ProDiff/run_a40_turbo.sh

#!/bin/bash

# ===================================
# A40 GPU超速优化训练脚本
# 目标：充分利用A40的46GB显存和300W功耗
# ===================================

# 配置参数
EXP_NAME="trajectory_a40_turbo_optimized"
DEVICE_ID=0
SAMPLING_TYPE="ddim"
LOG_DIR="./training_logs"

# 创建日志目录
mkdir -p ${LOG_DIR}

# 获取当前时间戳用于日志文件命名
TIMESTAMP=$(date +"%Y%m%d_%H%M%S")
LOG_FILE="${LOG_DIR}/training_${EXP_NAME}_${TIMESTAMP}.log"

# 显示启动信息
echo "====================================="
echo "ProDiff A40超速训练启动"
echo "====================================="
echo "实验名称: ${EXP_NAME}"
echo "GPU设备: ${DEVICE_ID}"
echo "采样类型: ${SAMPLING_TYPE}"
echo "Batch Size: 2560 (增加67%)"
echo "混合精度: 启用FP16"
echo "数据加载线程: 16"
echo "日志文件: ${LOG_FILE}"
echo "启动时间: $(date)"
echo "====================================="

# 检查GPU状态
echo "检查GPU状态..."
nvidia-smi

# 设置环境变量以优化A40性能
export CUDA_VISIBLE_DEVICES=${DEVICE_ID}
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:256  # 增加内存块大小
export OMP_NUM_THREADS=16  # 增加CPU线程数
export CUDA_LAUNCH_BLOCKING=0  # 异步执行
export PYTHONUNBUFFERED=1  # 实时输出

# A40专用优化设置
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1

# 使用nohup后台运行，同时输出到日志文件和终端
echo "开始超速训练..."
nohup python -u main.py \
    --exp_name ${EXP_NAME} \
    --device_id ${DEVICE_ID} \
    --sampling_type ${SAMPLING_TYPE} \
    --seed 42 \
    > ${LOG_FILE} 2>&1 &

# 获取进程ID
PID=$!
echo "训练进程PID: ${PID}"
echo "日志文件: ${LOG_FILE}"

# 保存PID到文件，方便后续管理
echo ${PID} > "${LOG_DIR}/training_${EXP_NAME}.pid"

echo ""
echo "====================================="
echo "超速训练已在后台启动！"
echo "====================================="
echo "预期性能提升:"
echo "  - Batch Size: 1536 → 2560 (+67%)"
echo "  - 混合精度训练: FP32 → FP16 (2x内存效率)"
echo "  - 数据加载: 12 → 16 线程 (+33%)"
echo "  - 预计速度提升: 60-80%"
echo ""
echo "监控命令:"
echo "  查看日志: tail -f ${LOG_FILE}"
echo "  查看进程: ps aux | grep ${PID}"
echo "  停止训练: kill ${PID}"
echo "  或者: kill \$(cat ${LOG_DIR}/training_${EXP_NAME}.pid)"
echo ""
echo "GPU监控:"
echo "  nvidia-smi"
echo "  watch -n 1 nvidia-smi"
echo "====================================="