本系列基于Qwen2.5-7B，學習如何使用vllm量化，并使用benchmark_serving.py、lm_eval 測試模型性能和評估模型準確度。
測試環境為：

OS: centos 7
GPU: nvidia l40
driver: 550.54.15
CUDA: 12.3

本文是該系列第3篇——INT4 W4A16

一、量化

from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset
from llmcompressor.transformers import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
from vllm import LLMmodel_path = "./Qwen2.5-7B"
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", torch_dtype="auto", trust_remote_code=False)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=False)NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048# Load and preprocess the dataset
ds = load_dataset("/data/ultrachat_200k", split="train_sft")
ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))def preprocess(example):return {"text": tokenizer.apply_chat_template(example["messages"], tokenize=False)}
ds = ds.map(preprocess)def tokenize(sample):return tokenizer(sample["text"], padding=False, max_length=MAX_SEQUENCE_LENGTH, truncation=True, add_special_tokens=False)
ds = ds.map(tokenize, remove_columns=ds.column_names)# Configure the quantization algorithms
recipe = GPTQModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"])# Apply quantization
oneshot(model=model,dataset=ds,recipe=recipe,max_seq_length=MAX_SEQUENCE_LENGTH,num_calibration_samples=NUM_CALIBRATION_SAMPLES,
)# Save the compressed model
SAVE_DIR = "./Qwen2.5-1.5B-W4A16-G128"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)

二、部署

 vllm  serve Qwen2.5-1.5B-W4A16-G128 --disable-log-requests 

三、benchmark

python /vllm/benchmarks/benchmark_serving.py --backend vllm --model Qwen2.5-1.5B-W4A16-G128 --endpoint /v1/completions --dataset-name sharegpt --dataset-path ./ShareGPT_Vicuna_unfiltered/ShareGPT_V3_unfiltered_cleaned_split.json  --num-prompts 100

============ Serving Benchmark Result ============
Successful requests:                     100
Benchmark duration (s):                  12.28
Total input tokens:                      23260
Total generated tokens:                  22059
Request throughput (req/s):              8.14
Output token throughput (tok/s):         1795.86
Total Token throughput (tok/s):          3689.50
---------------Time to First Token----------------
Mean TTFT (ms):                          1529.15
Median TTFT (ms):                        1562.42
P99 TTFT (ms):                           2646.95
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          44.98
Median TPOT (ms):                        22.79
P99 TPOT (ms):                           223.34
---------------Inter-token Latency----------------
Mean ITL (ms):                           19.59
Median ITL (ms):                         14.49
P99 ITL (ms):                            222.72
==================================================

四、lm_eval

4.1 gsm8k

lm_eval --model vllm  \--model_args pretrained="./w4a16/Qwen2.5-1.5B-W4A16-G128",add_bos_token=true,gpu_memory_utilization=0.5  \--tasks gsm8k \--num_fewshot 5  \--limit 250 \--batch_size 'auto'

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.784	±	0.0261
		strict-match	5	exact_match	↑	0.688	±	0.0294

4.2 mmlu

lm_eval --model vllm  \--model_args pretrained="./w4a16/Qwen2.5-1.5B-W4A16-G128",add_bos_token=true,gpu_memory_utilization=0.5  \--tasks mmlu \--num_fewshot 5  \--limit 250 \--batch_size 'auto'

Groups	Version	Filter	n-shot	Metric		Value		Stderr
mmlu	2	none		acc	↑	0.7527	±	0.0041
- humanities	2	none		acc	↑	0.7635	±	0.0081
- other	2	none		acc	↑	0.7485	±	0.0084
- social sciences	2	none		acc	↑	0.8333	±	0.0077
- stem	2	none		acc	↑	0.6846	±	0.0083