2025-10-13 13:22:50 +00:00
# Install and Use vLLM for Inference
2025-10-03 20:46:11 +00:00
> Use a container or build vLLM from source for Spark
## Table of Contents
- [Overview ](#overview )
2025-10-06 13:35:52 +00:00
- [Instructions ](#instructions )
2025-10-03 20:46:11 +00:00
- [Run on two Sparks ](#run-on-two-sparks )
2025-10-10 20:39:52 +00:00
- [Step 11. (Optional) Launch 405B inference server ](#step-11-optional-launch-405b-inference-server )
2025-10-10 00:11:49 +00:00
- [Troubleshooting ](#troubleshooting )
2025-10-03 20:46:11 +00:00
---
## Overview
2025-10-08 20:25:52 +00:00
## Basic idea
2025-10-10 20:39:52 +00:00
vLLM is an inference engine designed to run large language models efficiently. The key idea is **maximizing throughput and minimizing memory waste** when serving LLMs.
2025-10-08 20:25:52 +00:00
2025-10-10 20:39:52 +00:00
- It uses a memory-efficient attention algoritm called **PagedAttention** to handle long sequences without running out of GPU memory.
- New requests can be added to a batch already in process through **continuous batching** to keep GPUs fully utilized.
- It has an **OpenAI-compatible API** so applications built for the OpenAI API can switch to a vLLM backend with little or no modification.
2025-10-08 20:25:52 +00:00
2025-10-03 20:46:11 +00:00
## What you'll accomplish
2025-10-10 20:39:52 +00:00
You'll set up vLLM high-throughput LLM serving on DGX Spark with Blackwell architecture,
either using a pre-built Docker container or building from source with custom LLVM/Triton
2025-10-03 20:46:11 +00:00
support for ARM64.
## What to know before starting
- Experience building and configuring containers with Docker
2025-10-10 20:39:52 +00:00
- Familiarity with CUDA toolkit installation and version management
2025-10-03 20:46:11 +00:00
- Understanding of Python virtual environments and package management
- Knowledge of building software from source using CMake and Ninja
- Experience with Git version control and patch management
## Prerequisites
2025-10-06 13:35:52 +00:00
- DGX Spark device with ARM64 processor and Blackwell GPU architecture
2025-10-10 20:39:52 +00:00
- CUDA 13.0 toolkit installed: `nvcc --version` shows CUDA toolkit version.
2025-10-06 13:35:52 +00:00
- Docker installed and configured: `docker --version` succeeds
- NVIDIA Container Toolkit installed
- Python 3.12 available: `python3.12 --version` succeeds
- Git installed: `git --version` succeeds
- Network access to download packages and container images
2025-10-03 20:46:11 +00:00
## Time & risk
2025-10-08 22:00:07 +00:00
* **Duration:** 30 minutes for Docker approach
* **Risks:** Container registry access requires internal credentials
2025-10-10 00:11:49 +00:00
* **Rollback:** Container approach is non-destructive.
2025-10-03 20:46:11 +00:00
2025-10-06 13:35:52 +00:00
## Instructions
2025-10-10 20:39:52 +00:00
## Step 1. Pull vLLM container image
2025-10-06 13:35:52 +00:00
Find the latest container build from https://catalog.ngc.nvidia.com/orgs/nvidia/containers/vllm?version=25.09-py3
```
docker pull nvcr.io/nvidia/vllm:25.09-py3
```
## Step 2. Test vLLM in container
Launch the container and start vLLM server with a test model to verify basic functionality.
```bash
docker run -it --gpus all -p 8000:8000 \
nvcr.io/nvidia/vllm:25.09-py3 \
vllm serve "Qwen/Qwen2.5-Math-1.5B-Instruct"
```
Expected output should include:
- Model loading confirmation
- Server startup on port 8000
- GPU memory allocation details
In another terminal, test the server:
```bash
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen2.5-Math-1.5B-Instruct",
"messages": [{"role": "user", "content": "12*17"}],
"max_tokens": 500
}'
```
Expected response should contain `"content": "204"` or similar mathematical calculation.
2025-10-12 17:01:59 +00:00
## Step 3. Cleanup and rollback
2025-10-06 13:35:52 +00:00
For container approach (non-destructive):
```bash
2025-10-13 01:11:21 +00:00
docker rm $(docker ps -aq --filter ancestor=nvcr.io/nvidia/vllm:25.09-py3)
docker rmi nvcr.io/nvidia/vllm
2025-10-06 13:35:52 +00:00
```
To remove CUDA 12.9:
```bash
sudo /usr/local/cuda-12.9/bin/cuda-uninstaller
```
2025-10-12 17:01:59 +00:00
## Step 4. Next steps
2025-10-06 13:35:52 +00:00
- **Production deployment:** Configure vLLM with your specific model requirements
2025-10-10 20:39:52 +00:00
- **Performance tuning:** Adjust batch sizes and memory settings for your workload
2025-10-06 13:35:52 +00:00
- **Monitoring:** Set up logging and metrics collection for production use
- **Model management:** Explore additional model formats and quantization options
2025-10-03 20:46:11 +00:00
## Run on two Sparks
2025-10-10 20:39:52 +00:00
## Step 1. Configure network connectivity
2025-10-03 20:46:11 +00:00
2025-10-12 17:01:59 +00:00
Follow the network setup instructions from the [Connect two Sparks ](https://build.nvidia.com/spark/stack-sparks/stacked-sparks ) playbook to establish connectivity between your DGX Spark nodes.
2025-10-03 20:46:11 +00:00
2025-10-10 20:39:52 +00:00
This includes:
- Physical QSFP cable connection
- Network interface configuration (automatic or manual IP assignment)
- Passwordless SSH setup
- Network connectivity verification
2025-10-03 20:46:11 +00:00
2025-10-10 20:39:52 +00:00
## Step 2. Download cluster deployment script
2025-10-03 20:46:11 +00:00
Obtain the vLLM cluster deployment script on both nodes. This script orchestrates the Ray cluster setup required for distributed inference.
```bash
## 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
```
2025-10-10 20:39:52 +00:00
## Step 3. Pull the NVIDIA vLLM Image from NGC
2025-10-03 20:46:11 +00:00
First, you will need to configure docker to pull from NGC
If this is your first time using docker run:
```bash
sudo groupadd docker
sudo usermod -aG docker $USER
newgrp docker
```
2025-10-10 20:39:52 +00:00
After this, you should be able to run docker commands without using `sudo` .
2025-10-08 14:19:44 +00:00
2025-10-03 20:46:11 +00:00
```bash
docker pull nvcr.io/nvidia/vllm:25.09-py3
export VLLM_IMAGE=nvcr.io/nvidia/vllm:25.09-py3
```
2025-10-10 20:39:52 +00:00
## Step 4. Start Ray head node
2025-10-03 20:46:11 +00:00
Launch the Ray cluster head node on Node 1. This node coordinates the distributed inference and serves the API endpoint.
```bash
## On Node 1, start head node
export MN_IF_NAME=enP2p1s0f1np1
bash run_cluster.sh $VLLM_IMAGE 192.168.100.10 --head ~/.cache/huggingface \
-e VLLM_HOST_IP=192.168.100.10 \
-e UCX_NET_DEVICES=$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=192.168.100.10
```
2025-10-10 20:39:52 +00:00
## Step 5. Start Ray worker node
2025-10-03 20:46:11 +00:00
Connect Node 2 to the Ray cluster as a worker node. This provides additional GPU resources for tensor parallelism.
```bash
## On Node 2, join as worker
export MN_IF_NAME=enP2p1s0f1np1
bash run_cluster.sh $VLLM_IMAGE 192.168.100.10 --worker ~/.cache/huggingface \
-e VLLM_HOST_IP=192.168.100.11 \
-e UCX_NET_DEVICES=$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=192.168.100.10
```
2025-10-10 20:39:52 +00:00
## Step 6. Verify cluster status
2025-10-03 20:46:11 +00:00
Confirm both nodes are recognized and available in the Ray cluster.
```bash
## On Node 1 (head node)
docker exec node ray status
```
Expected output shows 2 nodes with available GPU resources.
2025-10-10 20:39:52 +00:00
## Step 7. Download Llama 3.3 70B model
2025-10-03 20:46:11 +00:00
Authenticate with Hugging Face and download the recommended production-ready model.
```bash
## On Node 1, authenticate and download
huggingface-cli login
huggingface-cli download meta-llama/Llama-3.3-70B-Instruct
```
2025-10-10 20:39:52 +00:00
## Step 8. Launch inference server for Llama 3.3 70B
2025-10-03 20:46:11 +00:00
Start the vLLM inference server with tensor parallelism across both nodes.
```bash
## On Node 1, enter container and start server
docker exec -it node /bin/bash
vllm serve meta-llama/Llama-3.3-70B-Instruct \
--tensor-parallel-size 2 --max_model_len 2048
```
2025-10-10 20:39:52 +00:00
## Step 9. Test 70B model inference
2025-10-03 20:46:11 +00:00
Verify the deployment with a sample inference request.
```bash
## 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
}'
```
Expected output includes a generated haiku response.
2025-10-10 20:39:52 +00:00
## Step 10. (Optional) Deploy Llama 3.1 405B model
2025-10-03 20:46:11 +00:00
2025-10-12 20:53:42 +00:00
> [!WARNING]
> 405B model has insufficient memory headroom for production use.
2025-10-03 20:46:11 +00:00
Download the quantized 405B model for testing purposes only.
```bash
## On Node 1, download quantized model
huggingface-cli download hugging-quants/Meta-Llama-3.1-405B-Instruct-AWQ-INT4
```
2025-10-10 20:39:52 +00:00
### Step 11. (Optional) Launch 405B inference server
2025-10-03 20:46:11 +00:00
Start the server with memory-constrained parameters for the large model.
```bash
## On Node 1, launch with restricted parameters
docker exec -it node /bin/bash
vllm serve hugging-quants/Meta-Llama-3.1-405B-Instruct-AWQ-INT4 \
--tensor-parallel-size 2 --max-model-len 256 --gpu-memory-utilization 1.0 \
--max-num-seqs 1 --max_num_batched_tokens 256
```
2025-10-10 20:39:52 +00:00
## Step 12. (Optional) Test 405B model inference
2025-10-03 20:46:11 +00:00
Verify the 405B deployment with constrained parameters.
```bash
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
}'
```
2025-10-10 20:39:52 +00:00
## Step 13. Validate deployment
2025-10-03 20:46:11 +00:00
Perform comprehensive validation of the distributed inference system.
```bash
## Check Ray cluster health
docker exec node ray status
## Verify server health endpoint
curl http://192.168.100.10:8000/health
## Monitor GPU utilization on both nodes
nvidia-smi
docker exec node nvidia-smi --query-gpu=memory.used,memory.total --format=csv
```
2025-10-10 20:39:52 +00:00
## Step 14. Cleanup and rollback
2025-10-03 20:46:11 +00:00
Remove temporary configurations and containers when testing is complete.
2025-10-12 20:53:42 +00:00
> [!WARNING]
> This will stop all inference services and remove cluster configuration.
2025-10-03 20:46:11 +00:00
```bash
## Stop containers on both nodes
docker stop node
docker rm node
## Remove network configuration on both nodes
sudo ip addr del 192.168.100.10/24 dev enP2p1s0f1np1 # Node 1
sudo ip addr del 192.168.100.11/24 dev enP2p1s0f1np1 # Node 2
sudo ip link set enP2p1s0f1np1 down
```
2025-10-10 20:39:52 +00:00
## Step 15. Next steps
2025-10-03 20:46:11 +00:00
Access the Ray dashboard for cluster monitoring and explore additional features:
```bash
## Ray dashboard available at:
http://192.168.100.10:8265
## Consider implementing for production:
## - Health checks and automatic restarts
2025-10-10 20:39:52 +00:00
## - Log rotation for long-running services
2025-10-03 20:46:11 +00:00
## - Persistent model caching across restarts
## - Alternative quantization methods (FP8, INT4)
```
2025-10-10 00:11:49 +00:00
## Troubleshooting
2025-10-12 17:01:59 +00:00
## Common issues for running on a single Spark
| Symptom | Cause | Fix |
|---------|--------|-----|
| CUDA version mismatch errors | Wrong CUDA toolkit version | Reinstall CUDA 12.9 using exact installer |
| Container registry authentication fails | Invalid or expired GitLab token | Generate new auth token |
| SM_121a architecture not recognized | Missing LLVM patches | Verify SM_121a patches applied to LLVM source |
## Common Issues for running on two Starks
2025-10-10 00:11:49 +00:00
| Symptom | Cause | Fix |
|---------|--------|-----|
| Node 2 not visible in Ray cluster | Network connectivity issue | Verify QSFP cable connection, check IP configuration |
2025-10-10 20:59:55 +00:00
| Cannot access gated repo for URL | Certain HuggingFace models have restricted access | Regenerate your [HuggingFace token ](https://huggingface.co/docs/hub/en/security-tokens ); and request access to the [gated model ](https://huggingface.co/docs/hub/en/models-gated#customize-requested-information ) on your web browser |
2025-10-10 00:11:49 +00:00
| Model download fails | Authentication or network issue | Re-run `huggingface-cli login` , check internet access |
2025-10-10 01:42:46 +00:00
| Cannot access gated repo for URL | Certain HuggingFace models have restricted access | Regenerate your HuggingFace token; and request access to the gated model on your web browser |
2025-10-10 00:11:49 +00:00
| CUDA out of memory with 405B | Insufficient GPU memory | Use 70B model or reduce max_model_len parameter |
| Container startup fails | Missing ARM64 image | Rebuild vLLM image following ARM64 instructions |
2025-10-12 20:13:25 +00:00
> [!NOTE]
> DGX Spark uses a Unified Memory Architecture (UMA), which enables dynamic memory sharing between the GPU and CPU.
2025-10-10 20:39:52 +00:00
> With many applications still updating to take advantage of UMA, you may encounter memory issues even when within
2025-10-10 00:11:49 +00:00
> the memory capacity of DGX Spark. If that happens, manually flush the buffer cache with:
```bash
sudo sh -c 'sync; echo 3 > /proc/sys/vm/drop_caches'
```