vLLM 推理

docker ps

sudo usermod -aG docker $USER
newgrp docker

export LATEST_VLLM_VERSION=<latest_container_version>
## example
## export LATEST_VLLM_VERSION=26.02-py3

export HF_MODEL_HANDLE=<HF_HANDLE>
## example
## export HF_MODEL_HANDLE=openai/gpt-oss-20b

docker pull nvcr.io/nvidia/vllm:${LATEST_VLLM_VERSION}

docker pull vllm/vllm-openai:gemma4-cu130

docker run -it --gpus all -p 8000:8000 \
nvcr.io/nvidia/vllm:${LATEST_VLLM_VERSION} \
vllm serve ${HF_MODEL_HANDLE}

docker run -it --gpus all -p 8000:8000 \
vllm/vllm-openai:gemma4-cu130 ${HF_MODEL_HANDLE}

curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
    "model": "'"${HF_MODEL_HANDLE}"'",
    "messages": [{"role": "user", "content": "12*17"}],
    "max_tokens": 500
}'

docker rm $(docker ps -aq --filter ancestor=nvcr.io/nvidia/vllm:${LATEST_VLLM_VERSION})
docker rmi nvcr.io/nvidia/vllm

## Download on both nodes
wget https://raw.githubusercontent.com/vllm-project/vllm/refs/heads/main/examples/online_serving/run_cluster.sh
chmod +x run_cluster.sh

sudo groupadd docker
sudo usermod -aG docker $USER
newgrp docker

docker pull nvcr.io/nvidia/vllm:25.11-py3
export VLLM_IMAGE=nvcr.io/nvidia/vllm:25.11-py3

## On Node 1, start head node

## Get the IP address of the high-speed interface
## Use the interface that shows "(Up)" from ibdev2netdev (enp1s0f0np0 or enp1s0f1np1)
export MN_IF_NAME=enp1s0f1np1
export VLLM_HOST_IP=$(ip -4 addr show $MN_IF_NAME | grep -oP '(?<=inet\s)\d+(\.\d+){3}')

echo "Using interface $MN_IF_NAME with IP $VLLM_HOST_IP"

bash run_cluster.sh $VLLM_IMAGE $VLLM_HOST_IP --head ~/.cache/huggingface \
  -e VLLM_HOST_IP=$VLLM_HOST_IP \
  -e UCX_NET_DEVICES=$MN_IF_NAME \
  -e NCCL_SOCKET_IFNAME=$MN_IF_NAME \
  -e OMPI_MCA_btl_tcp_if_include=$MN_IF_NAME \
  -e GLOO_SOCKET_IFNAME=$MN_IF_NAME \
  -e TP_SOCKET_IFNAME=$MN_IF_NAME \
  -e RAY_memory_monitor_refresh_ms=0 \
  -e MASTER_ADDR=$VLLM_HOST_IP

## On Node 2, join as worker

## Set the interface name (same as Node 1)
export MN_IF_NAME=enp1s0f1np1

## Get Node 2's own IP address
export VLLM_HOST_IP=$(ip -4 addr show $MN_IF_NAME | grep -oP '(?<=inet\s)\d+(\.\d+){3}')

## IMPORTANT: Set HEAD_NODE_IP to Node 1's IP address
## You must get this value from Node 1 (run: echo $VLLM_HOST_IP on Node 1)
export HEAD_NODE_IP=<NODE_1_IP_ADDRESS>

echo "Worker IP: $VLLM_HOST_IP, connecting to head node at: $HEAD_NODE_IP"

bash run_cluster.sh $VLLM_IMAGE $HEAD_NODE_IP --worker ~/.cache/huggingface \
  -e VLLM_HOST_IP=$VLLM_HOST_IP \
  -e UCX_NET_DEVICES=$MN_IF_NAME \
  -e NCCL_SOCKET_IFNAME=$MN_IF_NAME \
  -e OMPI_MCA_btl_tcp_if_include=$MN_IF_NAME \
  -e GLOO_SOCKET_IFNAME=$MN_IF_NAME \
  -e TP_SOCKET_IFNAME=$MN_IF_NAME \
  -e RAY_memory_monitor_refresh_ms=0 \
  -e MASTER_ADDR=$HEAD_NODE_IP

## On Node 1 (head node)
## Find the vLLM container name (it will be node-<random_number>)
export VLLM_CONTAINER=$(docker ps --format '{{.Names}}' | grep -E '^node-[0-9]+$')
echo "Found container: $VLLM_CONTAINER"

docker exec $VLLM_CONTAINER ray status

## From within the same container where `ray status` ran, run the following
hf auth login
hf download meta-llama/Llama-3.3-70B-Instruct

## On Node 1, enter container and start server
export VLLM_CONTAINER=$(docker ps --format '{{.Names}}' | grep -E '^node-[0-9]+$')
docker exec -it $VLLM_CONTAINER /bin/bash -c '
  vllm serve meta-llama/Llama-3.3-70B-Instruct \
    --tensor-parallel-size 2 --max_model_len 2048'

## Test from Node 1 or external client
curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-llama/Llama-3.3-70B-Instruct",
    "prompt": "Write a haiku about a GPU",
    "max_tokens": 32,
    "temperature": 0.7
  }'

## On Node 1, download quantized model
huggingface-cli download hugging-quants/Meta-Llama-3.1-405B-Instruct-AWQ-INT4

## On Node 1, launch with restricted parameters
export VLLM_CONTAINER=$(docker ps --format '{{.Names}}' | grep -E '^node-[0-9]+$')
docker exec -it $VLLM_CONTAINER /bin/bash -c '
  vllm serve hugging-quants/Meta-Llama-3.1-405B-Instruct-AWQ-INT4 \
    --tensor-parallel-size 2 --max-model-len 64 --gpu-memory-utilization 0.9 \
    --max-num-seqs 1 --max_num_batched_tokens 64'

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hugging-quants/Meta-Llama-3.1-405B-Instruct-AWQ-INT4",
    "prompt": "Write a haiku about a GPU",
    "max_tokens": 32,
    "temperature": 0.7
  }'

## Check Ray cluster health
export VLLM_CONTAINER=$(docker ps --format '{{.Names}}' | grep -E '^node-[0-9]+$')
docker exec $VLLM_CONTAINER ray status

## Verify server health endpoint
curl http://192.168.100.10:8000/health

## Monitor GPU utilization on both nodes
nvidia-smi
export VLLM_CONTAINER=$(docker ps --format '{{.Names}}' | grep -E '^node-[0-9]+$')
docker exec $VLLM_CONTAINER nvidia-smi --query-gpu=memory.used,memory.total --format=csv

## Ray dashboard available at:
http://<head-node-ip>:8265

## Consider implementing for production:
## - Health checks and automatic restarts
## - Log rotation for long-running services
## - Persistent model caching across restarts
## - Alternative quantization methods (FP8, INT4)

## Download on all nodes
wget https://raw.githubusercontent.com/vllm-project/vllm/refs/heads/main/examples/online_serving/run_cluster.sh
chmod +x run_cluster.sh

sudo groupadd docker
sudo usermod -aG docker $USER
newgrp docker

docker pull nvcr.io/nvidia/vllm:26.02-py3
export VLLM_IMAGE=nvcr.io/nvidia/vllm:26.02-py3

## On Node 1, start head node

## Get the IP address of the high-speed interface
## Use the interface that shows "(Up)" from ibdev2netdev (enp1s0f0np0 or enp1s0f1np1)
export MN_IF_NAME=enp1s0f1np1
export VLLM_HOST_IP=$(ip -4 addr show $MN_IF_NAME | grep -oP '(?<=inet\s)\d+(\.\d+){3}')

echo "Using interface $MN_IF_NAME with IP $VLLM_HOST_IP"

bash run_cluster.sh $VLLM_IMAGE $VLLM_HOST_IP --head ~/.cache/huggingface \
  -e VLLM_HOST_IP=$VLLM_HOST_IP \
  -e UCX_NET_DEVICES=$MN_IF_NAME \
  -e NCCL_SOCKET_IFNAME=$MN_IF_NAME \
  -e OMPI_MCA_btl_tcp_if_include=$MN_IF_NAME \
  -e GLOO_SOCKET_IFNAME=$MN_IF_NAME \
  -e TP_SOCKET_IFNAME=$MN_IF_NAME \
  -e RAY_memory_monitor_refresh_ms=0 \
  -e MASTER_ADDR=$VLLM_HOST_IP

## On other Nodes, join as workers

## Set the interface name (same as Node 1)
export MN_IF_NAME=enp1s0f1np1

## Get Node's own IP address
export VLLM_HOST_IP=$(ip -4 addr show $MN_IF_NAME | grep -oP '(?<=inet\s)\d+(\.\d+){3}')

## IMPORTANT: Set HEAD_NODE_IP to Node 1's IP address
## You must get this value from Node 1 (run: echo $VLLM_HOST_IP on Node 1)
export HEAD_NODE_IP=<NODE_1_IP_ADDRESS>

echo "Worker IP: $VLLM_HOST_IP, connecting to head node at: $HEAD_NODE_IP"

bash run_cluster.sh $VLLM_IMAGE $HEAD_NODE_IP --worker ~/.cache/huggingface \
  -e VLLM_HOST_IP=$VLLM_HOST_IP \
  -e UCX_NET_DEVICES=$MN_IF_NAME \
  -e NCCL_SOCKET_IFNAME=$MN_IF_NAME \
  -e OMPI_MCA_btl_tcp_if_include=$MN_IF_NAME \
  -e GLOO_SOCKET_IFNAME=$MN_IF_NAME \
  -e TP_SOCKET_IFNAME=$MN_IF_NAME \
  -e RAY_memory_monitor_refresh_ms=0 \
  -e MASTER_ADDR=$HEAD_NODE_IP

## On Node 1 (head node)
## Find the vLLM container name (it will be node-<random_number>)
export VLLM_CONTAINER=$(docker ps --format '{{.Names}}' | grep -E '^node-[0-9]+$')
echo "Found container: $VLLM_CONTAINER"

docker exec $VLLM_CONTAINER ray status

## On all nodes, from within the docker containers created in previous steps, run the following
hf auth login
hf download MiniMaxAI/MiniMax-M2.5

## On Node 1, enter container and start server
## Assuming that you run on a 4 node cluster, set --tensor-parallel-size as 4
export VLLM_CONTAINER=$(docker ps --format '{{.Names}}' | grep -E '^node-[0-9]+$')
docker exec -it $VLLM_CONTAINER /bin/bash -c '
  vllm serve MiniMaxAI/MiniMax-M2.5 \
    --tensor-parallel-size 4 --max-model-len 129000 --max-num-seqs 4 --trust-remote-code'

## Test from Node 1 or external client.
## If testing with external client change localhost to the Node 1 Mgmt IP address.
curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMaxAI/MiniMax-M2.5",
    "prompt": "Write a haiku about a GPU",
    "max_tokens": 32,
    "temperature": 0.7
  }'

## Check Ray cluster health
export VLLM_CONTAINER=$(docker ps --format '{{.Names}}' | grep -E '^node-[0-9]+$')
docker exec $VLLM_CONTAINER ray status

## Verify server health endpoint on Node 1
curl http://localhost:8000/health

## Monitor GPU utilization on all nodes
nvidia-smi
export VLLM_CONTAINER=$(docker ps --format '{{.Names}}' | grep -E '^node-[0-9]+$')
docker exec $VLLM_CONTAINER nvidia-smi --query-gpu=memory.used,memory.total --format=csv

## Ray dashboard available at:
http://<head-node-ip>:8265

## Consider implementing for production:
## - Health checks and automatic restarts
## - Log rotation for long-running services
## - Persistent model caching across restarts
## - Other models which can fit on the cluster with different quantization methods (FP8, NVFP4)

sudo sh -c 'sync; echo 3 > /proc/sys/vm/drop_caches'