vllm测试 0.7.2 Offline推理 benchmark_throughput_0.7.2.py 使用如下脚本可以减少不同参数推理时反复load模型 batch prompt_tokens completion_tokens可以用空格分隔传成字符串 其他参数与标准脚本一致 bash export HIP_VISIBLE_DEVICES=1 tp=1 model_path=/llm-models/qwen1.5/Qwen1.5-0.5B-Chat batch="1 2" prompt_tokens="16 64" completion_tokens="128 256" python benchmark_throughput_0.7.2.py --model ${model_path} --tensor-parallel-size ${tp} --num-prompts ${batch} --input-len ${prompt_tokens} --output-len ${completion_tokens} \ --dtype float16 --trust-remote-code --max-model-len 32768 --output-json ./test_0.5B-0.7.2.txt 按照如上传参，则计算的场景如下： bs input output 1 16 128 1 64 256 2 16 128 2 64 256 推理结果汇总在--output-json ./test_0.5B-0.7.2.txt当中,示例如下表： bash bs_in_out,elapsed_time,Throughput,total_tokens,output_tokens,ttft_mean,ttft_median,ttft_p99,tpop_mean,tpop_median,tpop_p99,output_token_throughput_mean,output_token_throughput_median,output_token_throughput_p99,inout_token_throughput_mean,inout_token_throughput_median,inout_token_throughput_p99 1_16_128,3.49,0.29,41.26,36.68,0.03801,0.03801,0.03801,0.0269,0.02691,0.02691,37.04,37.04,37.04,41.66,41.66,41.66 1_64_256,7.14,0.14,44.82,35.85,0.0291,0.0291,0.0291,0.0278,0.02776,0.02776,36.01,36.01,36.01,45.01,45.01,45.01 2_16_128,3.62,0.55,79.56,70.72,0.04829,0.04829,0.04893,0.028,0.02801,0.02801,35.51,35.51,35.51,39.94,39.94,39.95 2_64_256,7.31,0.27,87.55,70.04,0.04697,0.04697,0.04764,0.0284,0.02836,0.02836,35.17,35.17,35.18,43.97,43.97,43.97 Server推理 benchmark_servein_0.7.2.py backend_request_func.py 使用此方式可以避免sever推理时实际生成长度和指定长度不一致问题 bash #使用提供的脚本进行测试 #启动server vllm serve $MODEL_PATH --trust-remote-code --dtype $dtype --max-model-len $max_len -tp $tp --gpu-memory-utilization 0.97 #发送请求 #--distributed-executor-backend ray等其他参数根据实际情况添加 方式与平常一样，只是需要加上--ignore-eos python benchmark_servein_0.7.2.py --backend vllm --ignore-eos --dataset-name random --random-input-len $input_len --random-output-len $output_len --model $MODEL_PATH --num-prompts $num_prompts --endpoint /v1/completions prof offline_prof hipprof prof.py benchmark_throughput_0.7.2_hipprof.py bash #使用示例： 黄色背景为额外添加的部分 SGLANG_PROF_ROCTX=1 hipprof --trace-off python benchmark_throughput_0.7.2_hipprof.py --num-prompts 1 --input-len 2000 --output-len 1 --model /models/Llama-2-7b-hf --trust-remote-code --enforce-eager --dtype float16 > 7b-prefill-2000-test.log 2>&1 torchprof benchmark_throughput_0.7.2_torchprof.py bash #启动方式与平常使用一致 benchmark_throughput_0.7.2_torchprof.py --num-prompts 1 --input-len 2000 --output-len 1 --model /models/Llama-2-7b-hf --trust-remote-code --enforce-eager --dtype float16 > 7b-prefill-2000-test.log 2>&1 会打印prof信息，保存的json文件名为: {args.num_prompts}-{args.input_len}-{args.output_len}-{args.tensor_parallel_size}_dcu.json server_prof worker.py bash 替换/usr/local/lib/python3.10/site-packages/vllm/worker/worker.py #启动服务 loca_path为保存的json文件绝对路径 export VLLM_TORCH_PROFILER_DIR=$loca_path vllm serve $MODEL_PATH --trust-remote-code --dtype $dtype --max-model-len $max_len -tp $tp --gpu-memory-utilization 0.97 #发送请求 #--distributed-executor-backend ray等其他参数根据实际情况添加 python benchmark_servein_0.7.2.py --backend vllm --ignore-eos --profile --dataset-name random --random-input-len $input_len --random-output-len $output_len --model $MODEL_PATH --num-prompts $num_prompts --endpoint /v1/completions 0.6.2 Offline推理 benchmark_throughput_0.6.2.py 使用如下脚本可以减少不同参数推理时反复load模型 batch prompt_tokens completion_tokens可以用空格分隔传成字符串 其他参数与标准脚本一致 bash export HIP_VISIBLE_DEVICES=1 tp=1 model_path=/llm-models/qwen1.5/Qwen1.5-0.5B-Chat batch="1 2" prompt_tokens="16 64" completion_tokens="128 256" python benchmark_throughput_0.6.2.py --model ${model_path} --tensor-parallel-size ${tp} --num-prompts ${batch} --input-len ${prompt_tokens} --output-len ${completion_tokens} \ --dtype float16 --trust-remote-code --max-model-len 32768 --output-json ./test_0.5B-0.6.2.txt 按照如上传参，则计算的场景如下： bs input output 1 16 128 1 64 256 2 16 128 2 64 256 推理结果汇总在--output-json ./test_0.5B-0.6.2.txt当中,示例如下表： bash bs_in_out,elapsed_time,Throughput,total_tokens,output_tokens,ttft_mean,ttft_median,ttft_p99,tpop_mean,tpop_median,tpop_p99,output_token_throughput_mean,output_token_throughput_median,output_token_throughput_p99,inout_token_throughput_mean,inout_token_throughput_median,inout_token_throughput_p99 1_16_128,3.49,0.29,41.26,36.68,0.03801,0.03801,0.03801,0.0269,0.02691,0.02691,37.04,37.04,37.04,41.66,41.66,41.66 1_64_256,7.14,0.14,44.82,35.85,0.0291,0.0291,0.0291,0.0278,0.02776,0.02776,36.01,36.01,36.01,45.01,45.01,45.01 2_16_128,3.62,0.55,79.56,70.72,0.04829,0.04829,0.04893,0.028,0.02801,0.02801,35.51,35.51,35.51,39.94,39.94,39.95 2_64_256,7.31,0.27,87.55,70.04,0.04697,0.04697,0.04764,0.0284,0.02836,0.02836,35.17,35.17,35.18,43.97,43.97,43.97 Server推理 benchmark_servein_0.6.2.py backend_request_func.py 使用此方式可以减少server生成长度和指定长度差距过大 Bash #使用提供的脚本进行测试 #启动server vllm serve $MODEL_PATH --trust-remote-code --dtype $dtype --max-model-len $max_len -tp $tp --gpu-memory-utilization 0.97 #发送请求 #--distributed-executor-backend ray等其他参数根据实际情况添加 方式与平常一样，只是需要加上--ignore-eos python benchmark_servein_0.6.2.py --backend vllm --ignore-eos --dataset-name random --random-input-len $input_len --random-output-len $output_len --model $MODEL_PATH --num-prompts $num_prompts --endpoint /v1/completions prof offline_porf hipprof prof.py benchmark_throughput_0.6.2_hipprof.py bash #使用示例： 黄色背景为额外添加的部分 SGLANG_PROF_ROCTX=1 hipprof --trace-off python benchmark_throughput_0.6.2_hipprof.py --num-prompts 1 --input-len 2000 --output-len 1 --model /models/Llama-2-7b-hf --trust-remote-code --enforce-eager --dtype float16 > 7b-prefill-2000-test.log 2>&1 torchprof benchmark_throughput_0.6.2_torchprof.py bash #启动方式与平常使用一致 benchmark_throughput_0.6.2_torchprof.py --num-prompts 1 --input-len 2000 --output-len 1 --model /models/Llama-2-7b-hf --trust-remote-code --enforce-eager --dtype float16 > 7b-prefill-2000-test.log 2>&1 会打印prof信息，保存的json文件名为: {args.num_prompts}-{args.input_len}-{args.output_len}-{args.tensor_parallel_size}_dcu.json server_prof worker.py bash 替换/usr/local/lib/python3.10/site-packages/vllm/worker/worker.py #启动服务 loca_path为保存的json文件绝对路径 export VLLM_TORCH_PROFILER_DIR=$loca_path vllm serve $MODEL_PATH --trust-remote-code --dtype $dtype --max-model-len $max_len -tp $tp --gpu-memory-utilization 0.97 #发送请求 #--distributed-executor-backend ray等其他参数根据实际情况添加 python benchmark_servein_0.6.2.py --backend vllm --ignore-eos --profile --dataset-name random --random-input-len $input_len --random-output-len $output_len --model $MODEL_PATH --num-prompts $num_prompts --endpoint /v1/completions