mean=`grep -a "^Mean TTFT (ms):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-ttft_thre-${ttft_thre}-tpot_thre-${tpot_thre}-bs-${bs}.log | awk -F ' ' '{print $4}'` + TPOT_mean=`grep -a "^Mean TPOT (ms):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-ttft_thre-${ttft_thre}-tpot_thre-${tpot_thre}-bs-${bs}.log | awk -F ' ' '{print $4}'` + ITL_mean=`grep -a "^Mean ITL (ms):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-ttft_thre-${ttft_thre}-tpot_thre-${tpot_thre}-bs-${bs}.log | awk -F ' ' '{print $4}'` + duration=`grep -a "^Benchmark duration (s):" ${result_path}/${model_name}-tp-${tp}-input_len-${input_len}-output_len-${output_len}-ttft_thre-${ttft_thre}-tpot_thre-${tpot_thre}-bs-${bs}.log | awk -F ' ' '{print $4}'` + # total_time=$(python -c "print(${total_time} / 4)") + + echo "$model_name,$DCU,$tp,$dtype,$input_len,$output_len,$bs,$TTFT_mean,$TPOT_mean,$ITL_mean,$output_throughput,$total_throughput,$duration,$pkg_version" >> ${result_path}output.csv + sleep 10 + + condition=$(python <<- EOF +try: + ttft = float('$TTFT_mean') + tpot = float('$TPOT_mean') + ttft_thre = float('$ttft_thre') + tpot_thre = float('$tpot_thre') + + print(1 if (ttft <= ttft_thre and tpot <= tpot_thre) else 0) +except: + print(0) +EOF +) + if [ "$condition" -eq 1 ]; then + bs=$((bs + bs_interval)) + else + break + fi + done + done + break + else + current_time=$(date +%s) + elapsed_time=$((current_time - start_time)) + if [ $elapsed_time -ge $TIMEOUT ]; then + echo "ERR：PORT ${PORT} launch time out, exit!!!。" + exit 1 + fi + echo "PORT ${PORT} has not been launched yet, please wait...." + sleep $INTERVAL + fi + done + + pkill -f vllm + sleep 60 + done + +done <<< "$items" + diff --git a/online_apiserver_test_maxbs/run_apiserver.sh b/online_apiserver_test_maxbs/run_apiserver.sh new file mode 100644 index 0000000000000000000000000000000000000000..e580e6a2e8d6465c7cec657fa0ff7bdaaf77e3c8 --- /dev/null +++ b/online_apiserver_test_maxbs/run_apiserver.sh @@ -0,0 +1,69 @@ +#!/bin/bash + +## 产品部提供的vllm 0.9.2默认环境变量，用于测试非DS 671B模型 +export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +export ALLREDUCE_STREAM_WITH_COMPUTE=1 +export NCCL_MIN_NCHANNELS=16 +export NCCL_MAX_NCHANNELS=16 +# export VLLM_PCIE_USE_CUSTOM_ALLREDUCE=1 #K100-AI需要 + +## 根据TOPO进行NUMA绑核 +export VLLM_NUMA_BIND=1 +export VLLM_RANK0_NUMA=0 +export VLLM_RANK1_NUMA=0 +export VLLM_RANK2_NUMA=0 +export VLLM_RANK3_NUMA=0 +export VLLM_RANK4_NUMA=0 +export VLLM_RANK5_NUMA=0 +export VLLM_RANK6_NUMA=0 +export VLLM_RANK7_NUMA=0 + +## 项目上反馈的可能对性能有提升的环境变量，可酌情使用 +# export HSA_FORCE_FINE_GRAIN_PCIE=1 +# export NCCL_P2P_LEVEL=SYS +# export NCCL_LAUNCH_MODE=GROUP +# export VLLM_RPC_TIMEOUT=1800000 +# export VLLM_SPEC_DECODE_EAGER=1 +# export VLLM_ENFORCE_EAGER_BS_THRESHOLD=44 # K100AI设置为44，BW1000去掉该环境变量 + +# export VLLM_MLA_DISABLE=0 +# export VLLM_USE_FLASH_MLA=1 +# export VLLM_ZERO_OVERHEAD=1 # 该参数对某些模型性能有明显提升；但在某些环境下表现异常，根据实际情况斟酌选用；且使用后对Qwen3模型会关闭其thinking功能 + +# export W8A8_SUPPORT_METHODS=3 # 对W8A8的量化模型有提升 +# export ROCBLAS_INT8_ENABLE=0 # 对W8A8的量化模型有提升 + +# export VLLM_USE_FLASH_ATTN_PA=0 # 解决精度乱码问题 + +# 禁止生成core文件 +ulimit -c 0 + +# 模型参数 +model_name=$1 +model_path=$2 +tp=$3 +dtype=$4 +host=$5 +port=$6 +dst_path=$7 +case $dtype in + "float16" | "bfloat16") + extra="--dtype ${dtype}" ;; + "gptq-int8" | "gptq-int4") + export GPTQ_CK_GEMMBS=15000 + extra="--quantization gptq" ;; + "compressed-tensors-w8a8") + extra="--quantization compressed-tensors" ;; + "awq") + export AWQ_CK_GEMMBS=15000 + extra="--quantization awq" ;; +esac + +# vllm 0.9.2 默认使用V1 Engine及Prefix Caching功能开启（推荐），如果有测试需求可以自行切换V0 Engine或关闭Prefix Caching +# 使用V0 Engine：VLLM_USE_V1=0 +# 开启/关闭Prefix Caching：--enable-chunked-prefill / --no-enable-prefix-caching + +date=$(date +"%Y-%m-%d-%H-%M-%S") +python3 -m vllm.entrypoints.openai.api_server --served-model-name $model_name --gpu-memory-utilization 0.95 $extra \ + --model $model_path --tensor-parallel-size $tp --host $host --port $port --max-model-len 32768 \ + 2>&1 | tee ${dst_path}/launcher_${model_name}_tp_${tp}_dtype_${dtype}_${date}.log