项目现在更新了不少，来一把训练。目的是训练qwen2.5vl 7b，fork项目在此：https://github.com/xxddccaa/EasyR1QwenVL

dockerfile：

dockerfile
展开代码
FROM hiyouga/verl:ngc-th2.7.1-cu12.6-vllm0.10.0
# Install Chinese language support and fonts
RUN apt-get update && apt-get install -y \
    locales \
    fonts-noto-cjk \
    language-pack-zh-hans \
    openssh-server \
    zip \
    unzip \
    tree \
    vim \
    tzdata \
    apt-utils \
    htop \
    tmux \
    curl \
    wget \
    git \
    net-tools \
    libibverbs1 \
    libibverbs-dev \
    libgl1-mesa-glx libglx-mesa0 \
    && rm -rf /var/lib/apt/lists/*
WORKDIR /workplace
RUN git clone https://github.com/xxddccaa/EasyR1QwenVL.git
WORKDIR /workplace/EasyR1
RUN pip install -e . -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
RUN pip install swanlab -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple

bash
展开代码
docker build -f Dockerfile.qwen2vl . -t hub.i.basemind.com/g-xiedong-fine/easyr1:v1

js
展开代码
#!/bin/bash
cp /config/qwen2_5_vl_7b_grpo_train.sh /workplace/EasyR1/examples/
# 参数设置（可以改成从环境变量读取）
TOTAL_NODES=2                # 总节点数
GPU_PER_NODE=8               # 每台机器的 GPU 数量
#RAY_HEAD_IP="${MASTER_ADDR}"  # 主节点 IP
RAY_PORT="${MASTER_PORT}"              # Ray 默认端口
NODE_RANK="${RANK}"          # 当前节点 rank（0 是主节点）
RAY_HEAD_IP=$(getent hosts "${MASTER_ADDR}" | awk '{print $1}')

echo "RAY_HEAD_IP: $RAY_HEAD_IP"
echo "NODE_RANK: $NODE_RANK"
echo "TOTAL_NODES: $TOTAL_NODES"
echo "GPU_PER_NODE: $GPU_PER_NODE"
echo "RAY_PORT: $RAY_PORT"

export SWANLAB_API_KEY=xxxxx           # 设置在线跟踪模式API
export SWANLAB_LOG_DIR=/swanlab_log    # 设置本地日志存储路径
export SWANLAB_MODE=cloud     # 包含四种模式：cloud云端跟踪模式（默认）、cloud-only仅云端跟踪本地不保存文件、local本地跟踪模式、disabled完全不记录用于debug


# 检查是否是主节点
if [ "$NODE_RANK" -eq 0 ]; then
    echo "Starting Ray HEAD node on $(hostname -I | awk '{print $1}')"
    ray start --head --port="$RAY_PORT" --dashboard-host=0.0.0.0 --num-gpus="$GPU_PER_NODE"
    sleep 20
else
    sleep 10
    echo "Starting Ray WORKER node #$NODE_RANK, connecting to $RAY_HEAD_IP:$RAY_PORT"
    ray start --address="$RAY_HEAD_IP:$RAY_PORT" --num-gpus="$GPU_PER_NODE"
fi

# 等待所有节点加入（主节点检查）
if [ "$NODE_RANK" -eq 0 ]; then
    echo "Waiting for all nodes to join..."
    while true; do
        echo "Checking connected nodes..."
        LIVE_NODES=$(ray status | grep -c " 1 node_")
        echo "Current connected nodes: $LIVE_NODES / $TOTAL_NODES"
        
        # 打印ray status的完整输出以便调试
        echo "Ray cluster status:"
        ray status
        
        if [ "$LIVE_NODES" -eq "$TOTAL_NODES" ]; then
            echo "All $TOTAL_NODES nodes connected successfully!"
            break
        fi
        echo "Waiting 5 seconds before next check..."
        sleep 5
    done

    # 在主节点上启动训练脚本
    echo "Launching training script..."
    echo "Setting up SwanLab environment variables..."
    export SWANLAB_API_KEY=pM7Xvs5OS2EeXPO5gKXfJ           # 设置在线跟踪模式API
    export SWANLAB_LOG_DIR=/swanlab_log    # 设置本地日志存储路径
    export SWANLAB_MODE=cloud     # 包含四种模式：cloud云端跟踪模式（默认）、cloud-only仅云端跟踪本地不保存文件、local本地跟踪模式、disabled完全不记录用于debug
    echo "Starting training with qwen2_5_vl_7b_grpo_train.sh..."
    cd /workplace/EasyR1 && bash examples/qwen2_5_vl_7b_grpo_train.sh
fi

# 保持 Ray 运行（防止脚本退出）
sleep infinity