vllm量化02—awq

本系列基于Qwen2.5-7B，学习如何使用vllm量化，并使用benchmark_serving.py、lm_eval 测试模型性能和评估模型准确度。
测试环境为：

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

本文是该系列第2篇——awq量化

一、量化

使用AutoAWQ 量化

from awq import AutoAWQForCausalLM
from transformers import AutoTokenizermodel_path = "./Qwen2.5-7B"
quant_path = "./Qwen2.5-7B-awq-int4"
quant_config = { "zero_point": True, "q_group_size": 128, "w_bit": 4, "version": "GEMM" }# Load model
model = AutoAWQForCausalLM.from_pretrained(model_path, trust_remote_code=False)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=False)# Quantize
model.quantize(tokenizer, quant_config=quant_config)# Save quantized model
model.save_quantized(quant_path)
tokenizer.save_pretrained(quant_path)print(f'Model is quantized and saved at "{quant_path}"')

二、部署

 vllm  serve Qwen2.5-7B-awq-int4 --disable-log-requests --quantization awq --dtype="half"

三、benchmark

python /vllm/benchmarks/benchmark_serving.py --backend vllm --model Qwen2.5-7B-awq-int4 --endpoint /v1/completions --dataset-name sharegpt --dataset-path ./ShareGPT_Vicuna_unfiltered/ShareGPT_V3_unfiltered_cleaned_split.json  --num-prompts 100

结果：

============ Serving Benchmark Result ============
Successful requests:                     100
Benchmark duration (s):                  48.32
Total input tokens:                      23260
Total generated tokens:                  21979
Request throughput (req/s):              2.07
Output token throughput (tok/s):         454.85
Total Token throughput (tok/s):          936.22
---------------Time to First Token----------------
Mean TTFT (ms):                          2132.58
Median TTFT (ms):                        2277.96
P99 TTFT (ms):                           3749.79
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          95.75
Median TPOT (ms):                        69.59
P99 TPOT (ms):                           309.96
---------------Inter-token Latency----------------
Mean ITL (ms):                           65.81
Median ITL (ms):                         59.83
P99 ITL (ms):                            311.10

四、lm_eval

4.1 gsm8k

lm_eval --model vllm \ --model_args pretrained="Qwen2.5-7B-awq-int4",add_bos_token=true,gpu_memory_utilization=0.5,quantization="AWQ",dtype="half"  \--tasks mmlu  \--num_fewshot 5 \--limit 250

结果：

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.832	±	0.0237
		strict-match	5	exact_match	↑	0.740	±	0.0278

4.2 mmlu

lm_eval --model vllm  \--model_args pretrained="./Qwen2.5-7B-awq-int4/",add_bos_token=true,gpu_memory_utilization=0.5,quantization="AWQ",dtype="half"  \--tasks mmlu \--num_fewshot 5 \--limit 250 \--batch_size 'auto'

结果：

Groups	Version	Filter	n-shot	Metric		Value		Stderr
mmlu	2	none		acc	↑	0.7517	±	0.0041
- humanities	2	none		acc	↑	0.7576	±	0.0082
- other	2	none		acc	↑	0.7526	±	0.0084
- social sciences	2	none		acc	↑	0.8285	±	0.0077
- stem	2	none		acc	↑	0.6866	±	0.0083