15分钟在Rocky Linux 9.x和公有云英伟达环境下部署 vLLM 0.16.0

以下是一套经过测试的脚本。

Rocky linux 采用9.5，镜像制作中安装的包为 Server with GUI.

echo "mkdir -p /software/vllm"
cd /software/vllm

mkdir -p /yumrepo
cp vllm-rocky-9.5.tar /yumrepo
cd /yumrepo
tar -xvf vllm-rocky-9.5.tar

cd /etc/yum.repos.d
mv * /mnt

cat <<EOF > /etc/yum.repos.d/vllm.repo
[vllm]
name=vllm
baseurl=file:///yumrepo/vllm
enabled=1
gpgcheck=0

EOF

cd /yumrepo/vllm
dnf install *.rpm -y

rpm -qa | grep nvidia-driver
nvidia-smi

sudo systemctl start docker
systemctl --no-pager status docker
sudo systemctl enable docker
docker version

echo "reboot"

sudo grubby --args="nouveau.modeset=0 rd.driver.blacklist=nouveau" --update-kernel=ALL
sudo dracut --force
reboot

nvidia-smi

cd /software/vllm
cp model-vllm-Qwen2.5-0.5B-Instruct.tar /
cd /
tar -xvf model-vllm-Qwen2.5-0.5B-Instruct.tar
ls -l /model/vllm/
cat /model/vllm/config.json | grep -E

cd /software/vllm
docker load -i vllm-openai-0.16.0.tar

docker run --gpus all \
-v /model/vllm:/model/vllm \
-p 8000:8000 \
vllm/vllm-openai:latest \
--model /model/vllm \
--dtype auto \
--enforce-eager \
--max-model-len 2048