Skip to content

GitLab

Commit 38d80967 authored by zhuwenwen's avatar zhuwenwen
Browse files

Merge tag 'v0.10.2rc2' into v0.10.2rc2-ori

parents 33650733 880c741b
Hide whitespace changes
Inline Side-by-side
Showing with 806 additions and 2283 deletions
.gitignore
View file @ 38d80967
......@@ -4,7 +4,7 @@
# vllm-flash-attn built from source
vllm/vllm_flash_attn/*
# triton jit
# triton jit
.triton
# Byte-compiled / optimized / DLL files
......@@ -177,6 +177,14 @@ cython_debug/
# VSCode
.vscode/
# Claude
CLAUDE.md
.claude/
# Codex
AGENTS.md
.codex/
# DS Store
.DS_Store
......@@ -209,4 +217,4 @@ shellcheck*/
csrc/moe/marlin_moe_wna16/kernel_*
# Ignore ep_kernels_workspace folder
ep_kernels_workspace/
\ No newline at end of file
ep_kernels_workspace/
.yapfignore
View file @ 38d80967
collect_env.py
vllm/model_executor/layers/fla/ops/*.py
MANIFEST.in
View file @ 38d80967
......@@ -2,7 +2,6 @@ include LICENSE
include requirements/common.txt
include requirements/cuda.txt
include requirements/rocm.txt
include requirements/neuron.txt
include requirements/cpu.txt
include CMakeLists.txt
......
README.md
View file @ 38d80967
......@@ -14,19 +14,24 @@ Easy, fast, and cheap LLM serving for everyone
| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
</p>
---
Join us at the [PyTorch Conference, October 22-23](https://events.linuxfoundation.org/pytorch-conference/) and [Ray Summit, November 3-5](https://www.anyscale.com/ray-summit/2025) in San Francisco for our latest updates on vLLM and to meet the vLLM team! Register now for the largest vLLM community events of the year!
---
*Latest News* 🔥
- [2025/08] We hosted [vLLM Shenzhen Meetup](https://mp.weixin.qq.com/s/k8ZBO1u2_2odgiKWH_GVTQ) focusing on the ecosystem around vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Ua2SVKVSu-wp5vou_6ElraDt2bnKhiEA).
- [2025/08] We hosted [vLLM Singapore Meetup](https://www.sginnovate.com/event/vllm-sg-meet). We shared V1 updates, disaggregated serving and MLLM speedups with speakers from Embedded LLM, AMD, WekaIO, and A*STAR. Please find the meetup slides [here](https://drive.google.com/drive/folders/1ncf3GyqLdqFaB6IeB834E5TZJPLAOiXZ?usp=sharing).
- [2025/08] We hosted [vLLM Shanghai Meetup](https://mp.weixin.qq.com/s/pDmAXHcN7Iqc8sUKgJgGtg) focusing on building, developing, and integrating with vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1OvLx39wnCGy_WKq8SiVKf7YcxxYI3WCH).
- [2025/08] We hosted [vLLM Korea Meetup](https://luma.com/cgcgprmh) with Red Hat and Rebellions! We shared the latest advancements in vLLM along with project spotlights from the vLLM Korea community. Please find the meetup slides [here](https://drive.google.com/file/d/1bcrrAE1rxUgx0mjIeOWT6hNe2RefC5Hm/view).
- [2025/08] We hosted [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/dgkWg1WFpWGO2jCdTqQHxA) focusing on large-scale LLM deployment! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Pid6NSFLU43DZRi0EaTcPgXsAzDvbBqF) and the recording [here](https://www.chaspark.com/#/live/1166916873711665152).
- [2025/05] vLLM is now a hosted project under PyTorch Foundation! Please find the announcement [here](https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/).
- [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
<details>
<summary>Previous News</summary>
- [2025/08] We hosted [vLLM Korea Meetup](https://luma.com/cgcgprmh) with Red Hat and Rebellions! We shared the latest advancements in vLLM along with project spotlights from the vLLM Korea community. Please find the meetup slides [here](https://drive.google.com/file/d/1bcrrAE1rxUgx0mjIeOWT6hNe2RefC5Hm/view).
- [2025/08] We hosted [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/dgkWg1WFpWGO2jCdTqQHxA) focusing on large-scale LLM deployment! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Pid6NSFLU43DZRi0EaTcPgXsAzDvbBqF) and the recording [here](https://www.chaspark.com/#/live/1166916873711665152).
- [2025/05] We hosted [NYC vLLM Meetup](https://lu.ma/c1rqyf1f)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1_q_aW_ioMJWUImf1s1YM-ZhjXz8cUeL0IJvaquOYBeA/edit?usp=sharing).
- [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
- [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
......
benchmarks/README.md
View file @ 38d80967
......@@ -95,6 +95,24 @@ become available.
<td style="text-align: center;">✅</td>
<td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
</tr>
<tr>
<td><strong>HuggingFace-MTBench</strong></td>
<td style="text-align: center;">✅</td>
<td style="text-align: center;">✅</td>
<td><code>philschmid/mt-bench</code></td>
</tr>
<tr>
<td><strong>HuggingFace-Blazedit</strong></td>
<td style="text-align: center;">✅</td>
<td style="text-align: center;">✅</td>
<td><code>vdaita/edit_5k_char</code>, <code>vdaita/edit_10k_char</code></td>
</tr>
<tr>
<td><strong>Spec Bench</strong></td>
<td style="text-align: center;">✅</td>
<td style="text-align: center;">✅</td>
<td><code>wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl</code></td>
</tr>
<tr>
<td><strong>Custom</strong></td>
<td style="text-align: center;">✅</td>
......@@ -110,7 +128,12 @@ become available.
🚧: to be supported
**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`.
For local `dataset-path`, please set `hf-name` to its Hugging Face ID like
```bash
--dataset-path /datasets/VisionArena-Chat/ --hf-name lmarena-ai/VisionArena-Chat
```
## 🚀 Example - Online Benchmark
......@@ -234,6 +257,43 @@ vllm bench serve \
--num-prompts 2048
```
### Spec Bench 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}'
```
[SpecBench dataset](https://github.com/hemingkx/Spec-Bench)
Run all categories:
``` bash
# Download the dataset using:
# wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl
vllm bench serve \
--model meta-llama/Meta-Llama-3-8B-Instruct \
--dataset-name spec_bench \
--dataset-path "<YOUR_DOWNLOADED_PATH>/data/spec_bench/question.jsonl" \
--num-prompts -1
```
Available categories include `[writing, roleplay, reasoning, math, coding, extraction, stem, humanities, translation, summarization, qa, math_reasoning, rag]`.
Run only a specific category like "summarization":
``` bash
vllm bench serve \
--model meta-llama/Meta-Llama-3-8B-Instruct \
--dataset-name spec_bench \
--dataset-path "<YOUR_DOWNLOADED_PATH>/data/spec_bench/question.jsonl" \
--num-prompts -1
--spec-bench-category "summarization"
```
### Other HuggingFaceDataset Examples
```bash
......@@ -290,6 +350,18 @@ vllm bench serve \
--num-prompts 80
```
`vdaita/edit_5k_char` or `vdaita/edit_10k_char`:
``` bash
vllm bench serve \
--model Qwen/QwQ-32B \
--dataset-name hf \
--dataset-path vdaita/edit_5k_char \
--num-prompts 90 \
--blazedit-min-distance 0.01 \
--blazedit-max-distance 0.99
```
### Running With Sampling Parameters
When using OpenAI-compatible backends such as `vllm`, optional sampling
......@@ -689,7 +761,7 @@ python -m vllm.entrypoints.openai.api_server \
Send requests with images:
```bash
python benchmarks/benchmark_serving.py \
vllm bench serve \
--backend openai-chat \
--model Qwen/Qwen2.5-VL-7B-Instruct \
--dataset-name sharegpt \
......@@ -716,7 +788,7 @@ python -m vllm.entrypoints.openai.api_server \
Send requests with videos:
```bash
python benchmarks/benchmark_serving.py \
vllm bench serve \
--backend openai-chat \
--model Qwen/Qwen2.5-VL-7B-Instruct \
--dataset-name sharegpt \
......
benchmarks/auto_tune/README.md
View file @ 38d80967
......@@ -31,6 +31,12 @@ cd vllm
You must set the following variables at the top of the script before execution.
Note: You can also override the default values below via environment variables when running the script.
```bash
MODEL=meta-llama/Llama-3.3-70B-Instruct SYSTEM=TPU TP=8 DOWNLOAD_DIR='' INPUT_LEN=128 OUTPUT_LEN=2048 MAX_MODEL_LEN=2300 MIN_CACHE_HIT_PCT=0 MAX_LATENCY_ALLOWED_MS=100000000000 NUM_SEQS_LIST="128 256" NUM_BATCHED_TOKENS_LIST="1024 2048 4096" VLLM_LOGGING_LEVEL=DEBUG bash auto_tune.sh
```
| Variable | Description | Example Value |
| --- | --- | --- |
| `BASE` | **Required.** The absolute path to the parent directory of your vLLM repository directory. | `"$HOME"` |
......
benchmarks/auto_tune/auto_tune.sh
View file @ 38d80967
......@@ -5,25 +5,41 @@
TAG=$(date +"%Y_%m_%d_%H_%M")
SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
BASE="$SCRIPT_DIR/../../.."
MODEL="meta-llama/Llama-3.1-8B-Instruct"
SYSTEM="TPU"
TP=1
DOWNLOAD_DIR=""
INPUT_LEN=4000
OUTPUT_LEN=16
MAX_MODEL_LEN=4096
MIN_CACHE_HIT_PCT=0
MAX_LATENCY_ALLOWED_MS=100000000000
NUM_SEQS_LIST="128 256"
NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096"
VLLM_LOGGING_LEVEL=${VLLM_LOGGING_LEVEL:-INFO}
BASE=${BASE:-"$SCRIPT_DIR/../../.."}
MODEL=${MODEL:-"meta-llama/Llama-3.1-8B-Instruct"}
SYSTEM=${SYSTEM:-"TPU"}
TP=${TP:-1}
DOWNLOAD_DIR=${DOWNLOAD_DIR:-""}
INPUT_LEN=${INPUT_LEN:-4000}
OUTPUT_LEN=${OUTPUT_LEN:-16}
MAX_MODEL_LEN=${MAX_MODEL_LEN:-4096}
MIN_CACHE_HIT_PCT=${MIN_CACHE_HIT_PCT:-0}
MAX_LATENCY_ALLOWED_MS=${MAX_LATENCY_ALLOWED_MS:-100000000000}
NUM_SEQS_LIST=${NUM_SEQS_LIST:-"128 256"}
NUM_BATCHED_TOKENS_LIST=${NUM_BATCHED_TOKENS_LIST:-"512 1024 2048 4096"}
LOG_FOLDER="$BASE/auto-benchmark/$TAG"
RESULT="$LOG_FOLDER/result.txt"
PROFILE_PATH="$LOG_FOLDER/profile"
echo "result file: $RESULT"
echo "model: $MODEL"
echo "====================== AUTO TUNE PARAMETERS ===================="
echo "SCRIPT_DIR=$SCRIPT_DIR"
echo "BASE=$BASE"
echo "MODEL=$MODEL"
echo "SYSTEM=$SYSTEM"
echo "TP=$TP"
echo "DOWNLOAD_DIR=$DOWNLOAD_DIR"
echo "INPUT_LEN=$INPUT_LEN"
echo "OUTPUT_LEN=$OUTPUT_LEN"
echo "MAX_MODEL_LEN=$MAX_MODEL_LEN"
echo "MIN_CACHE_HIT_PCT=$MIN_CACHE_HIT_PCT"
echo "MAX_LATENCY_ALLOWED_MS=$MAX_LATENCY_ALLOWED_MS"
echo "NUM_SEQS_LIST=$NUM_SEQS_LIST"
echo "NUM_BATCHED_TOKENS_LIST=$NUM_BATCHED_TOKENS_LIST"
echo "VLLM_LOGGING_LEVEL=$VLLM_LOGGING_LEVEL"
echo "RESULT_FILE=$RESULT"
echo "====================== AUTO TUNEPARAMETERS ===================="
rm -rf $LOG_FOLDER
rm -rf $PROFILE_PATH
......@@ -213,7 +229,7 @@ run_benchmark() {
pkill -if vllm
sleep 10
printf '=%.0s' $(seq 1 20)
echo "===================="
return 0
}
......
benchmarks/benchmark_block_pool.py
View file @ 38d80967
......@@ -57,7 +57,7 @@ def invoke_main() -> None:
"--num-iteration",
type=int,
default=1000,
help="Number of iterations to run to stablize final data readings",
help="Number of iterations to run to stabilize final data readings",
)
parser.add_argument(
"--allocate-blocks",
......
benchmarks/benchmark_dataset.py
View file @ 38d80967
......@@ -403,7 +403,7 @@ class RandomDataset(BenchmarkDataset):
# [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.
# the encoded sequence is truncated before being decoded again.
total_input_len = prefix_len + int(input_lens[i])
re_encoded_sequence = tokenizer.encode(prompt, add_special_tokens=False)[
:total_input_len
......
benchmarks/benchmark_latency.py
View file @ 38d80967
# 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 sys
import argparse
import dataclasses
import json
import os
import time
from typing import Any, Optional
import numpy as np
from tqdm import tqdm
from typing_extensions import deprecated
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)
@deprecated(
"benchmark_latency.py is deprecated and will be removed in a "
"future version. Please use 'vllm bench latency' instead.",
)
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)
if __name__ == "__main__":
print("""DEPRECATED: This script has been moved to the vLLM CLI.
return parser
Please use the following command instead:
vllm bench latency
For help with the new command, run:
vllm bench latency --help
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)
Alternatively, you can run the new command directly with:
python -m vllm.entrypoints.cli.main bench latency --help
""")
sys.exit(1)
benchmarks/benchmark_ngram_proposer.py
View file @ 38d80967
......@@ -77,7 +77,7 @@ def invoke_main() -> None:
"--num-iteration",
type=int,
default=100,
help="Number of iterations to run to stablize final data readings",
help="Number of iterations to run to stabilize final data readings",
)
parser.add_argument(
"--num-req", type=int, default=128, help="Number of requests in the batch"
......
benchmarks/benchmark_serving.py
View file @ 38d80967
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
r"""Benchmark online serving throughput.
import sys
On the server side, run one of the following commands:
vLLM OpenAI API server
vllm serve <your_model> \
--swap-space 16
On the client side, run:
python benchmarks/benchmark_serving.py \
--backend <backend> \
--model <your_model> \
--dataset-name sharegpt \
--dataset-path <path to dataset> \
--request-rate <request_rate> \ # By default <request_rate> is inf
--num-prompts <num_prompts> # By default <num_prompts> 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 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 typing_extensions import deprecated
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
from vllm.benchmarks.serve import get_request
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 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)
or (
isinstance(test_mm_content, list)
and all(isinstance(item, dict) for item in test_mm_content)
)
), "multi_modal_data must be a dict or list[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_id = (
request.prompt,
request.prompt_len,
request.expected_output_len,
request.multi_modal_data,
request.request_id,
)
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,
request_id=request_id,
)
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 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))
if max_concurrency is not None:
print("{:<40} {:<10}".format("Maximum request concurrency:", max_concurrency))
if request_rate != float("inf"):
print("{:<40} {:<10.2f}".format("Request rate configured (RPS):", request_rate))
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)
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")
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)
@deprecated(
"benchmark_serving.py is deprecated and will be removed in a future "
"version. Please use 'vllm bench serve' instead.",
)
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,
request_id_prefix=args.request_id_prefix,
)
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,
request_id_prefix=args.request_id_prefix,
)
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,
request_id_prefix=args.request_id_prefix,
)
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,
no_stream=args.no_stream,
).sample(
num_requests=args.num_prompts,
tokenizer=tokenizer,
output_len=args.hf_output_len,
request_id_prefix=args.request_id_prefix,
)
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,
request_id_prefix=args.request_id_prefix,
),
"burstgpt": lambda: BurstGPTDataset(
random_seed=args.seed, dataset_path=args.dataset_path
).sample(
tokenizer=tokenizer,
num_requests=args.num_prompts,
request_id_prefix=args.request_id_prefix,
),
"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,
request_id_prefix=args.request_id_prefix,
),
}
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(
"--no-stream",
action="store_true",
help="Do not load the dataset in streaming mode.",
)
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",
)
parser.add_argument(
"--request-id-prefix",
type=str,
required=False,
default="benchmark-serving",
help="Specify the prefix of request id.",
)
# 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.",
)
if __name__ == "__main__":
print("""DEPRECATED: This script has been moved to the vLLM CLI.
return parser
Please use the following command instead:
vllm bench serve
For help with the new command, run:
vllm bench serve --help
if __name__ == "__main__":
parser = create_argument_parser()
args = parser.parse_args()
main(args)
Alternatively, you can run the new command directly with:
python -m vllm.entrypoints.cli.main bench serve --help
""")
sys.exit(1)
benchmarks/benchmark_serving_structured_output.py
View file @ 38d80967
......@@ -998,7 +998,7 @@ def create_argument_parser():
"--percentile-metrics",
type=str,
default="ttft,tpot,itl",
help="Comma-separated list of selected metrics to report percentils. "
help="Comma-separated list of selected metrics to report percentiles. "
"This argument specifies the metrics to report percentiles. "
'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
'Default value is "ttft,tpot,itl".',
......
benchmarks/benchmark_throughput.py
View file @ 38d80967
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Benchmark offline inference throughput."""
import sys
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 typing_extensions import deprecated
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"
# 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=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":
common_kwargs["no_stream"] = args.no_stream
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)
@deprecated(
"benchmark_throughput.py is deprecated and will be removed in a "
"future version. Please use 'vllm bench throughput' instead.",
)
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(
"--no-stream",
action="store_true",
help="Do not load the dataset in streaming mode.",
)
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: <prompt_or_response>]]]]",
)
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)
if __name__ == "__main__":
print("""DEPRECATED: This script has been moved to the vLLM CLI.
return parser
Please use the following command instead:
vllm bench throughput
For help with the new command, run:
vllm bench throughput --help
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)
Alternatively, you can run the new command directly with:
python -m vllm.entrypoints.cli.main bench throughput --help
""")
sys.exit(1)
benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh
View file @ 38d80967
......@@ -62,7 +62,7 @@ benchmark() {
--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}' &
'{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
CUDA_VISIBLE_DEVICES=1 python3 \
......@@ -72,7 +72,7 @@ benchmark() {
--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}' &
'{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
wait_for_server 8100
wait_for_server 8200
......
benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh
View file @ 38d80967
......@@ -69,7 +69,7 @@ launch_disagg_prefill() {
--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}' &
'{"kv_connector":"P2pNcclConnector","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 \
......@@ -78,7 +78,7 @@ launch_disagg_prefill() {
--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}' &
'{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
wait_for_server 8100
wait_for_server 8200
......
benchmarks/kernels/bench_block_fp8_gemm.py
View file @ 38d80967
......@@ -4,7 +4,10 @@
import torch
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
w8a8_block_fp8_matmul,
apply_w8a8_block_fp8_linear,
)
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
CUTLASS_BLOCK_FP8_SUPPORTED,
)
from vllm.platforms import current_platform
from vllm.triton_utils import triton as vllm_triton
......@@ -29,7 +32,7 @@ DEEPSEEK_V3_SHAPES = [
]
def build_w8a8_block_fp8_runner(M, N, K, block_size, device):
def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
"""Build runner function for w8a8 block fp8 matmul."""
factor_for_scale = 1e-2
......@@ -37,37 +40,54 @@ def build_w8a8_block_fp8_runner(M, N, K, block_size, device):
fp8_max, fp8_min = fp8_info.max, fp8_info.min
# Create random FP8 tensors
A_fp32 = (torch.rand(M, K, dtype=torch.float32, device=device) - 0.5) * 2 * fp8_max
A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
A_ref = (torch.rand(M, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
B_fp32 = (torch.rand(N, K, dtype=torch.float32, device=device) - 0.5) * 2 * fp8_max
B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
B_ref = (torch.rand(N, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
B = B_ref.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
# Create scales
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=device) * factor_for_scale
Bs = (
torch.rand(n_tiles, k_tiles, dtype=torch.float32, device=device)
* factor_for_scale
)
# SM90 CUTLASS requires row-major format for scales
if use_cutlass and current_platform.is_device_capability(90):
Bs = Bs.T.contiguous()
def run():
return w8a8_block_fp8_matmul(A, B, As, Bs, block_size, torch.bfloat16)
if use_cutlass:
return apply_w8a8_block_fp8_linear(
A_ref, B, block_size, Bs, cutlass_block_fp8_supported=True
)
else:
return apply_w8a8_block_fp8_linear(
A_ref, B, block_size, Bs, cutlass_block_fp8_supported=False
)
return run
# Determine available providers
available_providers = ["torch-bf16", "w8a8-block-fp8-triton"]
plot_title = "BF16 vs W8A8 Block FP8 GEMMs"
if CUTLASS_BLOCK_FP8_SUPPORTED:
available_providers.append("w8a8-block-fp8-cutlass")
@vllm_triton.testing.perf_report(
vllm_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=["torch-bf16", "w8a8-block-fp8"],
line_names=["torch-bf16", "w8a8-block-fp8"],
line_vals=available_providers,
line_names=available_providers,
ylabel="TFLOP/s (larger is better)",
plot_name="BF16 vs W8A8 Block FP8 GEMMs",
args={},
......@@ -85,11 +105,22 @@ def benchmark_tflops(batch_size, provider, N, K, block_size=(128, 128)):
ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
)
else: # w8a8-block-fp8
run_w8a8 = build_w8a8_block_fp8_runner(M, N, K, block_size, device)
elif provider == "w8a8-block-fp8-triton":
run_w8a8_triton = build_w8a8_block_fp8_runner(
M, N, K, block_size, device, use_cutlass=False
)
ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
lambda: run_w8a8_triton(), quantiles=quantiles
)
elif provider == "w8a8-block-fp8-cutlass":
run_w8a8_cutlass = build_w8a8_block_fp8_runner(
M, N, K, block_size, device, use_cutlass=True
)
ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
lambda: run_w8a8(), quantiles=quantiles
lambda: run_w8a8_cutlass(), quantiles=quantiles
)
else:
raise ValueError(f"Unknown provider: {provider}")
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)
......
benchmarks/kernels/benchmark_activation.py 0 → 100644
View file @ 38d80967
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# benchmark custom activation op performance
import itertools
import torch
import vllm.model_executor.layers.activation # noqa F401
from vllm.model_executor.custom_op import CustomOp
from vllm.platforms import current_platform
from vllm.triton_utils import triton
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
batch_size_range = [1, 16, 32, 64, 128]
seq_len_range = [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
intermediate_size = [3072, 9728, 12288]
configs = list(itertools.product(batch_size_range, seq_len_range, intermediate_size))
def benchmark_activation(
batch_size: int,
seq_len: int,
intermediate_size: int,
provider: str,
func_name: str,
dtype: torch.dtype,
):
device = "cuda"
num_tokens = batch_size * seq_len
dim = intermediate_size
current_platform.seed_everything(42)
torch.set_default_device(device)
if func_name == "gelu_and_mul":
layer = CustomOp.op_registry[func_name](approximate="none")
elif func_name == "gelu_and_mul_tanh":
layer = CustomOp.op_registry["gelu_and_mul"](approximate="tanh")
elif func_name == "fatrelu_and_mul":
threshold = 0.5
layer = CustomOp.op_registry[func_name](threshold)
else:
layer = CustomOp.op_registry[func_name]()
x = torch.randn(num_tokens, dim, dtype=dtype, device=device)
compiled_layer = torch.compile(layer.forward_native)
if provider == "custom":
fn = lambda: layer(x)
elif provider == "compiled":
fn = lambda: compiled_layer(x)
ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
fn, quantiles=[0.5, 0.2, 0.8]
)
return ms, max_ms, min_ms
if __name__ == "__main__":
parser = FlexibleArgumentParser(description="Benchmark the custom activation op.")
parser.add_argument(
"--func-name",
type=str,
choices=[
"mul_and_silu",
"silu_and_mul",
"gelu_and_mul",
"gelu_and_mul_tanh",
"fatrelu_and_mul",
"swigluoai_and_mul",
"gelu_new",
"gelu_fast",
"quick_gelu",
],
default="silu_and_mul",
)
parser.add_argument(
"--dtype", type=str, choices=["half", "bfloat16", "float"], default="bfloat16"
)
args = parser.parse_args()
assert args
func_name = args.func_name
dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype]
perf_report = triton.testing.perf_report(
triton.testing.Benchmark(
x_names=["batch_size", "seq_len", "intermediate_size"],
x_vals=configs,
line_arg="provider",
line_vals=["custom", "compiled"],
line_names=["Custom OP", "Compiled"],
styles=[("blue", "-"), ("green", "-")],
ylabel="ms",
plot_name=f"{func_name}-op-performance",
args={},
)
)
perf_report(
lambda batch_size, seq_len, intermediate_size, provider: benchmark_activation(
batch_size, seq_len, intermediate_size, provider, func_name, dtype
)
).run(print_data=True)
benchmarks/kernels/benchmark_device_communicators.py 0 → 100644
View file @ 38d80967
#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Benchmark script for device communicators:
CustomAllreduce (oneshot, twoshot), PyNcclCommunicator,
and SymmMemCommunicator (multimem, two-shot).
Usage:
torchrun --nproc_per_node=<N> benchmark_device_communicators.py [options]
Example:
torchrun --nproc_per_node=2 benchmark_device_communicators.py
--sequence-lengths 512 1024 2048 --num-warmup 10 --num-trials 100
"""
import json
import os
import time
from contextlib import nullcontext
from typing import Callable, Optional
import torch
import torch.distributed as dist
from torch.distributed import ProcessGroup
from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
from vllm.distributed.device_communicators.symm_mem import SymmMemCommunicator
from vllm.logger import init_logger
from vllm.utils import FlexibleArgumentParser
logger = init_logger(__name__)
# Default sequence lengths to benchmark
DEFAULT_SEQUENCE_LENGTHS = [128, 512, 1024, 2048, 4096, 8192]
# Fixed hidden size and dtype for all benchmarks
HIDDEN_SIZE = 8192
BENCHMARK_DTYPE = torch.bfloat16
# CUDA graph settings
CUDA_GRAPH_CAPTURE_CYCLES = 10
class CommunicatorBenchmark:
"""Benchmark class for testing device communicators."""
def __init__(
self,
rank: int,
world_size: int,
device: torch.device,
cpu_group: ProcessGroup,
sequence_lengths: list[int],
):
self.rank = rank
self.world_size = world_size
self.device = device
self.cpu_group = cpu_group
# Calculate max_size_override based on largest sequence length
max_seq_len = max(sequence_lengths)
max_tensor_elements = max_seq_len * HIDDEN_SIZE
self.max_size_override = max_tensor_elements * BENCHMARK_DTYPE.itemsize + 1
# Initialize communicators
self.custom_allreduce = None
self.pynccl_comm = None
self.symm_mem_comm = None
self.symm_mem_comm_multimem = None
self.symm_mem_comm_two_shot = None
self._init_communicators()
def _init_communicators(self):
"""Initialize all available communicators."""
try:
self.custom_allreduce = CustomAllreduce(
group=self.cpu_group,
device=self.device,
max_size=self.max_size_override,
)
if not self.custom_allreduce.disabled:
logger.info("Rank %s: CustomAllreduce initialized", self.rank)
else:
logger.info("Rank %s: CustomAllreduce disabled", self.rank)
except Exception as e:
logger.warning(
"Rank %s: Failed to initialize CustomAllreduce: %s", self.rank, e
)
self.custom_allreduce = None
try:
self.pynccl_comm = PyNcclCommunicator(
group=self.cpu_group, device=self.device
)
if not self.pynccl_comm.disabled:
logger.info("Rank %s: PyNcclCommunicator initialized", self.rank)
else:
logger.info("Rank %s: PyNcclCommunicator disabled", self.rank)
self.pynccl_comm = None
except Exception as e:
logger.warning(
"Rank %s: Failed to initialize PyNcclCommunicator: %s", self.rank, e
)
self.pynccl_comm = None
# Initialize variants for SymmMemCommunicator
try:
self.symm_mem_comm_multimem = SymmMemCommunicator(
group=self.cpu_group,
device=self.device,
force_multimem=True,
max_size_override=self.max_size_override,
)
if not self.symm_mem_comm_multimem.disabled:
logger.info(
"Rank %s: SymmMemCommunicator (multimem) initialized", self.rank
)
else:
self.symm_mem_comm_multimem = None
except Exception as e:
logger.warning(
"Rank %s: Failed to initialize SymmMemCommunicator (multimem): %s",
self.rank,
e,
)
self.symm_mem_comm_multimem = None
try:
self.symm_mem_comm_two_shot = SymmMemCommunicator(
group=self.cpu_group,
device=self.device,
force_multimem=False,
max_size_override=self.max_size_override,
)
if not self.symm_mem_comm_two_shot.disabled:
logger.info(
"Rank %s: SymmMemCommunicator (two_shot) initialized", self.rank
)
else:
self.symm_mem_comm_two_shot = None
except Exception as e:
logger.warning(
"Rank %s: Failed to initialize SymmMemCommunicator (two_shot): %s",
self.rank,
e,
)
self.symm_mem_comm_two_shot = None
def benchmark_allreduce(
self, sequence_length: int, num_warmup: int, num_trials: int
) -> dict[str, float]:
"""Benchmark allreduce operations for all available communicators."""
results = {}
# Define communicators with their benchmark functions
communicators = []
if self.custom_allreduce is not None:
comm = self.custom_allreduce
# CustomAllreduce one-shot
communicators.append(
(
"ca_1stage",
lambda t, c=comm: c.custom_all_reduce(t),
lambda t, c=comm: c.should_custom_ar(t),
comm.capture(),
"1stage", # env variable value
)
)
# CustomAllreduce two-shot
communicators.append(
(
"ca_2stage",
lambda t, c=comm: c.custom_all_reduce(t),
lambda t, c=comm: c.should_custom_ar(t),
comm.capture(),
"2stage", # env variable value
)
)
if self.pynccl_comm is not None:
comm = self.pynccl_comm
communicators.append(
(
"pynccl",
lambda t, c=comm: c.all_reduce(t),
lambda t: True, # Always available if initialized
nullcontext(),
None, # no env variable needed
)
)
if self.symm_mem_comm_multimem is not None:
comm = self.symm_mem_comm_multimem
communicators.append(
(
"symm_mem_multimem",
lambda t, c=comm: c.all_reduce(t),
lambda t, c=comm: c.should_use_symm_mem(t),
nullcontext(),
None, # no env variable needed
)
)
if self.symm_mem_comm_two_shot is not None:
comm = self.symm_mem_comm_two_shot
communicators.append(
(
"symm_mem_two_shot",
lambda t, c=comm: c.all_reduce(t),
lambda t, c=comm: c.should_use_symm_mem(t),
nullcontext(),
None, # no env variable needed
)
)
# Benchmark each communicator
for name, allreduce_fn, should_use_fn, context, env_var in communicators:
# Set environment variable if needed
if env_var is not None:
os.environ["VLLM_CUSTOM_ALLREDUCE_ALGO"] = env_var
else:
# Clear the environment variable to avoid interference
os.environ.pop("VLLM_CUSTOM_ALLREDUCE_ALGO", None)
latency = self.benchmark_allreduce_single(
sequence_length,
allreduce_fn,
should_use_fn,
context,
num_warmup,
num_trials,
)
if latency is not None:
results[name] = latency
return results
def benchmark_allreduce_single(
self,
sequence_length: int,
allreduce_fn: Callable[[torch.Tensor], Optional[torch.Tensor]],
should_use_fn: Callable[[torch.Tensor], bool],
context,
num_warmup: int,
num_trials: int,
) -> Optional[float]:
"""Benchmark method with CUDA graph optimization."""
try:
# Create test tensor (2D: sequence_length x hidden_size)
tensor = torch.randn(
sequence_length, HIDDEN_SIZE, dtype=BENCHMARK_DTYPE, device=self.device
)
if not should_use_fn(tensor):
return None
torch.cuda.synchronize()
stream = torch.cuda.Stream()
with torch.cuda.stream(stream):
graph_input = tensor.clone()
# Warmup before capture
for _ in range(3):
allreduce_fn(graph_input)
# Capture the graph using context manager
with context:
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph):
for _ in range(CUDA_GRAPH_CAPTURE_CYCLES):
allreduce_fn(graph_input)
torch.cuda.synchronize()
for _ in range(num_warmup):
graph.replay()
torch.cuda.synchronize()
torch.cuda.synchronize()
start_time = time.perf_counter()
for _ in range(num_trials):
graph.replay()
torch.cuda.synchronize()
end_time = time.perf_counter()
# Convert to ms and divide by CUDA_GRAPH_CAPTURE_CYCLES
return (
(end_time - start_time) / num_trials / CUDA_GRAPH_CAPTURE_CYCLES * 1000
)
except Exception as e:
logger.error("CUDA graph benchmark failed: %s", e)
raise RuntimeError(
f"CUDA graph benchmark failed for communicator: {e}"
) from e
def _calculate_speedup_info(comm_results: dict[str, float]) -> str:
"""Calculate speedup information for a single tensor size."""
if not comm_results:
return "N/A"
# Find the fastest communicator
fastest_comm = min(comm_results.keys(), key=lambda k: comm_results[k])
fastest_time = comm_results[fastest_comm]
# Calculate speedup vs PyNccl if available
if "pynccl" in comm_results:
pynccl_time = comm_results["pynccl"]
speedup = pynccl_time / fastest_time
return f"{fastest_comm} ({speedup:.2f}x)"
else:
return f"{fastest_comm} (N/A)"
def print_results(
results: dict[str, dict[str, float]], sequence_lengths: list[int], world_size: int
):
"""Print benchmark results in a formatted table."""
print(f"\n{'=' * 130}")
print("Device Communicator Benchmark Results")
print(
f"World Size: {world_size}, Data Type: {BENCHMARK_DTYPE}, "
f"Hidden Size: {HIDDEN_SIZE}"
)
print(f"{'=' * 130}")
# Get all communicator names
all_comms = set()
for size_results in results.values():
all_comms.update(size_results.keys())
all_comms = sorted(list(all_comms))
# Print header
header = f"{'Tensor Shape':<20}{'Tensor Size':<15}"
for comm in all_comms:
header += f"{comm:<20}"
header += f"{'Best (Speedup vs PyNccl)':<30}"
print(header)
print("-" * len(header))
# Print results for each sequence length
for seq_len in sequence_lengths:
if seq_len in results:
# Calculate tensor size in elements and bytes
tensor_elements = seq_len * HIDDEN_SIZE
tensor_bytes = tensor_elements * BENCHMARK_DTYPE.itemsize
# Format tensor size (MB)
tensor_size_mb = tensor_bytes / (1024 * 1024)
tensor_size_str = f"{tensor_size_mb:.2f} MB"
# Format tensor shape
tensor_shape = f"({seq_len}, {HIDDEN_SIZE})"
row = f"{tensor_shape:<20}{tensor_size_str:<15}"
for comm in all_comms:
if comm in results[seq_len]:
row += f"{results[seq_len][comm]:<20.3f}"
else:
row += f"{'N/A':<20}"
# Calculate speedup information
speedup_info = _calculate_speedup_info(results[seq_len])
row += f"{speedup_info:<30}"
print(row)
print(f"{'=' * 130}")
print("All times are in milliseconds (ms) per allreduce operation")
print("Speedup column shows: fastest_algorithm (speedup_vs_pynccl)")
def main():
parser = FlexibleArgumentParser(description="Benchmark device communicators")
parser.add_argument(
"--sequence-lengths",
type=int,
nargs="+",
default=DEFAULT_SEQUENCE_LENGTHS,
help="Sequence lengths to benchmark (tensor shape: seq_len x hidden_size)",
)
parser.add_argument(
"--num-warmup", type=int, default=5, help="Number of warmup iterations"
)
parser.add_argument(
"--num-trials", type=int, default=50, help="Number of benchmark trials"
)
parser.add_argument("--output-json", type=str, help="Output results to JSON file")
args = parser.parse_args()
# Initialize distributed
if not dist.is_initialized():
dist.init_process_group(backend="gloo")
rank = dist.get_rank()
world_size = dist.get_world_size()
# Set device
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
# Get CPU process group
cpu_group = dist.new_group(backend="gloo")
# Disable USE_SYMM_MEM to avoid affecting the max_sizes
# in symm_mem and custom_all_reduce for benchmark
os.environ["VLLM_ALLREDUCE_USE_SYMM_MEM"] = "0"
# Initialize benchmark
benchmark = CommunicatorBenchmark(
rank, world_size, device, cpu_group, args.sequence_lengths
)
# Run benchmarks
all_results = {}
for seq_len in args.sequence_lengths:
if rank == 0:
logger.info(
"Benchmarking sequence length: %s (tensor shape: %s x %s)",
seq_len,
seq_len,
HIDDEN_SIZE,
)
results = benchmark.benchmark_allreduce(
sequence_length=seq_len,
num_warmup=args.num_warmup,
num_trials=args.num_trials,
)
all_results[seq_len] = results
# Synchronize between ranks
dist.barrier()
# Print results (only rank 0)
if rank == 0:
print_results(all_results, args.sequence_lengths, world_size)
# Save to JSON if requested
if args.output_json:
# Add speedup information to results
enhanced_results = {}
for seq_len, comm_results in all_results.items():
enhanced_results[seq_len] = {
"timings": comm_results,
"speedup_info": _calculate_speedup_info(comm_results),
}
output_data = {
"world_size": world_size,
"dtype": str(BENCHMARK_DTYPE),
"hidden_size": HIDDEN_SIZE,
"sequence_lengths": args.sequence_lengths,
"num_warmup": args.num_warmup,
"num_trials": args.num_trials,
"cuda_graph_capture_cycles": CUDA_GRAPH_CAPTURE_CYCLES,
"results": enhanced_results,
}
with open(args.output_json, "w") as f:
json.dump(output_data, f, indent=2)
logger.info("Results saved to %s", args.output_json)
# Cleanup
if cpu_group != dist.group.WORLD:
dist.destroy_process_group(cpu_group)
if __name__ == "__main__":
main()
benchmarks/kernels/benchmark_lora.py
View file @ 38d80967
......@@ -637,7 +637,7 @@ def bench_optype(
# 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
# Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are set up
for kwargs in kwargs_list:
op_type.bench_fn()(**kwargs)
torch.cuda.synchronize()
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment