手把手实测！openEuler 24.03 部署 SGLang，Benchmark 数据出炉

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 ...

docker run -d \
    --name sglang-ascend-910b \
    --privileged \
    --ipc=host \
    --device=/dev/davinci2 \
    --device=/dev/davinci_manager \
    --device=/dev/devmm_svm \
    --device=/dev/hisi_hdc \
    -v /usr/local/sbin:/usr/local/sbin:ro \
    -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 /home/models:/models \
    -v $(pwd)/sglang-workspace:/workspace/home \
    -v $(pwd)/pip-cache:/root/.cache/pip \
    -w /workspace \
    -p 8002:8002 \
    --shm-size=16g \
    --ulimit memlock=-1 \
    --ulimit stack=67108864 \
    -e OMP_NUM_THREADS=8 \
    swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.0.0-910b-openeuler24.03-py3.11-devel \
    sleep infinity

# 容器内执行
python3.11 -m venv --upgrade /workspace/home/venv311
source /workspace/home/venv311/bin/activate
pip install -U pip -i https://mirrors.huaweicloud.com/repository/pypi/simple
python --version
# 预期输出：Python 3.11.x

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

printf 'torch==2.10.0+cpu\ntorch_npu==2.10.0\nnumpy==1.26.4\n' > /tmp/sglang_constraints.txt
CONSTRAINTS='-c /tmp/sglang_constraints.txt'

PIP_INDEX="https://mirrors.huaweicloud.com/repository/pypi/simple"

pip install 'setuptools<80' -i $PIP_INDEX

# torch CPU 版 wheel（从 PyTorch 官方 CPU 索引安装）
pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cpu

# torch_npu（昇腾适配层）
pip install torch_npu==2.10.0 -i $PIP_INDEX

pip install triton-ascend==3.2.1.dev20260530 \
    --extra-index-url https://mirrors.huaweicloud.com/ascend/repos/pypi/nightly \
    --trusted-host mirrors.huaweicloud.com \
    -i $PIP_INDEX

# 下载 zip（通过 GitHub 镜像代理加速）
mkdir -p /workspace/home
curl -fL --retry 3 \
    -o /workspace/home/sgl-kernel-npu.zip \
    "https://ghfast.top/https://github.com/sgl-project/sgl-kernel-npu/releases/download/2026.05.01.post3/sgl-kernel-npu-2026.05.01.post3-torch2.10.0-py311-cann9.0.0-a3-aarch64.zip"

# 安装 unzip（容器内默认无此工具）
command -v unzip >/dev/null 2>&1 || { dnf install -y unzip 2>/dev/null || yum install -y unzip; }

# 解压并安装
mkdir -p /workspace/home/sgl-kernel-npu
unzip -o /workspace/home/sgl-kernel-npu.zip -d /workspace/home/sgl-kernel-npu
cd /workspace/home/sgl-kernel-npu
pip install deep_ep*.whl sgl_kernel_npu*.whl attentions*.whl torch_memory_saver*.whl
cd -

# 验证
python -c 'import sgl_kernel_npu; print("sgl_kernel_npu import OK")'

# 克隆源码（通过镜像代理加速，浅克隆节省空间）
git clone --branch v0.5.16 --depth 1 \
    https://ghfast.top/https://github.com/sgl-project/sglang.git \
    /workspace/home/sglang
cd /workspace/home/sglang

# NPU 版安装：用 pyproject_npu.toml 覆盖 pyproject.toml
# 注意用 cp 而非 mv——保留 npu.toml 模板，二次部署时 git fetch 后仍可复用
cp -f python/pyproject_npu.toml python/pyproject.toml

# 安装（constraints 防止拉 CUDA 包）
pip install $CONSTRAINTS -e 'python[all_npu]' -i $PIP_INDEX

if pip list 2>/dev/null | grep -q '^opencv-python '; then
    pip uninstall -y opencv-python
    pip install $CONSTRAINTS opencv-python-headless -i $PIP_INDEX
fi

pip install aiohttp -i $PIP_INDEX

export HF_ENDPOINT="https://hf-mirror.com"
mkdir -p /workspace/home/sglang_bench_data
curl -fL --retry 3 \
    -o /workspace/home/sglang_bench_data/ShareGPT_V3_unfiltered_cleaned_split.json \
    "https://hf-mirror.com/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json"

python -c 'import torch, torch_npu, sglang, sgl_kernel_npu; print("torch:", torch.__version__); print("torch_npu:", torch_npu.__version__); print("sglang:", sglang.__version__)'

source /usr/local/Ascend/ascend-toolkit/set_env.sh
export ASCEND_RT_VISIBLE_DEVICES=0

sglang serve /models/Qwen3-0.6B \
    --host 0.0.0.0 \
    --port 8002 \
    --attention-backend ascend \
    --device npu \
    --tp-size 1 \x
    --mem-fraction-static 0.9 \
    --trust-remote-code \
    > /workspace/logs/sglang_server.log 2>&1 &

export HF_ENDPOINT="https://hf-mirror.com"

while ! curl -fsS http://127.0.0.1:8002/health 2>/dev/null; do
    echo "等待服务就绪..."
    sleep 10
done
echo "服务已就绪！"

curl -sS http://127.0.0.1:8002/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "/models/Qwen3-0.6B",
        "messages": [{"role": "user", "content": "请用一句话介绍华为昇腾。"}],
        "max_tokens": 64,
        "temperature": 0,
        "chat_template_kwargs": {"enable_thinking": false}
    }'

export HF_ENDPOINT="https://hf-mirror.com"
export HF_DATASETS_OFFLINE=0

# 如果硬件算力足够，适当增加输入输出的长度和并发
python -m sglang.benchmark.serving \
    --backend sglang \
    --model /models/Qwen3.5-9B \
    --dataset-name random \
    --dataset-path /workspace/home/sglang_bench_data/ShareGPT_V3_unfiltered_cleaned_split.json \
    --random-input-len  2048 \
    --random-output-len 512 \
    --num-prompts 200 \
    --max-concurrency 64 \
    --port 8002 \
    --output-file /workspace/logs/sglang_benchmark.json

# 优雅退出（SGLang 官方推荐 SIGINT）
kill -SIGINT $(pgrep -f 'sglang')

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