实战：vLLM多机多卡部署大模型

简介

两台GPU服务器，每台GPU服务器8卡GPU，400G IB网络，模型deepseek R1 671B

1. Docker容器中使用GPU

必须确保已安装并配置 NVIDIA Docker。你可以安装 nvidia-docker 来确保 GPU 驱动能够被 Docker 使用

#安装 nvidia-docker：
sudo apt-get install nvidia-docker2

#然后重启 Docker：
sudo systemctl restart docker

2.下载vllm-openai镜像

当前latest版本： v0.7.2

docker pull vllm/vllm-openai:latest

3. 启动vllm-openai容器

会启动一个node容器

3.1 启动脚本

# master  服务器1执行
sudo bash run_cluster.sh \
                vllm/vllm-openai \  # 镜像名称
                172.16.3.38 \  # head服务器IP
                --head \   # 代表head
                /home/llm/ai/model \  # huggingface 模型路径
                -v /home/llm/ai/model:/home/llm/ai/model \   # 宿主机本地模型映射到容器内
                -e GLOO_SOCKET_IFNAME=enp4s0 \  # 服务器IP对应的网卡名称
                -e NCCL_SOCKET_IFNAME=enp4s0 \  # 服务器IP对应的网卡名称
                -e VLLM_HOST_IP=172.16.3.38

# worker  服务器2执行
bash run_cluster.sh \
                vllm/vllm-openai \
                172.16.3.38 \
                --worker \  # 代表 从服务器
                /home/llm/ai/model \
                -v /home/llm/ai/model:/home/llm/ai/model \
                -e GLOO_SOCKET_IFNAME=enp5s0 \
                -e NCCL_SOCKET_IFNAME=enp5s0 \
                -e VLLM_HOST_IP=172.16.3.37

run_cluster.sh

#!/bin/bash

# Check for minimum number of required arguments
if [ $# -lt 4 ]; then
    echo "Usage: $0 docker_image head_node_address --head|--worker path_to_hf_home [additional_args...]"
    exit 1
fi

# Assign the first three arguments and shift them away
DOCKER_IMAGE="$1"
HEAD_NODE_ADDRESS="$2"
NODE_TYPE="$3"  # Should be --head or --worker
PATH_TO_HF_HOME="$4"
shift 4

# Additional arguments are passed directly to the Docker command
ADDITIONAL_ARGS=("$@")

# Validate node type
if [ "${NODE_TYPE}" != "--head" ] && [ "${NODE_TYPE}" != "--worker" ]; then
    echo "Error: Node type must be --head or --worker"
    exit 1
fi

# Define a function to cleanup on EXIT signal
cleanup() {
    docker stop node
    docker rm node
}
trap cleanup EXIT

# Command setup for head or worker node
RAY_START_CMD="ray start --block"
if [ "${NODE_TYPE}" == "--head" ]; then
    RAY_START_CMD+=" --head --port=6379"
else
    RAY_START_CMD+=" --address=${HEAD_NODE_ADDRESS}:6379"
fi

# Run the docker command with the user specified parameters and additional arguments
docker run \
    --entrypoint /bin/bash \
    --network host \
    --name node \
    --shm-size 10.24g \
    --gpus all \
    -v "${PATH_TO_HF_HOME}:/root/.cache/huggingface" \
    "${ADDITIONAL_ARGS[@]}" \
    "${DOCKER_IMAGE}" -c "${RAY_START_CMD}"

4. 进入到容器内部加载模型

# 进入容器内部
docker exec -it node /bin/bash

# tensor-parallel-size 每台服务器显卡数量
# pipeline-parallel-size 服务器数量
vllm serve /home/llm/ai/model/Qwen/Qwen1___5-1___8B-Chat --port 8080 --tensor-parallel-size 1 --pipeline-parallel-size 2  --dtype float16

5. 查看集群状态

• 容器内执行 可以看到两个节点
  ray status