droid/scripts/run_preprocessing_egodex.sh

#!/bin/bash

# ================= 配置区域 =================
DATA_ROOT="/mnt/world_foundational_model/gkz/data/EgoDex"
OUTPUT_DIR="/mnt/kevin/data/egodex_preprocessed/data"
SCRIPT_PATH="/mnt/kevin/code/wmrl/Dual-Dynamics-Models/DROID-main/scripts/preprocess_egodex.py"
LOG_DIR="/mnt/kevin/data/egodex_preprocessed/logs"
PYTHON="/mnt/kevin/envs/miniconda3/envs/atm_ati_vdm_droid/bin/python"

# ================= 分布式节点配置 =================
# 从环境变量读取节点信息。如果未设置，默认是单机模式
# NODE_RANK: 当前机器是第几台 (0, 1, 2...)
# NNODES: 总共多少台机器
NODE_RANK=${RANK:-"0"}
NNODES=${WORLD_SIZE:-"1"}

mkdir -p "$LOG_DIR"

# 自动获取当前机器的 GPU 数量
GPUS_PER_NODE=$(nvidia-smi --query-gpu=name --format=csv,noheader | wc -l)

# 计算整个集群的总进程数 (用于数据切分)
TOTAL_WORLD_SIZE=$((NNODES * GPUS_PER_NODE))

echo "=========================================================="
echo "Distributed Processing Status"
echo "Node Rank (This Machine) : $NODE_RANK / $NNODES"
echo "GPUs on this Node        : $GPUS_PER_NODE"
echo "Total Workers (Cluster)  : $TOTAL_WORLD_SIZE"
echo "=========================================================="

# 循环启动当前机器上的每个 GPU
for (( i=0; i<GPUS_PER_NODE; i++ ))
do
    # 计算全局唯一的 Rank ID
    # Global Rank = (当前节点号 * 单节点卡数) + 本地卡号
    GLOBAL_RANK=$((NODE_RANK * GPUS_PER_NODE + i))

    echo "Launching worker: Local GPU $i | Global Rank $GLOBAL_RANK"

    nohup $PYTHON "$SCRIPT_PATH" \
      --data_root "$DATA_ROOT" \
      --output_dir "$OUTPUT_DIR" \
      --gpu_id "$i" \
      --world_size "$TOTAL_WORLD_SIZE" \
      --global_rank "$GLOBAL_RANK" \
      > "$LOG_DIR/worker_global_${GLOBAL_RANK}.log" 2>&1 &

    # 错峰启动，避免 CPU 瞬时负载过高
    sleep 2
done

echo "All workers on Node $NODE_RANK launched!"
echo "Log sample: tail -f $LOG_DIR/worker_global_$((NODE_RANK * GPUS_PER_NODE)).log"
wait