批量跑EvalScope推理性能压测

#!/bin/bash

# 并发列表（你可以自己改）
PARALLEL_LIST=(131 132 133 134)

# 固定参数
URL="http://10.63.114.139:8080/v1/chat/completions"
MODEL="/data/Qwen3-32B"
TOKENIZER="/home/tzc/quanzhongwenjian/Qwen3-32B"

for PARALLEL in "${PARALLEL_LIST[@]}"
do
    NUMBER=$((PARALLEL * 5))

    LOG_FILE="qwen3-32b_p${PARALLEL}_n${NUMBER}.log"

    echo "========================================="
    echo "Running test: parallel=${PARALLEL}, number=${NUMBER}"
    echo "Log: ${LOG_FILE}"
    echo "========================================="

    evalscope perf \
      --api openai \
      --url ${URL} \
      --model ${MODEL} \
      --tokenizer-path ${TOKENIZER} \
      --dataset random \
      --number ${NUMBER} \
      --parallel ${PARALLEL} \
      --rate 10 \
      --min-prompt-length 2048 \
      --max-prompt-length 2048 \
      --min-tokens 2048 \
      --max-tokens 2048 \
      --extra-args '{"ignore_eos":true, "temperature":0.0, "top_p":1.0, "skip_special_tokens": true}' \
      --stream \
      --no-test-connection \
      --prefix-length 0 \
      > ${LOG_FILE} 2>&1

    echo "Finished: parallel=${PARALLEL}"
done

echo "All tests done."

执行脚本

nohup bash run_perf.sh > run_all.log 2>&1 &

结果整理脚本

root@# cat parse_perf_txt1.sh
#!/bin/bash

OUTPUT_FILE="result.txt"

# 表头（对齐 + 空格分隔）
printf "%-10s %-10s %-12s %-10s %-15s %-30s %-30s\n" \
"parallel" "number" "TTFT(ms)" "TPOT(ms)" "请求频率(req/s)" "单卡整体吞吐(tok/s)" "单卡输出吞吐(tok/s)" > $OUTPUT_FILE

for file in qwen3-32b_p*_n*.log; do
    echo "Processing $file..."

    # 只处理跑完的日志
    if ! grep -q "Save the summary to:" "$file"; then
        echo "Skip $file (not finished)"
        continue
    fi

    parallel=$(echo $file | sed -E 's/.*_p([0-9]+)_n([0-9]+).*/\1/')
    number=$(echo $file | sed -E 's/.*_p([0-9]+)_n([0-9]+).*/\2/')

    # 提取数值（只取数字，规避 | 等符号）
    TTFT=$(grep "Average time to first token" "$file" | tail -1 | grep -oE "[0-9]+\.[0-9]+")
    TPOT=$(grep "Average time per output token" "$file" | tail -1 | grep -oE "[0-9]+\.[0-9]+")
    REQ=$(grep "Request throughput" "$file" | tail -1 | grep -oE "[0-9]+\.[0-9]+")
    TOTAL_TPS=$(grep "Total token throughput" "$file" | tail -1 | grep -oE "[0-9]+\.[0-9]+")
    OUTPUT_TPS=$(grep "Output token throughput" "$file" | tail -1 | grep -oE "[0-9]+\.[0-9]+")

    # 判空
    if [[ -z "$TTFT" || -z "$TPOT" || -z "$REQ" ]]; then
        echo "Skip $file (parse failed)"
        continue
    fi

    # 单位换算
    TTFT_MS=$(awk -v x=$TTFT 'BEGIN{printf "%.2f", x*1000}')
    TPOT_MS=$(awk -v x=$TPOT 'BEGIN{printf "%.2f", x*1000}')
    SINGLE_TOTAL=$(awk -v x=$TOTAL_TPS 'BEGIN{printf "%.2f", x/4}')
    SINGLE_OUTPUT=$(awk -v x=$OUTPUT_TPS 'BEGIN{printf "%.2f", x/4}')

    # 去掉所有可能的内部空格（保险）
    TOTAL_TPS=$(echo $TOTAL_TPS | tr -d ' ')
    OUTPUT_TPS=$(echo $OUTPUT_TPS | tr -d ' ')

    # 输出（列之间是对齐空格，字段内部无空格）
    printf "%-10s %-10s %-12s %-10s %-15s %-30s %-30s\n" \
    "$parallel" "$number" "$TTFT_MS" "$TPOT_MS" "$REQ" \
    "${SINGLE_TOTAL}(=${TOTAL_TPS}/4)" \
    "${SINGLE_OUTPUT}(=${OUTPUT_TPS}/4)" >> $OUTPUT_FILE

done

echo "Done! Output -> $OUTPUT_FILE"

在上一个脚本执行目录执行：

root@# bash parse_perf_txt1.sh 
Processing qwen3-32b_p131_n655.log...
Processing qwen3-32b_p132_n660.log...
Processing qwen3-32b_p133_n665.log...
Processing qwen3-32b_p134_n670.log...
Done! Output -> result.txt
root@# cat result.txt 
parallel   number     TTFT(ms)     TPOT(ms)   请求频率(req/s) 单卡整体吞吐(tok/s)      单卡输出吞吐(tok/s)     
131        655        3827.90      51.30      1.1122          1172.03(=4688.11/4)            569.46(=2277.85/4)            
132        660        4428.90      51.20      1.1134          1173.24(=4692.98/4)            570.05(=2280.21/4)            
133        665        5395.30      51.20      1.1136          1173.34(=4693.35/4)            570.15(=2280.61/4)            
134        670        6239.60      51.00      1.1194          1179.44(=4717.77/4)            573.14(=2292.56/4)