Skip to content

GitLab

"x/imagegen/vscode:/vscode.git/clone" did not exist on "33ee7168ba1e16c813b52dc2c9417efa1e2e9f20"
Commit cea23494 authored by laibao's avatar laibao
Browse files

更新README.md，修改docker镜像和依赖库版本，删除多个示例文件以简化项目结构。

parent 8cb87a8a
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2920 additions and 194 deletions
README.md
View file @ cea23494
......@@ -34,7 +34,7 @@ Qwen1.5是阿里云开源大型语言模型系列，是Qwen2.0的beta版本。
提供[光源](https://www.sourcefind.cn/#/image/dcu/custom)拉取推理的docker镜像：
```
docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.3.0-ubuntu22.04-dtk24.04.3-py3.10
docker pull image.sourcefind.cn:5000/dcu/admin/base/custom:vllm0.8.5-ubuntu22.04-dtk25.04.1-rc5-das1.6-py3.10-20250724
# <Image ID>用上面拉取docker镜像的ID替换
# <Host Path>主机端路径
# <Container Path>容器映射路径
......@@ -61,12 +61,13 @@ conda create -n qwen1.5_vllm python=3.10
关于本项目DCU显卡所需的特殊深度学习库可从[光合](https://developer.sourcefind.cn/tool/)开发者社区下载安装。
* DTK驱动：dtk24.04.3
* Pytorch: 2.3.0
* triton: 2.1.0
* lmslim: 0.1.2
* DTK驱动：dtk25.04.01
* Pytorch: 2.4.0
* triton: 3.0.0
* lmslim: 0.2.1
* flash_attn: 2.6.1
* vllm: 0.6.2
* flash_mla: 1.0.0
* vllm: 0.8.5
* python: python3.10
`Tips：需先安装相关依赖，最后安装vllm包`
......@@ -107,7 +108,7 @@ export VLLM_RANK7_NUMA=7
### 离线批量推理
```bash
python examples/offline_inference.py
VLLM_USE_FLASH_ATTN_PA=1 pythonexamples/offline_inference/basic/basic.py
```
其中，`prompts`为提示词；`temperature`为控制采样随机性的值，值越小模型生成越确定，值变高模型生成更随机，0表示贪婪采样，默认为1；`max_tokens=16`为生成长度，默认为1；
......@@ -118,18 +119,17 @@ python examples/offline_inference.py
1、指定输入输出
```bash
python benchmarks/benchmark_throughput.py --num-prompts 1 --input-len 32 --output-len 128 --model Qwen/Qwen1.5-7B-Chat -tp 1 --trust-remote-code --enforce-eager --dtype float16
VLLM_USE_FLASH_ATTN_PA=1 python benchmarks/benchmark_throughput.py --num-prompts 1 --input-len 32 --output-len 128 --model /your/model/path -tp 1 --trust-remote-code --enforce-eager --dtype float16
```
其中 `--num-prompts`是batch数，`--input-len`是输入seqlen，`--output-len`是输出token长度，`--model`为模型路径，`-tp`为使用卡数，`dtype="float16"`为推理数据类型，如果模型权重是bfloat16,需要修改为float16推理。若指定 `--output-len 1`即为首字延迟。`-q gptq`为使用gptq量化模型进行推理。
2、使用数据集
下载数据集：
[sharegpt_v3_unfiltered_cleaned_split](https://huggingface.co/datasets/learnanything/sharegpt_v3_unfiltered_cleaned_split)
```bash
python benchmarks/benchmark_throughput.py --num-prompts 1 --model Qwen/Qwen1.5-7B-Chat --dataset ShareGPT_V3_unfiltered_cleaned_split.json -tp 1 --trust-remote-code --enforce-eager --dtype float16
VLLM_USE_FLASH_ATTN_PA=1 python benchmarks/benchmark_throughput.py --num-prompts 1 --model /your/model/path --dataset-name sharegpt --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json -tp 1 --trust-remote-code --enforce-eager --dtype float16
```
其中 `--num-prompts`是batch数，`--model`为模型路径，`--dataset`为使用的数据集，`-tp`为使用卡数，`dtype="float16"`为推理数据类型，如果模型权重是bfloat16,需要修改为float16推理。`-q gptq`为使用gptq量化模型进行推理。
......@@ -139,23 +139,23 @@ python benchmarks/benchmark_throughput.py --num-prompts 1 --model Qwen/Qwen1.5-7
1、启动服务端：
```bash
python -m vllm.entrypoints.openai.api_server --model Qwen/Qwen1.5-7B-Chat --dtype float16 --enforce-eager -tp 1
VLLM_USE_FLASH_ATTN_PA=1 vllm serve --model /your/model/path --enforce-eager --dtype float16 --trust-remote-code --tensor-parallel-size 1
```
2、启动客户端：
```bash
python benchmarks/benchmark_serving.py --model Qwen/Qwen1.5-7B-Chat --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 1 --trust-remote-code
```
python benchmarks/benchmark_serving.py --model /your/model/path --dataset-name sharegpt --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 1 --trust-remote-code
```
参数同使用数据集，离线批量推理性能测试，具体参考[benchmarks/benchmark_serving.py](benchmarks/benchmark_serving.py)
参数同使用数据集，离线批量推理性能测试，具体参考[benchmarks/benchmark_serving.py](/codes/modelzoo/qwen1.5_vllm/-/blob/master/benchmarks/benchmark_serving.py)
### OpenAI兼容服务
启动服务：
```bash
vllm serve Qwen/Qwen1.5-7B-Chat --enforce-eager --dtype float16 --trust-remote-code --port 8000
VLLM_USE_FLASH_ATTN_PA=1 vllm serve /your/model/path --enforce-eager --dtype float16 --trust-remote-code
```
这里serve之后为加载模型路径，`--dtype`为数据类型：float16，默认情况使用tokenizer中的预定义聊天模板，`--chat-template`可以添加新模板覆盖默认模板,`-q gptq`为使用gptq量化模型进行推理,`-q awqq`为使用awq量化模型进行推理。
......@@ -172,14 +172,14 @@ curl http://localhost:8000/v1/models
curl http://localhost:8000/v1/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen1.5-7B",
"model": "/your/model/path",
"prompt": "What is deep learning?",
"max_tokens": 7,
"temperature": 0
}'
```
或者使用[examples/openai_completion_client.py](examples/openai_completion_client.py)
或者使用[examples/online_serving/openai_completion_client.py](examples/online_serving/openai_completion_client.py)
### OpenAI Chat API和vllm结合使用
......@@ -187,7 +187,8 @@ curl http://localhost:8000/v1/completions \
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen1.5-7B-Chat",
"model": "/your/model/path",
"max_tokens": 128,
"messages": [
{"role": "system", "content": "What is deep learning?"},
{"role": "user", "content": "What is deep learning?"}
......@@ -195,7 +196,7 @@ curl http://localhost:8000/v1/chat/completions \
}'
```
或者使用[examples/openai_chatcompletion_client.py](examples/openai_chatcompletion_client.py)
或者使用[examples/online_serving/openai_chat_completion_client.py](examples/online_serving/openai_chat_completion_client.py)
### **gradio和vllm结合使用**
......@@ -230,7 +231,7 @@ ssh -L 8000:计算节点IP:8000 -L 8001:计算节点IP:8001 用户名@登录节
3.启动OpenAI兼容服务
```
vllm serve Qwen/Qwen1.5-7B-Chat --enforce-eager --dtype float16 --trust-remote-code --port 8000 --host "0.0.0.0"
VLLM_USE_FLASH_ATTN_PA=1 vllm serve /your/model/path --enforce-eager --dtype float16 --trust-remote-code --port 8000 --host "0.0.0.0"
```
4.启动gradio服务
......
examples/cpu_offload.py deleted 100644 → 0
View file @ 8cb87a8a
from vllm import LLM, SamplingParams
# Sample prompts.
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
# Create an LLM.
llm = LLM(model="meta-llama/Llama-2-13b-chat-hf", cpu_offload_gb=10)
# Generate texts from the prompts. The output is a list of RequestOutput objects
# that contain the prompt, generated text, and other information.
outputs = llm.generate(prompts, sampling_params)
# Print the outputs.
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
examples/fp8/README.md 0 → 100644
View file @ cea23494
# FP8 KV Cache
This utility extracts the KV cache scaling factors from a quantized HF (Hugging Face) model. The extracted scaling factors are saved to a JSON file, which can later be used by vLLM (variable-length language model) during runtime. This tool is particularly useful when the KV cache data type is FP8 and is intended for use on ROCm (hcu) platforms.
## Prerequisites
- Python 3.x
- PyTorch
- NumPy
- Hugging Face Transformers
- Hugging Face Hub
- AMMO
Before incorporating the FP8 datatype for inference workloads, you must adhere to the following steps:
1. Install all necessary prerequisites and dependencies.
2. Convert HF model into a quantized HF model.
3. Extract KV Cache Scaling Factors from quantized HF model.
4. Load KV Cache Scaling Factors into VLLM.
### 2. Convert HF model into a quantized HF model.
Note: The following steps are adapted from the [TensorRT-LLM repository](https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/quantization/README.md).
`quantize.py` (examples/fp8/quantizer/quantize.py) uses the quantization toolkit (AMMO) to calibrate the PyTorch models and export TensorRT-LLM checkpoints. Each TensorRT-LLM checkpoint contains a config file (in .json format) and one or several rank weight files (in .safetensors format).
The detailed quantization toolkit (AMMO) conversion guide for FP8 can be found at `examples/fp8/quantizer/README.md`.
### 3. Extract KV Cache Scaling Factors from quantized HF model.
`extract_scales.py` (examples/fp8/extract_scales.py) can be utilized to extract the KV cache scaling factors from your quantized HF model, however at the moment, this tool exclusively supports Llama 2 models. It is also important to note the following:
1. **File Structure**: The utility operates under the assumption that all parameters, including KV cache scaling factors, corresponding to a particular Tensor Parallelism (TP) rank are stored in a single file. These files must adhere to a specific naming convention where the TP rank is immediately identified after a specific keyword (e.g., "rank") in the filename.
2. **TP Decomposition**: The utility assumes consistency between the TP decomposition employed by the quantizer tool and that used by vLLM.
3. **AMMO Compatibility**: Currently, the generated KV cache scaling factors for AMMO remain uniform across all TP ranks.
```python
# prerequisites:
# - Quantized HF LLaMa 2 model
python3 examples/fp8/extract_scales.py --help
Usage: extract_scales.py [-h] --quantized_model QUANTIZED_MODEL [--load_format {auto,safetensors,npz,pt}] [--output_dir OUTPUT_DIR] [--output_name OUTPUT_NAME] [--tp_size TP_SIZE]
KV Scale Extraction Example
optional arguments:
--quantized_model: Specify either the local path to, or name of, a quantized HF model. It is expected that the quantization format is FP8_E4M3, for use on ROCm (hcu).
Optional arguments:
--cache_dir: Specify a cache directory to use in the event of a HF model download. (Default: None)
--load_format: Specify the format of the model's tensor files containing the KV cache scaling factors. (Choices: auto, safetensors, npz, pt; Default: auto)
--revision: Specify the model's revision number. (Default: None)
--output_dir: Specify the output directory. By default the KV cache scaling factors will be saved in the model directory. (Default: None)
--output_name: Specify the output filename. (Default: kv_cache_scales.json)
--tp_size: Specify the tensor-parallel (TP) size that the quantized model should correspond to. If specified, during KV cache scaling factor extraction the observed TP size will be checked against this and an error will be raised if there is a mismatch. (Default: None)
```
```python
Example:
python3 examples/fp8/extract_scales.py --quantized_model <QUANTIZED_MODEL_DIR> --tp_size <TENSOR_PARALLEL_SIZE> --output_dir <PATH_TO_OUTPUT_DIR>
```
### 4. Load KV Cache Scaling Factors into VLLM.
This script evaluates the inference throughput of language models using various backends such as vLLM. It measures the time taken to process a given number of prompts and generate sequences for each prompt. The recently generated KV cache scaling factors are now integrated into the benchmarking process and allow for KV cache scaling factors to be utilized for FP8.
```python
# prerequisites:
# - LLaMa 2 kv_cache_scales.json file
python3 benchmarks/benchmark_throughput.py --help
usage: benchmark_throughput.py [-h] [--backend {vllm,hf,mii}] [--dataset DATASET] [--input-len INPUT_LEN] [--output-len OUTPUT_LEN] [--model MODEL]
[--tokenizer TOKENIZER] [--quantization {awq,gptq,None}] [--tensor-parallel-size TENSOR_PARALLEL_SIZE] [--n N]
[--use-beam-search] [--num-prompts NUM_PROMPTS] [--seed SEED] [--hf-max-batch-size HF_MAX_BATCH_SIZE] [--trust-remote-code]
[--max-model-len MAX_MODEL_LEN] [--dtype {auto,half,float16,bfloat16,float,float32}] [--enforce-eager] [--kv-cache-dtype {auto,fp8}]
[--quantization-param-path KV_CACHE_quantization_param_path]
Benchmark Throughput Example
optional arguments:
-h, --help show this help message and exit
--backend {vllm,hf,mii}
--dataset DATASET Path to the dataset.
--input-len INPUT_LEN Input prompt length for each request
--output-len OUTPUT_LEN Output length for each request. Overrides the output length from the dataset.
--model MODEL
--tokenizer TOKENIZER
--quantization {awq,gptq,None}, -q {awq,gptq,None}
--tensor-parallel-size TENSOR_PARALLEL_SIZE, -tp TENSOR_PARALLEL_SIZE
--n N Number of generated sequences per prompt.
--use-beam-search
--num-prompts NUM_PROMPTS Number of prompts to process.
--seed SEED
--hf-max-batch-size HF_MAX_BATCH_SIZE Maximum batch size for HF backend.
--trust-remote-code trust remote code from huggingface
--max-model-len MAX_MODEL_LEN Maximum length of a sequence (including prompt and output). If None, will be derived from the model.
--dtype {auto,half,float16,bfloat16,float,float32} data type for model weights and activations. The "auto" option will use FP16 precision for FP32 and FP16 models, and BF16 precision for BF16 models.
--enforce-eager enforce eager execution
--kv-cache-dtype {auto,fp8} Data type for kv cache storage. If "auto", will use model data type. FP8_E5M2 (without scaling) is only supported on cuda version greater than 11.8. On ROCm (hcu), FP8_E4M3 is instead supported ```for common inference criteria.
--quantization-param-path QUANT_PARAM_JSON Path to the JSON file containing the KV cache scaling factors. This should generally be supplied, when KV cache dtype is FP8. Otherwise, KV cache scaling factors default to 1.0, which may cause accuracy issues. FP8_E5M2 (without scaling) is only supported on cuda version greater than 11.8. On ROCm (hcu), FP8_E4M3 is instead supported for common inference criteria.
```
```
Example:
python3 benchmarks/benchmark_throughput.py --input-len <INPUT_LEN> --output-len <OUTPUT_LEN> -tp <TENSOR_PARALLEL_SIZE> --kv-cache-dtype fp8 --quantization-param-path <path/to/kv_cache_scales.json> --model <path-to-llama2>
```python
examples/fp8/extract_scales.py 0 → 100644
View file @ cea23494
import argparse
import glob
import json
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import numpy as np
import torch
from safetensors.torch import safe_open
from vllm.model_executor.layers.quantization.schema import QuantParamSchema
# Adapted from vllm/model_executor/model_loader/weight_utils.py
# The main differences are that we add the NPZ format and simplify
# its functionality drastically for our purposes (e.g. we assume that
# the quantized model exists locally and there is no need to download it)
def _prepare_hf_weights(
quantized_model_dir: str,
load_format: str = "auto",
fall_back_to_pt: bool = True,
) -> Tuple[List[str], bool]:
if not os.path.isdir(quantized_model_dir):
raise FileNotFoundError(
f"The quantized model directory `{quantized_model_dir}` "
"does not exist.")
use_safetensors = False
# Some quantized models use .pt files for storing the weights.
if load_format == "auto":
allow_patterns = ["*.safetensors", "*.bin"]
elif load_format == "safetensors":
use_safetensors = True
allow_patterns = ["*.safetensors"]
elif load_format == "pt":
allow_patterns = ["*.pt"]
elif load_format == "npz":
allow_patterns = ["*.npz"]
else:
raise ValueError(f"Unknown load_format: {load_format}")
if fall_back_to_pt:
allow_patterns += ["*.pt"]
hf_weights_files: List[str] = []
for pattern in allow_patterns:
hf_weights_files += glob.glob(
os.path.join(quantized_model_dir, pattern))
if len(hf_weights_files) > 0:
if pattern == "*.safetensors":
use_safetensors = True
break
if not use_safetensors:
# Exclude files that are not needed for inference.
# https://github.com/huggingface/transformers/blob/v4.34.0/src/transformers/trainer.py#L227-L233
blacklist = [
"training_args.bin",
"optimizer.bin",
"optimizer.pt",
"scheduler.pt",
"scaler.pt",
]
hf_weights_files = [
f for f in hf_weights_files
if not any(f.endswith(x) for x in blacklist)
]
if len(hf_weights_files) == 0:
raise RuntimeError(
f"Cannot find any model weights with `{quantized_model_dir}`")
return hf_weights_files, use_safetensors
# Adapted from vllm/model_executor/model_loader/weight_utils.py
def _hf_tensorfile_iterator(filename: str, load_format: str,
use_safetensors: bool):
if load_format == "npz":
assert not use_safetensors
with np.load(filename) as data:
for name in data.files:
param = torch.from_numpy(data[name])
yield name, param
elif use_safetensors:
with safe_open(filename, framework="pt") as f:
for name in f.keys(): # NOQA: SIM118
param = f.get_tensor(name)
yield name, param
else:
state = torch.load(filename, map_location="cpu")
for name, param in state.items():
yield name, param
del state
torch.cuda.empty_cache()
def _kv_scales_extractor(
hf_tensor_files: List[str],
use_safetensors: bool,
rank_keyword: str = "rank",
expected_tp_size: Optional[int] = None) -> Dict[int, Dict[int, float]]:
"""
Given a list of files containing tensor data, attempt to extract KV cache
scales from these files. Intended as a helper function taking in the output
from _prepare_hf_weights.
Args:
rank_keyword Matches the number immediately after this keyword in the
tensor filename to determine the TP rank corresponding
to said tensor file
expected_tp_size If specified, the TP size of the tensor files is checked
against this and an error is raised if they don't match.
Returns a dictionary mapping TP ranks to their relevant KV cache scales.
The per-rank scales are themselves represented as a dictionary of layer
indices to the respective per-layer scale.
"""
for char in rank_keyword:
assert not char.isdecimal(
), f"Rank keyword {rank_keyword} contains a numeric character!"
rank_scales_map: Dict[int, Dict[int, float]] = {}
for tensor_file in hf_tensor_files:
try:
rank_idx = tensor_file.find(rank_keyword)
if rank_idx != -1:
start_idx = rank_idx + len(rank_keyword)
stop_idx = start_idx
while stop_idx < len(
tensor_file) and tensor_file[stop_idx].isdecimal():
stop_idx += 1
if stop_idx == start_idx:
raise RuntimeError("Did not find rank # in filename.")
rank = int(tensor_file[start_idx:stop_idx])
elif len(hf_tensor_files) == 1:
# Since there is only one tensor file, we can assume
# that it's intended for TP rank 0
rank = 0
else:
raise RuntimeError(
f"Filename does not contain '{rank_keyword}'.")
except RuntimeError:
print("Unable to determine TP rank "
f"corresponding to file '{tensor_file}'")
raise
if rank not in rank_scales_map:
layer_scales_map: Dict[int, float] = {}
rank_scales_map[rank] = layer_scales_map
else:
raise RuntimeError(
f"Tensor file '{tensor_file}' shares TP rank {rank} "
"with another tensor file.")
module_delimiter = ":" if args.load_format == "npz" else "."
for name, param in _hf_tensorfile_iterator(tensor_file,
args.load_format,
use_safetensors):
if "kv_cache_scaling_factor" in name:
nums = [
int(s) for s in name.split(module_delimiter)
if s.isdecimal()
]
assert len(
nums) == 1, f"Could not determine layer idx for {name}"
layer_idx = nums[0]
assert layer_idx not in layer_scales_map, f"Duplicate scaling"\
f" factor corresponding to layer {layer_idx}"
try:
layer_scales_map[layer_idx] = param.item()
except RuntimeError:
print(
"This utility supports only per-tensor scalar scales "
f"for now. The tensor\n {name} = {param} \nis an "
"invalid scale factor.")
raise
if all(
len(layer_scales_map) == 0
for layer_scales_map in rank_scales_map.values()):
# Note: this is true even if the rank_scales_map is empty
print("WARNING: No KV cache scale factors found. No output saved.")
return None
empirical_tp_world_size = max(rank_scales_map.keys()) + 1
if expected_tp_size is not None:
assert expected_tp_size == empirical_tp_world_size, \
f"User expected TP world size = {expected_tp_size} " \
"from model but tool is expecting TP world size = " \
f"{empirical_tp_world_size} from model instead."
for i in range(empirical_tp_world_size):
assert i in rank_scales_map, "Expected TP world size = "\
f"{empirical_tp_world_size} but did not find KV " \
f"cache scaling factors for TP rank {i}"
print(f"Found TP world size = {empirical_tp_world_size} "
"when extracting KV cache scales!")
return rank_scales_map
def _metadata_extractor(quantized_model_dir: str,
metadata_extract_fns: \
Dict[str, Callable[[Dict[str, Any]], Any]]) \
-> Dict[str, Any]:
"""
Given a directory containing quantized model files, this function
aims to extract metadata from the JSON files within this directory.
Each JSON file is expected to represent a dictionary in JSON
format (referred to as a "JSON-dictionary"). Metadata extraction is
defined by a dictionary called metadata_extract_fns, where each
metadata field name is mapped to an extraction function.
These extraction functions are designed to take a JSON-dictionary
as their only argument and return the corresponding metadata.
While extraction functions are permitted to raise exceptions, they
should only raise a KeyError or ValueError if the metadata field
cannot be extracted from the current JSON-dictionary, yet there's
a possibility of finding it in another JSON-dictionary.
The function returns a dictionary that maps metadata fields to
their extracted data. The keys of this dictionary correspond exactly
to those in metadata_extract_fns. If any fields fail to be extracted,
their corresponding values are set to None, and a warning is printed.
"""
if not os.path.isdir(quantized_model_dir):
raise FileNotFoundError(
f"The quantized model directory `{quantized_model_dir}` "
"does not exist.")
metadata_files = glob.glob(os.path.join(quantized_model_dir, "*.json"))
result: Dict[str, Any] = {}
for file in metadata_files:
with open(file) as f:
try:
metadata = json.load(f)
except json.JSONDecodeError:
print(f"Could not parse `{file}` as a valid metadata file,"
" skipping it.")
continue
if not isinstance(metadata, dict):
print(f"The file `{file}` does not correspond to a "
"JSON-serialized dictionary, skipping it.")
continue
for metadata_name, extract_fn in metadata_extract_fns.items():
try:
metadata_info = extract_fn(metadata)
if metadata_name not in result:
result[metadata_name] = metadata_info
elif metadata_info != result[metadata_name]:
raise RuntimeError(
"Metadata mismatch! Originally found "
f"{metadata_name} = {result[metadata_name]} but "
f"now found {metadata_name} = {metadata_info} in "
f"`{file}`")
except KeyError:
# It is possible that a given file does not contain some
# of our selected metadata as it could be located in some
# other metadata file.
# 'EFINAE': extract_fn failure is not an error.
pass
except ValueError:
# See above.
pass
# Warn if we cannot find any of the requested metadata
for metadata_name in metadata_extract_fns:
if metadata_name not in result:
print("WARNING: Unable to find requested metadata field "
f"`{metadata_name}`, setting it to None.")
result[metadata_name] = None
return result
def main(args):
metadata_extract_fns = {
"model_type": lambda json_dict: json_dict["layers"][0]["decoder_type"],
"tp_size": lambda json_dict: int(json_dict["tensor_parallel"]),
"model_dtype": lambda json_dict: json_dict["dtype"]
}
recovered_metadata = _metadata_extractor(args.quantized_model,
metadata_extract_fns)
if args.tp_size is not None:
metadata_tp_size = recovered_metadata["tp_size"]
if metadata_tp_size is not None:
assert args.tp_size == metadata_tp_size, \
f"User expected TP world size = {args.tp_size} " \
f"but found TP world size = {metadata_tp_size} from metadata!"
expected_tp_size = args.tp_size or recovered_metadata["tp_size"]
rank_keyword = "rank"
hf_tensor_files, use_safetensors = _prepare_hf_weights(
args.quantized_model, args.load_format)
rank_scales_map = _kv_scales_extractor(hf_tensor_files, use_safetensors,
rank_keyword, expected_tp_size)
# Postprocess: formatting to the current schema. Consider pulling it
# out into a dedicated function should it ever become more complicated.
rank_scales_map = {
rank: {k: scale[k]
for k in sorted(scale.keys())}
for rank, scale in rank_scales_map.items()
}
# TODO: Expand this with activation and weights scaling factors when
# they are used in the future
schema = QuantParamSchema(
model_type=recovered_metadata["model_type"],
kv_cache={
"dtype": ("float8_e4m3fn" if len(rank_scales_map) > 0 else
recovered_metadata["model_dtype"]),
"scaling_factor":
rank_scales_map
},
)
if args.output_dir is None:
output_file = os.path.join(args.quantized_model, args.output_name)
else:
if not os.path.isdir(args.output_dir):
os.makedirs(args.output_dir, exist_ok=True)
output_file = os.path.join(args.output_dir, args.output_name)
with open(output_file, 'w') as f:
f.write(schema.model_dump_json(indent=4))
print(f"Completed! KV cache scaling factors saved to {output_file}")
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="This simple utility extracts the "
"KV cache scaling factors from a quantized HF model "
"and saves them to a JSON file compatible with later "
"use by vLLM (pass this file to the appropriate "
"runtime typically using the argument "
"--quantization-param-path <filename>). This is only used "
"if the KV cache dtype is FP8 and on ROCm (hcu).")
parser.add_argument(
"--quantized-model",
help="Specify the directory containing a single quantized HF model. "
"It is expected that the quantization format is FP8_E4M3, for use "
"on ROCm (hcu).",
required=True)
parser.add_argument(
"--load_format",
help="Optionally specify the format of the model's tensor files "
"containing the KV cache scaling factors.",
choices=["auto", "safetensors", "npz", "pt"],
default="auto")
parser.add_argument(
"--output-dir",
help="Optionally specify the output directory. By default the "
"KV cache scaling factors will be saved in the model directory, "
"however you can override this behavior here.",
default=None)
parser.add_argument(
"--output-name",
help="Optionally specify the output filename.",
# TODO: Change this once additional scaling factors are enabled
default="kv_cache_scales.json")
parser.add_argument(
"--tp-size",
help="Optionally specify the tensor-parallel (TP) size that the "
"quantized model should correspond to. If specified, during KV "
"cache scaling factor extraction the observed TP size will be "
"checked against this and an error will be raised if there is "
"a mismatch. If not specified, the quantized model's expected "
"TP size is instead inferred from the largest TP rank observed. "
"The expected TP size is cross-checked against the TP ranks "
"observed in the quantized model and an error is raised if any "
"discrepancies are found.",
default=None,
type=int)
args = parser.parse_args()
main(args)
examples/gguf_inference.py deleted 100644 → 0
View file @ 8cb87a8a
from huggingface_hub import hf_hub_download
from vllm import LLM, SamplingParams
def run_gguf_inference(model_path):
PROMPT_TEMPLATE = "<|system|>\n{system_message}</s>\n<|user|>\n{prompt}</s>\n<|assistant|>\n" # noqa: E501
system_message = "You are a friendly chatbot who always responds in the style of a pirate." # noqa: E501
# Sample prompts.
prompts = [
"How many helicopters can a human eat in one sitting?",
"What's the future of AI?",
]
prompts = [
PROMPT_TEMPLATE.format(system_message=system_message, prompt=prompt)
for prompt in prompts
]
# Create a sampling params object.
sampling_params = SamplingParams(temperature=0, max_tokens=128)
# Create an LLM.
llm = LLM(model=model_path,
tokenizer="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
gpu_memory_utilization=0.95)
outputs = llm.generate(prompts, sampling_params)
# Print the outputs.
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
if __name__ == "__main__":
repo_id = "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF"
filename = "tinyllama-1.1b-chat-v1.0.Q4_0.gguf"
model = hf_hub_download(repo_id, filename=filename)
run_gguf_inference(model)
examples/gradio_openai_chatbot_webserver.py deleted 100644 → 0
View file @ 8cb87a8a
import argparse
import gradio as gr
from openai import OpenAI
# Argument parser setup
parser = argparse.ArgumentParser(
description='Chatbot Interface with Customizable Parameters')
parser.add_argument('--model-url',
type=str,
default='http://localhost:8000/v1',
help='Model URL')
parser.add_argument('-m',
'--model',
type=str,
required=True,
help='Model name for the chatbot')
parser.add_argument('--temp',
type=float,
default=0.8,
help='Temperature for text generation')
parser.add_argument('--stop-token-ids',
type=str,
default='',
help='Comma-separated stop token IDs')
parser.add_argument("--host", type=str, default=None)
parser.add_argument("--port", type=int, default=8001)
# Parse the arguments
args = parser.parse_args()
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = args.model_url
# Create an OpenAI client to interact with the API server
client = OpenAI(
api_key=openai_api_key,
base_url=openai_api_base,
)
def predict(message, history):
# Convert chat history to OpenAI format
history_openai_format = [{
"role": "system",
"content": "You are a great ai assistant."
}]
for human, assistant in history:
history_openai_format.append({"role": "user", "content": human})
history_openai_format.append({
"role": "assistant",
"content": assistant
})
history_openai_format.append({"role": "user", "content": message})
# Create a chat completion request and send it to the API server
stream = client.chat.completions.create(
model=args.model, # Model name to use
messages=history_openai_format, # Chat history
temperature=args.temp, # Temperature for text generation
stream=True, # Stream response
extra_body={
'repetition_penalty':
1,
'stop_token_ids': [
int(id.strip()) for id in args.stop_token_ids.split(',')
if id.strip()
] if args.stop_token_ids else []
})
# Read and return generated text from response stream
partial_message = ""
for chunk in stream:
partial_message += (chunk.choices[0].delta.content or "")
yield partial_message
# Create and launch a chat interface with Gradio
gr.ChatInterface(predict).queue().launch(server_name=args.host,
server_port=args.port,
share=True)
examples/medusa/medusa_benchmark_throughput.py
View file @ cea23494
......@@ -42,9 +42,6 @@ def sample_requests(
# Only keep the first two turns of each conversation.
dataset = [data["prompt"] for data in dataset]
# Shuffle the dataset.
random.shuffle(dataset)
# Filter out sequences that are too long or too short
filtered_dataset: List[Tuple[str, int, int]] = []
for i in range(len(dataset)):
......@@ -95,7 +92,6 @@ def run_vllm(
spec_decoding_acceptance_method: str = None,
enable_lora: bool = False,
max_lora_rank: int = 32,
merge_lora: bool = False,
lora_extra_vocab_size: int = 0,
lora_target_modules: List[str] = None,
num_speculative_heads: int = 5,
......@@ -134,7 +130,6 @@ def run_vllm(
spec_decoding_acceptance_method=spec_decoding_acceptance_method,
enable_lora=enable_lora,
max_lora_rank=max_lora_rank,
merge_lora=merge_lora,
lora_extra_vocab_size=lora_extra_vocab_size,
lora_target_modules=lora_target_modules,
num_speculative_heads=num_speculative_heads,
......@@ -229,7 +224,6 @@ async def run_vllm_async(
spec_decoding_acceptance_method: str = None,
enable_lora: bool = False,
max_lora_rank: int = 32,
merge_lora: bool = False,
lora_extra_vocab_size: int = 0,
lora_target_modules: List[str] = None,
num_speculative_heads: int = 5,
......@@ -270,7 +264,6 @@ async def run_vllm_async(
spec_decoding_acceptance_method=spec_decoding_acceptance_method,
enable_lora=enable_lora,
max_lora_rank=max_lora_rank,
merge_lora=merge_lora,
lora_extra_vocab_size=lora_extra_vocab_size,
lora_target_modules=lora_target_modules,
num_speculative_heads=num_speculative_heads,
......@@ -345,7 +338,7 @@ def main(args: argparse.Namespace):
args.disable_async_output_proc, False, args.max_num_seqs,
args.speculative_model, args.speculative_draft_tensor_parallel_size,
args.speculative_disable_by_batch_size, args.spec_decoding_acceptance_method,
args.enable_lora, args.max_lora_rank, args.merge_lora, args.lora_extra_vocab_size,
args.enable_lora, args.max_lora_rank, args.lora_extra_vocab_size,
args.lora_target_modules, args.num_speculative_heads,
args.num_speculative_tokens
]
......@@ -363,7 +356,7 @@ def main(args: argparse.Namespace):
args.disable_async_output_proc, args.max_num_seqs,
args.speculative_model, args.speculative_draft_tensor_parallel_size,
args.speculative_disable_by_batch_size, args.spec_decoding_acceptance_method,
args.enable_lora, args.max_lora_rank, args.merge_lora, args.lora_extra_vocab_size,
args.enable_lora, args.max_lora_rank, args.lora_extra_vocab_size,
args.lora_target_modules, args.num_speculative_heads,
args.num_speculative_tokens
]
......
examples/mla/test_triton_decode_attention.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
import os
import json
import pytest
import torch
import triton
from triton_decode_attention import decode_attentionv1_fwd, decode_attentionv2_fwd
def cdiv(a, b):
return (a + b - 1) // b
@pytest.mark.parametrize("B", [1])
# @pytest.mark.parametrize("L", [100])
@pytest.mark.parametrize("L", [1,100,400,700,1000,1300,1600,1900,2200,2500,2800,3100,3400,3700,4000,4300,4600,4900,5000,5500,6000,6500,7000,7500,8000,8500,9000,9500,10000,10500,11000,11500,12000,12500,13000,13500,14000,14500,15000,15500,16000,16500,17000,17500,18000,18500,19000,19500,20000,20500,21000,21500,22000,22500,23000,23500,24000,24500,25000,25500,26000,26500,27000,27500,28000,28500,29000,29500,30000,30500,31000,31500,32000,32500])
@pytest.mark.parametrize("H_Q", [4, 8, 16])
@pytest.mark.parametrize("H_KV", [1])
@pytest.mark.parametrize("D_QK", [576])
@pytest.mark.parametrize("D_V", [512])
@pytest.mark.parametrize("CACHE_SIZE", [16384])
@pytest.mark.parametrize("PAGE_SIZE", [16])
def test_decode_attention(B, L, H_Q, H_KV, D_QK, D_V, CACHE_SIZE, PAGE_SIZE):
assert CACHE_SIZE % PAGE_SIZE == 0
dtype = torch.bfloat16
seq_len = L # This represents the number of tokens already in the sequence
sm_scale = 1.0 / (D_QK**0.5)
num_kv_splits = 4
num_pages_per_batch = cdiv(seq_len, PAGE_SIZE) #这里为向上取整,65,(1027+16-1)//16
req_to_page = torch.randint(0,
CACHE_SIZE // PAGE_SIZE,
(B, num_pages_per_batch, 1), #shape为（B, num_pages_per_batch, 1）的tensor,大小取值为0 至cache_size//page_size
device="cuda")
req_to_token = req_to_page * PAGE_SIZE
req_to_token = req_to_token.expand(B, num_pages_per_batch, PAGE_SIZE)
req_to_token = req_to_token + torch.arange(PAGE_SIZE, device="cuda").view(
1, 1, -1)
req_to_token = req_to_token.view(B, -1)
req_to_token = req_to_token[:, :seq_len].contiguous()
# q represents the new token being generated, one per batch
q = torch.randn(B, H_Q, D_QK, dtype=dtype, device="cuda")
# k_buffer and v_buffer represent all previous tokens
# Page size is 1.
k_buffer = torch.randn(CACHE_SIZE, H_KV, D_QK, dtype=dtype, device="cuda")
v_buffer = torch.randn(CACHE_SIZE, H_KV, D_V, dtype=dtype, device="cuda")
# o will have the same shape as q
o = torch.zeros(B, H_Q, D_V, dtype=dtype, device="cuda")
b_seq_len = torch.full((B, ), seq_len, device="cuda")
attn_logits = torch.empty(
(B, H_Q, num_kv_splits, D_V + 1),
dtype=torch.float32,
device="cuda",
)
b_req_idx = torch.arange(B, device="cuda").to(torch.int32)
# Call the original implementation.
decode_attentionv2_fwd(
q,
k_buffer,
v_buffer,
o,
req_to_token,
b_seq_len,
attn_logits,
num_kv_splits,
sm_scale,
)
# Page size can be larger than 1.
k_buffer = k_buffer.view(CACHE_SIZE // PAGE_SIZE, PAGE_SIZE, H_KV, D_QK)
v_buffer = v_buffer.view(CACHE_SIZE // PAGE_SIZE, PAGE_SIZE, H_KV, D_V)
o1 = torch.zeros_like(o)
configs = {
"v2_tc": {"stage1": {}, "stage2": {}},
"v1_2stages_tc": {"stage1": {}, "stage2": {}},
}
ms = {
"v1_2stages_tc": 10000.0,
"v2_tc": 10000.0,
}
final_best_config = {
"kernel_kind": "",
"best_config": {},
"best_us": 0.0,
}
v2_tc_stage1_best_config, v2_tc_stage2_best_config = decode_attentionv2_fwd(
q,
k_buffer,
v_buffer,
o1,
req_to_page,
b_seq_len,
attn_logits,
num_kv_splits,
sm_scale,
PAGE_SIZE,
)
assert torch.allclose(o, o1, atol=1e-2, rtol=1e-2)
quantiles = [0.5, 0.2, 0.8]
v2_tc_ms, v2_tc_min_ms, v2_tc_max_ms = triton.testing.do_bench(lambda:
decode_attentionv2_fwd(
q,
k_buffer,
v_buffer,
o1,
req_to_page,
b_seq_len,
attn_logits,
num_kv_splits,
sm_scale,
PAGE_SIZE,
), quantiles=quantiles)
for key, value in v2_tc_stage1_best_config.kwargs.items():
configs["v2_tc"]["stage1"][key] = value
configs["v2_tc"]["stage1"]["num_stages"] = v2_tc_stage1_best_config.num_stages
configs["v2_tc"]["stage1"]["num_warps"] = v2_tc_stage1_best_config.num_warps
for key, value in v2_tc_stage2_best_config.kwargs.items():
configs["v2_tc"]["stage2"][key] = value
configs["v2_tc"]["stage2"]["num_stages"] = v2_tc_stage2_best_config.num_stages
configs["v2_tc"]["stage2"]["num_warps"] = v2_tc_stage2_best_config.num_warps
ms["v2_tc"] = v2_tc_ms
print(f"v2_tc best configs is {configs['v2_tc']}")
print("print mla decode attention v2 kernel [B, L, H_Q, H_KV, D_QK, D_V, CACHE_SIZE, PAGE_SIZE] min cost :",[B, L, H_Q, H_KV, D_QK, D_V, CACHE_SIZE, PAGE_SIZE], v2_tc_ms)
o2 = torch.zeros_like(o)
v1_tc_stage1_best_config, v1_tc_stage2_best_config = decode_attentionv1_fwd(
q,
k_buffer,
v_buffer,
o2,
req_to_page,
b_seq_len,
attn_logits,
num_kv_splits,
sm_scale,
PAGE_SIZE,
)
assert torch.allclose(o, o2, atol=1e-2, rtol=1e-2)
v1_tc_ms, v1_tc_min_ms, v1_tc_max_ms = triton.testing.do_bench(lambda:
decode_attentionv1_fwd(
q,
k_buffer,
v_buffer,
o1,
req_to_page,
b_seq_len,
attn_logits,
num_kv_splits,
sm_scale,
PAGE_SIZE,
), quantiles=quantiles)
for key, value in v1_tc_stage1_best_config.kwargs.items():
configs["v1_2stages_tc"]["stage1"][key] = value
configs["v1_2stages_tc"]["stage1"]["num_stages"] = v1_tc_stage1_best_config.num_stages
configs["v1_2stages_tc"]["stage1"]["num_warps"] = v1_tc_stage1_best_config.num_warps
configs["v1_2stages_tc"]["stage1"]["num_ldmatrixes"] = v1_tc_stage1_best_config.num_ldmatrixes
for key, value in v1_tc_stage2_best_config.kwargs.items():
configs["v1_2stages_tc"]["stage2"][key] = value
configs["v1_2stages_tc"]["stage2"]["num_stages"] = v1_tc_stage2_best_config.num_stages
configs["v1_2stages_tc"]["stage2"]["num_warps"] = v1_tc_stage2_best_config.num_warps
configs["v1_2stages_tc"]["stage2"]["num_ldmatrixes"] = v1_tc_stage1_best_config.num_ldmatrixes
ms["v1_2stages_tc"] = v1_tc_ms
min_key, min_ms = min(ms.items(), key=lambda x: x[1])
final_best_config["kernel_kind"] = min_key
final_best_config["best_config"] = configs[min_key]
final_best_config["best_us"] = min_ms * 1000
print(f"v1_2stages_tc best configs is {configs['v1_2stages_tc']}")
print("print mla decode attention v1 kernel [B, L, H_Q, H_KV, D_QK, D_V, CACHE_SIZE, PAGE_SIZE] min cost :",[B, L, H_Q, H_KV, D_QK, D_V, CACHE_SIZE, PAGE_SIZE], v1_tc_ms)
print(f"Tuned_decode_attention choose {min_key} kernel, min cost {min_ms} ms, best config of {min_key} kernel is {configs[min_key]}")
assert torch.allclose(o, o2, atol=1e-2, rtol=1e-2)
#**************save config**************#
batch = b_req_idx.shape[0]
mean_seq_len = int((b_seq_len.sum() / max(1, batch)).item())
device_name = torch.cuda.get_device_name().replace(" ", "_")
if "K100_AI" in device_name:
# return f"QH={QH}_KVH={KVH}_QKD={QKD}_VD={VD}_{cache_dtype}_K100AI.json"
file_name = f"QH={H_Q}_KVH={H_KV}_QKD={D_QK}_VD={D_V}_fp16_K100AI.json"
elif "BW" in device_name:
# return f"QH={QH}_KVH={KVH}_QKD={QKD}_VD={VD}_{cache_dtype}_BW.json"
file_name = f"QH={H_Q}_KVH={H_KV}_QKD={D_QK}_VD={D_V}_fp16_BW.json"
else:
raise ValueError(f"Unsurpport device name: {device_name}")
if os.path.exists(file_name):
with open(file_name, 'r') as file:
config_info = json.load(file)
else:
config_info = {}
# 如果 config_info 中没有当前的 batch，初始化它为一个空字典
# if f"{batch}" not in config_info:
# config_info[f"{batch}"] = {}
# 把新的 mean_seq_len 配置加入到当前 batch 中
# config_info[f"{batch}"][f"{mean_seq_len}"] = final_best_config
config_info[f"{mean_seq_len}"] = final_best_config
# 保存最佳配置
with open(file_name, 'w') as file:
json.dump(config_info, file, indent=1)
#**************save config**************#
examples/mla/triton_decode_attention.py 0 → 100644
View file @ cea23494
This diff is collapsed.
examples/offline_inference.py deleted 100644 → 0
View file @ 8cb87a8a
from vllm import LLM, SamplingParams
if __name__ == '__main__':
# Sample prompts.
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(temperature=0.8, top_p=0.95, max_tokens=16)
# Create an LLM.
llm = LLM(model="facebook/opt-125m",tensor_parallel_size=1, distributed_executor_backend="ray", dtype="float16",trust_remote_code=True, enforce_eager=True)
# Generate texts from the prompts. The output is a list of RequestOutput objects
# that contain the prompt, generated text, and other information.
outputs = llm.generate(prompts, sampling_params)
# Print the outputs.
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
examples/offline_inference/audio_language.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
This example shows how to use vLLM for running offline inference
with the correct prompt format on audio language models.
For most models, the prompt format should follow corresponding examples
on HuggingFace model repository.
"""
import os
from dataclasses import asdict
from typing import NamedTuple, Optional
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
from vllm import LLM, EngineArgs, SamplingParams
from vllm.assets.audio import AudioAsset
from vllm.lora.request import LoRARequest
from vllm.utils import FlexibleArgumentParser
audio_assets = [AudioAsset("mary_had_lamb"), AudioAsset("winning_call")]
question_per_audio_count = {
0: "What is 1+1?",
1: "What is recited in the audio?",
2: "What sport and what nursery rhyme are referenced?",
}
class ModelRequestData(NamedTuple):
engine_args: EngineArgs
prompt: str
stop_token_ids: Optional[list[int]] = None
lora_requests: Optional[list[LoRARequest]] = None
# NOTE: The default `max_num_seqs` and `max_model_len` may result in OOM on
# lower-end GPUs.
# Unless specified, these settings have been tested to work on a single L4.
# Granite Speech
def run_granite_speech(question: str, audio_count: int) -> ModelRequestData:
# NOTE - the setting in this example are somehat different than what is
# optimal for granite speech, and it is generally recommended to use beam
# search. Check the model README for suggested settings.
# https://huggingface.co/ibm-granite/granite-speech-3.3-8b
model_name = "ibm-granite/granite-speech-3.3-8b"
engine_args = EngineArgs(
model=model_name,
trust_remote_code=True,
max_model_len=2048,
max_num_seqs=2,
enable_lora=True,
max_lora_rank=64,
limit_mm_per_prompt={"audio": audio_count},
)
# The model has an audio-specific lora directly in its model dir;
# it should be enabled whenever you pass audio inputs to the model.
speech_lora_path = model_name
audio_placeholder = "<|audio|>" * audio_count
prompts = f"<|start_of_role|>system<|end_of_role|>Knowledge Cutoff Date: April 2024.\nToday's Date: December 19, 2024.\nYou are Granite, developed by IBM. You are a helpful AI assistant<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>{audio_placeholder}{question}<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>" # noqa: E501
return ModelRequestData(
engine_args=engine_args,
prompt=prompts,
lora_requests=[LoRARequest("speech", 1, speech_lora_path)],
)
# MiniCPM-O
def run_minicpmo(question: str, audio_count: int) -> ModelRequestData:
model_name = "openbmb/MiniCPM-o-2_6"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
engine_args = EngineArgs(
model=model_name,
trust_remote_code=True,
max_model_len=4096,
max_num_seqs=2,
limit_mm_per_prompt={"audio": audio_count},
)
stop_tokens = ["<|im_end|>", "<|endoftext|>"]
stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
audio_placeholder = "(<audio>./</audio>)" * audio_count
audio_chat_template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n<|spk_bos|><|spk|><|spk_eos|><|tts_bos|>' }}{% endif %}" # noqa: E501
messages = [{"role": "user", "content": f"{audio_placeholder}\n{question}"}]
prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
chat_template=audio_chat_template,
)
return ModelRequestData(
engine_args=engine_args,
prompt=prompt,
stop_token_ids=stop_token_ids,
)
# Phi-4-multimodal-instruct
def run_phi4mm(question: str, audio_count: int) -> ModelRequestData:
"""
Phi-4-multimodal-instruct supports both image and audio inputs. Here, we
show how to process audio inputs.
"""
model_path = snapshot_download("microsoft/Phi-4-multimodal-instruct")
# Since the vision-lora and speech-lora co-exist with the base model,
# we have to manually specify the path of the lora weights.
speech_lora_path = os.path.join(model_path, "speech-lora")
placeholders = "".join([f"<|audio_{i + 1}|>" for i in range(audio_count)])
prompts = f"<|user|>{placeholders}{question}<|end|><|assistant|>"
engine_args = EngineArgs(
model=model_path,
trust_remote_code=True,
max_model_len=12800,
max_num_seqs=2,
enable_lora=True,
max_lora_rank=320,
limit_mm_per_prompt={"audio": audio_count},
)
return ModelRequestData(
engine_args=engine_args,
prompt=prompts,
lora_requests=[LoRARequest("speech", 1, speech_lora_path)],
)
# Qwen2-Audio
def run_qwen2_audio(question: str, audio_count: int) -> ModelRequestData:
model_name = "Qwen/Qwen2-Audio-7B-Instruct"
engine_args = EngineArgs(
model=model_name,
max_model_len=4096,
max_num_seqs=5,
limit_mm_per_prompt={"audio": audio_count},
)
audio_in_prompt = "".join(
[
f"Audio {idx + 1}: <|audio_bos|><|AUDIO|><|audio_eos|>\n"
for idx in range(audio_count)
]
)
prompt = (
"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
"<|im_start|>user\n"
f"{audio_in_prompt}{question}<|im_end|>\n"
"<|im_start|>assistant\n"
)
return ModelRequestData(
engine_args=engine_args,
prompt=prompt,
)
# Qwen2.5-Omni
def run_qwen2_5_omni(question: str, audio_count: int):
model_name = "Qwen/Qwen2.5-Omni-7B"
engine_args = EngineArgs(
model=model_name,
max_model_len=4096,
max_num_seqs=5,
limit_mm_per_prompt={"audio": audio_count},
)
audio_in_prompt = "".join(
["<|audio_bos|><|AUDIO|><|audio_eos|>\n" for idx in range(audio_count)]
)
default_system = (
"You are Qwen, a virtual human developed by the Qwen Team, Alibaba "
"Group, capable of perceiving auditory and visual inputs, as well as "
"generating text and speech."
)
prompt = (
f"<|im_start|>system\n{default_system}<|im_end|>\n"
"<|im_start|>user\n"
f"{audio_in_prompt}{question}<|im_end|>\n"
"<|im_start|>assistant\n"
)
return ModelRequestData(
engine_args=engine_args,
prompt=prompt,
)
# Ultravox 0.5-1B
def run_ultravox(question: str, audio_count: int) -> ModelRequestData:
model_name = "fixie-ai/ultravox-v0_5-llama-3_2-1b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
messages = [{"role": "user", "content": "<|audio|>\n" * audio_count + question}]
prompt = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
engine_args = EngineArgs(
model=model_name,
max_model_len=4096,
max_num_seqs=5,
trust_remote_code=True,
limit_mm_per_prompt={"audio": audio_count},
)
return ModelRequestData(
engine_args=engine_args,
prompt=prompt,
)
# Whisper
def run_whisper(question: str, audio_count: int) -> ModelRequestData:
assert audio_count == 1, "Whisper only support single audio input per prompt"
model_name = "openai/whisper-large-v3-turbo"
prompt = "<|startoftranscript|>"
engine_args = EngineArgs(
model=model_name,
max_model_len=448,
max_num_seqs=5,
limit_mm_per_prompt={"audio": audio_count},
)
return ModelRequestData(
engine_args=engine_args,
prompt=prompt,
)
model_example_map = {
"granite_speech": run_granite_speech,
"minicpmo": run_minicpmo,
"phi4_mm": run_phi4mm,
"qwen2_audio": run_qwen2_audio,
"qwen2_5_omni": run_qwen2_5_omni,
"ultravox": run_ultravox,
"whisper": run_whisper,
}
def parse_args():
parser = FlexibleArgumentParser(
description="Demo on using vLLM for offline inference with "
"audio language models"
)
parser.add_argument(
"--model-type",
"-m",
type=str,
default="ultravox",
choices=model_example_map.keys(),
help='Huggingface "model_type".',
)
parser.add_argument(
"--num-prompts", type=int, default=1, help="Number of prompts to run."
)
parser.add_argument(
"--num-audios",
type=int,
default=1,
choices=[0, 1, 2],
help="Number of audio items per prompt.",
)
parser.add_argument(
"--seed",
type=int,
default=None,
help="Set the seed when initializing `vllm.LLM`.",
)
return parser.parse_args()
def main(args):
model = args.model_type
if model not in model_example_map:
raise ValueError(f"Model type {model} is not supported.")
audio_count = args.num_audios
req_data = model_example_map[model](
question_per_audio_count[audio_count], audio_count
)
# Disable other modalities to save memory
default_limits = {"image": 0, "video": 0, "audio": 0}
req_data.engine_args.limit_mm_per_prompt = default_limits | dict(
req_data.engine_args.limit_mm_per_prompt or {}
)
engine_args = asdict(req_data.engine_args) | {"seed": args.seed}
llm = LLM(**engine_args)
# We set temperature to 0.2 so that outputs can be different
# even when all prompts are identical when running batch inference.
sampling_params = SamplingParams(
temperature=0.2, max_tokens=64, stop_token_ids=req_data.stop_token_ids
)
mm_data = {}
if audio_count > 0:
mm_data = {
"audio": [
asset.audio_and_sample_rate for asset in audio_assets[:audio_count]
]
}
assert args.num_prompts > 0
inputs = {"prompt": req_data.prompt, "multi_modal_data": mm_data}
if args.num_prompts > 1:
# Batch inference
inputs = [inputs] * args.num_prompts
# Add LoRA request if applicable
lora_request = (
req_data.lora_requests * args.num_prompts if req_data.lora_requests else None
)
outputs = llm.generate(
inputs,
sampling_params=sampling_params,
lora_request=lora_request,
)
for o in outputs:
generated_text = o.outputs[0].text
print(generated_text)
if __name__ == "__main__":
args = parse_args()
main(args)
examples/offline_inference/automatic_prefix_caching.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Demonstration script for Automatic Prefix Caching (APC) in vLLM.
Automatic Prefix Caching (APC) allows the vLLM engine to reuse cached
KV (key-value) pairs from previous prompts if a new query shares the same
prefix. This reduces redundant computation and improves inference speed.
To enable APC, set `enable_prefix_caching=True` when initializing the
vLLM engine.
This script uses a long Markdown table as the shared prompt prefix and
compares the generation time for two queries that share the same prefix
but ask different questions.
Run:
python examples/offline_inference/automatic_prefix_caching.py
"""
import time
from vllm import LLM, SamplingParams
# ruff: noqa: E501
# A prompt containing a large markdown table. The table is randomly generated by GPT-4.
LONG_PROMPT = (
"You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as follows.\n# Table\n"
+ """
| ID | Name | Age | Occupation | Country | Email | Phone Number | Address |
|-----|---------------|-----|---------------|---------------|------------------------|----------------|------------------------------|
| 1 | John Doe | 29 | Engineer | USA | john.doe@example.com | 555-1234 | 123 Elm St, Springfield, IL |
| 2 | Jane Smith | 34 | Doctor | Canada | jane.smith@example.com | 555-5678 | 456 Oak St, Toronto, ON |
| 3 | Alice Johnson | 27 | Teacher | UK | alice.j@example.com | 555-8765 | 789 Pine St, London, UK |
| 4 | Bob Brown | 45 | Artist | Australia | bob.b@example.com | 555-4321 | 321 Maple St, Sydney, NSW |
| 5 | Carol White | 31 | Scientist | New Zealand | carol.w@example.com | 555-6789 | 654 Birch St, Wellington, NZ |
| 6 | Dave Green | 28 | Lawyer | Ireland | dave.g@example.com | 555-3456 | 987 Cedar St, Dublin, IE |
| 7 | Emma Black | 40 | Musician | USA | emma.b@example.com | 555-1111 | 246 Ash St, New York, NY |
| 8 | Frank Blue | 37 | Chef | Canada | frank.b@example.com | 555-2222 | 135 Spruce St, Vancouver, BC |
| 9 | Grace Yellow | 50 | Engineer | UK | grace.y@example.com | 555-3333 | 864 Fir St, Manchester, UK |
| 10 | Henry Violet | 32 | Artist | Australia | henry.v@example.com | 555-4444 | 753 Willow St, Melbourne, VIC|
| 11 | Irene Orange | 26 | Scientist | New Zealand | irene.o@example.com | 555-5555 | 912 Poplar St, Auckland, NZ |
| 12 | Jack Indigo | 38 | Teacher | Ireland | jack.i@example.com | 555-6666 | 159 Elm St, Cork, IE |
| 13 | Karen Red | 41 | Lawyer | USA | karen.r@example.com | 555-7777 | 357 Cedar St, Boston, MA |
| 14 | Leo Brown | 30 | Chef | Canada | leo.b@example.com | 555-8888 | 246 Oak St, Calgary, AB |
| 15 | Mia Green | 33 | Musician | UK | mia.g@example.com | 555-9999 | 975 Pine St, Edinburgh, UK |
| 16 | Noah Yellow | 29 | Doctor | Australia | noah.y@example.com | 555-0000 | 864 Birch St, Brisbane, QLD |
| 17 | Olivia Blue | 35 | Engineer | New Zealand | olivia.b@example.com | 555-1212 | 753 Maple St, Hamilton, NZ |
| 18 | Peter Black | 42 | Artist | Ireland | peter.b@example.com | 555-3434 | 912 Fir St, Limerick, IE |
| 19 | Quinn White | 28 | Scientist | USA | quinn.w@example.com | 555-5656 | 159 Willow St, Seattle, WA |
| 20 | Rachel Red | 31 | Teacher | Canada | rachel.r@example.com | 555-7878 | 357 Poplar St, Ottawa, ON |
| 21 | Steve Green | 44 | Lawyer | UK | steve.g@example.com | 555-9090 | 753 Elm St, Birmingham, UK |
| 22 | Tina Blue | 36 | Musician | Australia | tina.b@example.com | 555-1213 | 864 Cedar St, Perth, WA |
| 23 | Umar Black | 39 | Chef | New Zealand | umar.b@example.com | 555-3435 | 975 Spruce St, Christchurch, NZ|
| 24 | Victor Yellow | 43 | Engineer | Ireland | victor.y@example.com | 555-5657 | 246 Willow St, Galway, IE |
| 25 | Wendy Orange | 27 | Artist | USA | wendy.o@example.com | 555-7879 | 135 Elm St, Denver, CO |
| 26 | Xavier Green | 34 | Scientist | Canada | xavier.g@example.com | 555-9091 | 357 Oak St, Montreal, QC |
| 27 | Yara Red | 41 | Teacher | UK | yara.r@example.com | 555-1214 | 975 Pine St, Leeds, UK |
| 28 | Zack Blue | 30 | Lawyer | Australia | zack.b@example.com | 555-3436 | 135 Birch St, Adelaide, SA |
| 29 | Amy White | 33 | Musician | New Zealand | amy.w@example.com | 555-5658 | 159 Maple St, Wellington, NZ |
| 30 | Ben Black | 38 | Chef | Ireland | ben.b@example.com | 555-7870 | 246 Fir St, Waterford, IE |
"""
)
def get_generation_time(llm, sampling_params, prompts):
# time the generation
start_time = time.time()
output = llm.generate(prompts, sampling_params=sampling_params)
end_time = time.time()
# print the output and generation time
print("-" * 30)
print(f"Output: {output[0].outputs[0].text}")
print(f"Generation time: {end_time - start_time} seconds.")
print("-" * 30)
def main():
# set enable_prefix_caching=True to enable APC
llm = LLM(model="lmsys/longchat-13b-16k", enable_prefix_caching=True)
sampling_params = SamplingParams(temperature=0, max_tokens=100)
# Querying the age of John Doe
get_generation_time(
llm,
sampling_params,
LONG_PROMPT
+ "Question: what is the age of John Doe? Your answer: The age of John Doe is ",
)
# Querying the age of Zack Blue
# This query will be faster since vllm avoids computing the KV cache of LONG_PROMPT again.
get_generation_time(
llm,
sampling_params,
LONG_PROMPT
+ "Question: what is the age of Zack Blue? Your answer: The age of Zack Blue is ",
)
if __name__ == "__main__":
main()
examples/offline_inference/basic/README.md 0 → 100644
View file @ cea23494
# Basic
The `LLM` class provides the primary Python interface for doing offline inference, which is interacting with a model without using a separate model inference server.
## Usage
The first script in this example shows the most basic usage of vLLM. If you are new to Python and vLLM, you should start here.
```bash
python examples/offline_inference/basic/basic.py
```
The rest of the scripts include an [argument parser](https://docs.python.org/3/library/argparse.html), which you can use to pass any arguments that are compatible with [`LLM`](https://docs.vllm.ai/en/latest/api/offline_inference/llm.html). Try running the script with `--help` for a list of all available arguments.
```bash
python examples/offline_inference/basic/classify.py
```
```bash
python examples/offline_inference/basic/embed.py
```
```bash
python examples/offline_inference/basic/score.py
```
The chat and generate scripts also accept the [sampling parameters](https://docs.vllm.ai/en/latest/api/inference_params.html#sampling-parameters): `max_tokens`, `temperature`, `top_p` and `top_k`.
```bash
python examples/offline_inference/basic/chat.py
```
```bash
python examples/offline_inference/basic/generate.py
```
## Features
In the scripts that support passing arguments, you can experiment with the following features.
### Default generation config
The `--generation-config` argument specifies where the generation config will be loaded from when calling `LLM.get_default_sampling_params()`. If set to ‘auto’, the generation config will be loaded from model path. If set to a folder path, the generation config will be loaded from the specified folder path. If it is not provided, vLLM defaults will be used.
> If max_new_tokens is specified in generation config, then it sets a server-wide limit on the number of output tokens for all requests.
Try it yourself with the following argument:
```bash
--generation-config auto
```
### Quantization
#### AQLM
vLLM supports models that are quantized using AQLM.
Try one yourself by passing one of the following models to the `--model` argument:
- `ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf`
- `ISTA-DASLab/Llama-2-7b-AQLM-2Bit-2x8-hf`
- `ISTA-DASLab/Llama-2-13b-AQLM-2Bit-1x16-hf`
- `ISTA-DASLab/Mixtral-8x7b-AQLM-2Bit-1x16-hf`
- `BlackSamorez/TinyLlama-1_1B-Chat-v1_0-AQLM-2Bit-1x16-hf`
> Some of these models are likely to be too large for a single GPU. You can split them across multiple GPUs by setting `--tensor-parallel-size` to the number of required GPUs.
#### GGUF
vLLM supports models that are quantized using GGUF.
Try one yourself by downloading a quantized GGUF model and using the following arguments:
```python
from huggingface_hub import hf_hub_download
repo_id = "bartowski/Phi-3-medium-4k-instruct-GGUF"
filename = "Phi-3-medium-4k-instruct-IQ2_M.gguf"
print(hf_hub_download(repo_id, filename=filename))
```
```bash
--model {local-path-printed-above} --tokenizer microsoft/Phi-3-medium-4k-instruct
```
### CPU offload
The `--cpu-offload-gb` argument can be seen as a virtual way to increase the GPU memory size. For example, if you have one 24 GB GPU and set this to 10, virtually you can think of it as a 34 GB GPU. Then you can load a 13B model with BF16 weight, which requires at least 26GB GPU memory. Note that this requires fast CPU-GPU interconnect, as part of the model is loaded from CPU memory to GPU memory on the fly in each model forward pass.
Try it yourself with the following arguments:
```bash
--model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10
```
examples/offline_inference/basic/basic.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm import LLM, SamplingParams
# Sample prompts.
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
"Hello, my name is",
]
# Create a sampling params object.
sampling_params = SamplingParams(temperature=0.8, top_p=0.95, max_tokens=16)
def main():
# Create an LLM.
llm = LLM(model="/mnt/data/llm-models/qwen3/Qwen3-8B",tensor_parallel_size=1, dtype="float16",trust_remote_code=True, enforce_eager=True, block_size=64, enable_prefix_caching=False)
# Generate texts from the prompts.
# The output is a list of RequestOutput objects
# that contain the prompt, generated text, and other information.
outputs = llm.generate(prompts, sampling_params)
# Print the outputs.
print("\nGenerated Outputs:\n" + "-" * 60)
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}")
print(f"Output: {generated_text!r}")
print("-" * 60)
if __name__ == "__main__":
main()
examples/offline_inference/basic/chat.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm import LLM, EngineArgs
from vllm.utils import FlexibleArgumentParser
def create_parser():
parser = FlexibleArgumentParser()
# Add engine args
EngineArgs.add_cli_args(parser)
parser.set_defaults(model="meta-llama/Llama-3.2-1B-Instruct")
# Add sampling params
sampling_group = parser.add_argument_group("Sampling parameters")
sampling_group.add_argument("--max-tokens", type=int)
sampling_group.add_argument("--temperature", type=float)
sampling_group.add_argument("--top-p", type=float)
sampling_group.add_argument("--top-k", type=int)
# Add example params
parser.add_argument("--chat-template-path", type=str)
return parser
def main(args: dict):
# Pop arguments not used by LLM
max_tokens = args.pop("max_tokens")
temperature = args.pop("temperature")
top_p = args.pop("top_p")
top_k = args.pop("top_k")
chat_template_path = args.pop("chat_template_path")
# Create an LLM
llm = LLM(**args)
# Create sampling params object
sampling_params = llm.get_default_sampling_params()
if max_tokens is not None:
sampling_params.max_tokens = max_tokens
if temperature is not None:
sampling_params.temperature = temperature
if top_p is not None:
sampling_params.top_p = top_p
if top_k is not None:
sampling_params.top_k = top_k
def print_outputs(outputs):
print("\nGenerated Outputs:\n" + "-" * 80)
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}\n")
print(f"Generated text: {generated_text!r}")
print("-" * 80)
print("=" * 80)
# In this script, we demonstrate how to pass input to the chat method:
conversation = [
{"role": "system", "content": "You are a helpful assistant"},
{"role": "user", "content": "Hello"},
{"role": "assistant", "content": "Hello! How can I assist you today?"},
{
"role": "user",
"content": "Write an essay about the importance of higher education.",
},
]
outputs = llm.chat(conversation, sampling_params, use_tqdm=False)
print_outputs(outputs)
# You can run batch inference with llm.chat API
conversations = [conversation for _ in range(10)]
# We turn on tqdm progress bar to verify it's indeed running batch inference
outputs = llm.chat(conversations, sampling_params, use_tqdm=True)
print_outputs(outputs)
# A chat template can be optionally supplied.
# If not, the model will use its default chat template.
if chat_template_path is not None:
with open(chat_template_path) as f:
chat_template = f.read()
outputs = llm.chat(
conversations,
sampling_params,
use_tqdm=False,
chat_template=chat_template,
)
if __name__ == "__main__":
parser = create_parser()
args: dict = vars(parser.parse_args())
main(args)
examples/offline_inference/basic/classify.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from argparse import Namespace
from vllm import LLM, EngineArgs
from vllm.utils import FlexibleArgumentParser
def parse_args():
parser = FlexibleArgumentParser()
parser = EngineArgs.add_cli_args(parser)
# Set example specific arguments
parser.set_defaults(
model="jason9693/Qwen2.5-1.5B-apeach", task="classify", enforce_eager=True
)
return parser.parse_args()
def main(args: Namespace):
# Sample prompts.
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
# Create an LLM.
# You should pass task="classify" for classification models
model = LLM(**vars(args))
# Generate logits. The output is a list of ClassificationRequestOutputs.
outputs = model.classify(prompts)
# Print the outputs.
print("\nGenerated Outputs:\n" + "-" * 60)
for prompt, output in zip(prompts, outputs):
probs = output.outputs.probs
probs_trimmed = (str(probs[:16])[:-1] + ", ...]") if len(probs) > 16 else probs
print(
f"Prompt: {prompt!r} \n"
f"Class Probabilities: {probs_trimmed} (size={len(probs)})"
)
print("-" * 60)
if __name__ == "__main__":
args = parse_args()
main(args)
examples/offline_inference/basic/embed.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from argparse import Namespace
from vllm import LLM, EngineArgs
from vllm.utils import FlexibleArgumentParser
def parse_args():
parser = FlexibleArgumentParser()
parser = EngineArgs.add_cli_args(parser)
# Set example specific arguments
parser.set_defaults(
model="intfloat/e5-mistral-7b-instruct",
task="embed",
enforce_eager=True,
max_model_len=1024,
)
return parser.parse_args()
def main(args: Namespace):
# Sample prompts.
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
# Create an LLM.
# You should pass task="embed" for embedding models
model = LLM(**vars(args))
# Generate embedding. The output is a list of EmbeddingRequestOutputs.
outputs = model.embed(prompts)
# Print the outputs.
print("\nGenerated Outputs:\n" + "-" * 60)
for prompt, output in zip(prompts, outputs):
embeds = output.outputs.embedding
embeds_trimmed = (
(str(embeds[:16])[:-1] + ", ...]") if len(embeds) > 16 else embeds
)
print(f"Prompt: {prompt!r} \nEmbeddings: {embeds_trimmed} (size={len(embeds)})")
print("-" * 60)
if __name__ == "__main__":
args = parse_args()
main(args)
examples/offline_inference/basic/generate.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm import LLM, EngineArgs
from vllm.utils import FlexibleArgumentParser
def create_parser():
parser = FlexibleArgumentParser()
# Add engine args
EngineArgs.add_cli_args(parser)
parser.set_defaults(model="meta-llama/Llama-3.2-1B-Instruct")
# Add sampling params
sampling_group = parser.add_argument_group("Sampling parameters")
sampling_group.add_argument("--max-tokens", type=int)
sampling_group.add_argument("--temperature", type=float)
sampling_group.add_argument("--top-p", type=float)
sampling_group.add_argument("--top-k", type=int)
return parser
def main(args: dict):
# Pop arguments not used by LLM
max_tokens = args.pop("max_tokens")
temperature = args.pop("temperature")
top_p = args.pop("top_p")
top_k = args.pop("top_k")
# Create an LLM
llm = LLM(**args)
# Create a sampling params object
sampling_params = llm.get_default_sampling_params()
if max_tokens is not None:
sampling_params.max_tokens = max_tokens
if temperature is not None:
sampling_params.temperature = temperature
if top_p is not None:
sampling_params.top_p = top_p
if top_k is not None:
sampling_params.top_k = top_k
# Generate texts from the prompts. The output is a list of RequestOutput
# objects that contain the prompt, generated text, and other information.
prompts = [
"Hello, my name is",
"The president of the United States is",
"The capital of France is",
"The future of AI is",
]
outputs = llm.generate(prompts, sampling_params)
# Print the outputs.
print("-" * 50)
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
print("-" * 50)
if __name__ == "__main__":
parser = create_parser()
args: dict = vars(parser.parse_args())
main(args)
examples/offline_inference/basic/score.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from argparse import Namespace
from vllm import LLM, EngineArgs
from vllm.utils import FlexibleArgumentParser
def parse_args():
parser = FlexibleArgumentParser()
parser = EngineArgs.add_cli_args(parser)
# Set example specific arguments
parser.set_defaults(
model="BAAI/bge-reranker-v2-m3", task="score", enforce_eager=True
)
return parser.parse_args()
def main(args: Namespace):
# Sample prompts.
text_1 = "What is the capital of France?"
texts_2 = [
"The capital of Brazil is Brasilia.",
"The capital of France is Paris.",
]
# Create an LLM.
# You should pass task="score" for cross-encoder models
model = LLM(**vars(args))
# Generate scores. The output is a list of ScoringRequestOutputs.
outputs = model.score(text_1, texts_2)
# Print the outputs.
print("\nGenerated Outputs:\n" + "-" * 60)
for text_2, output in zip(texts_2, outputs):
score = output.outputs.score
print(f"Pair: {[text_1, text_2]!r} \nScore: {score}")
print("-" * 60)
if __name__ == "__main__":
args = parse_args()
main(args)
examples/offline_inference/batch_llm_inference.py 0 → 100644
View file @ cea23494
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
This example shows how to use Ray Data for data parallel batch inference.
Ray Data is a data processing framework that can handle large datasets
and integrates tightly with vLLM for data-parallel inference.
As of Ray 2.44, Ray Data has a native integration with
vLLM (under ray.data.llm).
Ray Data provides functionality for:
* Reading and writing to cloud storage (S3, GCS, etc.)
* Automatic sharding and load-balancing across a cluster
* Optimized configuration of vLLM using continuous batching
* Compatible with tensor/pipeline parallel inference as well.
Learn more about Ray Data's LLM integration:
https://docs.ray.io/en/latest/data/working-with-llms.html
"""
import ray
from packaging.version import Version
from ray.data.llm import build_llm_processor, vLLMEngineProcessorConfig
assert Version(ray.__version__) >= Version("2.44.1"), (
"Ray version must be at least 2.44.1"
)
# Uncomment to reduce clutter in stdout
# ray.init(log_to_driver=False)
# ray.data.DataContext.get_current().enable_progress_bars = False
# Read one text file from S3. Ray Data supports reading multiple files
# from cloud storage (such as JSONL, Parquet, CSV, binary format).
ds = ray.data.read_text("s3://anonymous@air-example-data/prompts.txt")
print(ds.schema())
size = ds.count()
print(f"Size of dataset: {size} prompts")
# Configure vLLM engine.
config = vLLMEngineProcessorConfig(
model_source="unsloth/Llama-3.1-8B-Instruct",
engine_kwargs={
"enable_chunked_prefill": True,
"max_num_batched_tokens": 4096,
"max_model_len": 16384,
},
concurrency=1, # set the number of parallel vLLM replicas
batch_size=64,
)
# Create a Processor object, which will be used to
# do batch inference on the dataset
vllm_processor = build_llm_processor(
config,
preprocess=lambda row: dict(
messages=[
{"role": "system", "content": "You are a bot that responds with haikus."},
{"role": "user", "content": row["text"]},
],
sampling_params=dict(
temperature=0.3,
max_tokens=250,
),
),
postprocess=lambda row: dict(
answer=row["generated_text"],
**row, # This will return all the original columns in the dataset.
),
)
ds = vllm_processor(ds)
# Peek first 10 results.
# NOTE: This is for local testing and debugging. For production use case,
# one should write full result out as shown below.
outputs = ds.take(limit=10)
for output in outputs:
prompt = output["prompt"]
generated_text = output["generated_text"]
print(f"Prompt: {prompt!r}")
print(f"Generated text: {generated_text!r}")
# Write inference output data out as Parquet files to S3.
# Multiple files would be written to the output destination,
# and each task would write one or more files separately.
#
# ds.write_parquet("s3://<your-output-bucket>")
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