run.sh

export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export ROCBLAS_COMPUTETYPE_FP16R=0
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export NCCL_ALGO=Ring
export NCCL_LAUNCH_MODE=GROUP
export NCCL_NCHANNELS_PER_PEER=16
export NCCL_MAX_NCHANNELS=16
export NCCL_MIN_NCHANNELS=16
export NCCL_IB_TIMEOUT=22
export CUDA_DEVICE_MAX_CONNECTIONS=1
export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
export NCCL_P2P_LEVEL=SYS
export NCCL_NET_GDR_LEVEL=7
export NCCL_NET_GDR_READ=1
export RCCL_SDMA_COPY_ENABLE=0
export SENDRECV_STREAM_WITH_COMPUTE=1
export NCCL_TOPO_FILE="/workspace/test/topo.xml"
export LD_LIBRARY_PATH=/usr/local/lib/python3.10/site-packages/torch/lib/:$LD_LIBRARY_PATH
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export VLLM_PCIE_USE_CUSTOM_ALLREDUCE=1
export VLLM_NUMA_BIND=1
export VLLM_RANK0_NUMA=0
export VLLM_RANK1_NUMA=1
export VLLM_RANK2_NUMA=2
export VLLM_RANK3_NUMA=3
export VLLM_RANK4_NUMA=4
export VLLM_RANK5_NUMA=5
export VLLM_RANK6_NUMA=6
export VLLM_RANK7_NUMA=7

export VLLM_RPC_TIMEOUT=100000



#!/bin/bash

# 模型配置文件路径
MODELS_CONFIG="/workspace/test/models-to-test.cfg"

# 结果目录
RESULTS_DIR="/workspace/test/results"

# 读取配置文件，跳过注释和空行
while IFS= read -r line || [[ -n "$line" ]]; do
    # 跳过注释行和空行
    if [[ "$line" =~ ^# ]] || [[ -z "$line" ]]; then
        continue
    fi
    
    # 解析配置行
    IFS=';' read -ra CONFIG <<< "$line"
    
    model_name="${CONFIG[0]}"
    model_path="${CONFIG[1]}"
    tp="${CONFIG[2]}"
    batch="${CONFIG[3]//,/ }"        # 将逗号替换为空格
    prompt_tokens="${CONFIG[4]//,/ }"
    completion_tokens="${CONFIG[5]//,/ }"
    dtype="${CONFIG[6]}"
    max_model_len="${CONFIG[7]}"
    gpu_memory_utilization="${CONFIG[8]}"
    
    echo "开始测试模型: $model_name"
    echo "模型路径: $model_path"
    echo "参数配置:"
    echo "  tensor_parallel_size: $tp"
    echo "  batch_sizes: $batch"
    echo "  prompt_tokens: $prompt_tokens"
    echo "  completion_tokens: $completion_tokens"
    echo "  dtype: $dtype"
    echo "  max_model_len: $max_model_len"
    echo "  gpu_memory_utilization: $gpu_memory_utilization"
    
    # 创建模型专属结果目录
    model_result_dir="${RESULTS_DIR}/${model_name}"
    mkdir -p "$model_result_dir"
    
    # 运行基准测试
    python /workspace/test/benchmark_throughput_0.7.2.py \
        --model "$model_path" \
        --tensor-parallel-size "$tp" \
        --num-prompts $batch \
        --input-len $prompt_tokens \
        --output-len $completion_tokens \
        --dtype "$dtype" \
        --trust-remote-code \
        --max-model-len "$max_model_len" \
        --gpu-memory-utilization "$gpu_memory_utilization" \
        --output-json "${model_result_dir}/${model_name}_tp${tp}.txt" \
        2>&1 | tee "${model_result_dir}/${model_name}_tp${tp}.log"
    
    echo "完成测试模型: $model_name"
    echo "结果保存在: $model_result_dir"
    echo "----------------------------------------"
done < "$MODELS_CONFIG"