Benchmark

执行Benchmark

find /usr/local/lib/python3.11/site-packages/mindie* -name *config.json |xargs chmod -R 640

# 开启日志打屏,关闭时日志默认保存在~/mindie/log/debug/mindie-benchmarkxxxxx.log
export MINDIE_LOG_TO_STDOUT="benchmark:1; client:0"

benchmark --DatasetType synthetic --ModelName deepseek-r1-qwen-32b --ModelPath /data/llm_models/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --TestType vllm_client --Http http://127.0.0.1:1025 --ManagementHttp http://127.0.0.1:1026 --Concurrency 10 --TaskKind stream --MaxOutputLen 512 --Tokenizer True

结果解析

GeneratedTokenSpeed = GeneratedTokens / GenerateTime
Throughput = RequestCount / TimeElapsed
OutputGenerateSpeed = GeneratedTokens * RequestCount / TimeElapsed
InputGenerateSpeed = InputTokens * RequestCount / TimeElapsed
TotalGenerateSpeed = OutputGenerateSpeed + InputGenerateSpeed
GenerateSpeedPerClient = OutputGenerateSpeed / Concurrency