一文实测！openEuler 24.03 搭建 vLLM, Benchmark数据直观出炉

docker pull swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11-devel

# 查看 NPU 设备
ls /dev/davinci*
# 预期输出：/dev/davinci0  /dev/davinci2  /dev/davinci_manager ...

# 查看 NPU 状态
npu-smi info

# 以 Qwen3-0.6B 为例
ls /home/models/Qwen3-0.6B/
# 预期包含：config.json  tokenizer.json  model*.safetensors ...

IMAGE="swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11-devel"
CONTAINER_NAME="vllm-ascend-910b"

NPU_CARDS=("davinci2")
TP_SIZE=1

MODEL_HOST_DIR="/home/models"
MODEL_CONTAINER_DIR="/models"
MODEL_NAME="Qwen3.5-9B"

HOST_PORT=8000
CONTAINER_PORT=8000

SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
LOG_HOST_DIR="${SCRIPT_DIR}/logs"
LOG_CONTAINER_DIR="/workspace/logs"

PIP_CACHE_HOST_DIR="${SCRIPT_DIR}/pip-cache"
PIP_CACHE_CONTAINER_DIR="/root/.cache/pip"

docker run -d \
    --name "${CONTAINER_NAME}" \
    "${device_args[@]}" \
    --device=/dev/davinci_manager \
    --device=/dev/devmm_svm \
    --device=/dev/hisi_hdc \
    -v /usr/local/dcmi:/usr/local/dcmi:ro \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi:ro \
    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro \
    -v /etc/ascend_install.info:/etc/ascend_install.info:ro \
    -v "${MODEL_HOST_DIR}:${MODEL_CONTAINER_DIR}" \
    -v "${LOG_HOST_DIR}:${LOG_CONTAINER_DIR}" \
    -v "${PIP_CACHE_HOST_DIR}:${PIP_CACHE_CONTAINER_DIR}" \
    -v /usr/local/Ascend/firmware:/usr/local/Ascend/firmware \
    -w /workspace \
    -p "${HOST_PORT}:${CONTAINER_PORT}" \
    --shm-size=16g \
    --ulimit memlock=-1 \
    --ulimit stack=67108864 \
    -e OMP_NUM_THREADS=8 \
    "${IMAGE}" \
    sleep infinity >/dev/null

log "容器 ${CONTAINER_NAME} 已启动"

docker exec -it vllm-ascend-910b bash
npu-smi info

sed -i -e 's|https://repo.openeuler.org/|https://mirrors.huaweicloud.com/openeuler/|' \
       -e '/^metalink=/s/^/#/' /etc/yum.repos.d/openEuler.repo

pip install -U pip -i https://mirrors.huaweicloud.com/repository/pypi/simple

pip install torch==2.9.0 torch_npu==2.9.0 \
    -i https://mirrors.huaweicloud.com/repository/pypi/simple

pip install vllm==0.18.0 \
    -i https://mirrors.huaweicloud.com/repository/pypi/simple

pip install vllm-ascend==0.18.0 \
    -i https://mirrors.huaweicloud.com/repository/pypi/simple

pip install --no-deps triton-ascend==3.2.1 \
    -i https://mirrors.huaweicloud.com/repository/pypi/simple \
    --extra-index-url https://mirrors.huaweicloud.com/ascend/repos/pypi

# 找到 npu_utils.cpp 路径
NPU_UTILS_CPP=$(python -c "import triton.backends.ascend as m, os; print(os.path.join(os.path.dirname(m.__file__), 'npu_utils.cpp'))")

# 检查并补丁
if grep -q "RT_LIMIT_TYPE_SIMT_WARP_STACK_SIZE" "$NPU_UTILS_CPP"; then
    cp "$NPU_UTILS_CPP" "$NPU_UTILS_CPP.orig"
    sed -i "/RT_LIMIT_TYPE_SIMT_WARP_STACK_SIZE/d; /RT_LIMIT_TYPE_SIMT_DVG_WARP_STACK_SIZE/d" "$NPU_UTILS_CPP"
    echo "已补丁: $NPU_UTILS_CPP"
else
    echo "无需补丁"
fi

# 验证 triton NPU 驱动可正常初始化
python -c "from triton.runtime import driver; print('triton npu driver OK:', driver.active.get_current_target())"

pip install aiohttp -i https://mirrors.huaweicloud.com/repository/pypi/simple

python -c 'import torch, torch_npu, vllm; print("torch:", torch.__version__); print("torch_npu:", torch_npu.__version__); print("vllm:", vllm.__version__)'
pip show vllm-ascend | grep -E '^(Name|Version)'

vllm serve /models/Qwen3.5-9B \
    --host 0.0.0.0 \
    --port 8000 \
    --tensor-parallel-size 1 \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.9 \
    --trust-remote-code \
    > /workspace/logs/vllm_server.log 2>&1 &

tail -f /workspace/logs/vllm_server.log

# 循环检测，直到返回 200
while ! curl -fsS http://127.0.0.1:8000/v1/models 2>/dev/null; do
    echo "等待服务就绪..."
    sleep 10
done
echo "服务已就绪！"
curl -s http://127.0.0.1:8000/v1/models | python -m json.tool

curl -sS http://127.0.0.1:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "/models/Qwen3.5-9B",
        "messages": [{"role": "user", "content": "请用一句话介绍华为昇腾。/no_think"}],
        "max_tokens": 2000,
        "temperature": 0,
        "chat_template_kwargs":{"enable_thinking":false}
    }'

# 可自行根据系统资源情况调整参数
vllm bench serve \
    --backend vllm \
    --model /models/Qwen3.5-9B \
    --endpoint /v1/completions \
    --host 127.0.0.1 \
    --port 8000 \
    --dataset-name random \
    --num-prompts 200 \
    --random-input-len  4096 \
    --random-output-len 1024 \
    --max-concurrency 32 \
    --ignore-eos \
    --save-result \
    --result-dir /workspace/logs \
    --result-filename vllm_benchmark.json

# 停止推理服务
pkill -f 'vllm serve'

# 退出容器后删除容器
exit
docker stop vllm-ascend-910b
docker rm vllm-ascend-910b