Running DeepSeek R1 671B with Distributed vLLM

This tutorial guides you through the process of configuring and running the unquantized DeepSeek R1 671B using Distributed vLLM on a GPUStack cluster. Due to the extremely large size of the model, distributed inference across multiple workers is usually required.

GPUStack enables easy setup and orchestration of distributed inference using vLLM, making it possible to run massive models like DeepSeek R1 with minimal manual configuration.

Prerequisites

Before you begin, make sure the following requirements are met:

• You have access to a sufficient number of Linux nodes, each equipped with the required GPUs. For example:

GPU	Number of Nodes
H100/H800:8	2
A100/A800-80GB:8	4
A100/A800:8	8

• High-speed interconnects such as NVLink or InfiniBand are recommended for optimal performance.
• Model files should be downloaded to the same path on each node. While GPUStack supports on-the-fly model downloading, pre-downloading is recommended as it can be time consuming depending on the network speed.

Note

• In this tutorial, we assume a setup of 4 nodes, each equipped with 8 A800-80GB GPUs and connected via 200G InfiniBand.
• A100/A800 GPUs do not support the FP8 precision originally used by DeepSeek R1. Hence, we use the BF16 version from Unsloth.

Step 1: Install GPUStack Server

In this tutorial, we will use Docker to install GPUStack. You can also use other installation methods if you prefer.

Use the following command to start the GPUStack server:

docker run -d --name gpustack \
    --restart=unless-stopped \
    --gpus all \
    --network=host \
    --ipc=host \
    -v gpustack-data:/var/lib/gpustack \
    -v /path/to/your/model:/path/to/your/model \
    -e NCCL_SOCKET_IFNAME=eth2 \
    -e GLOO_SOCKET_IFNAME=eth2 \
    gpustack/gpustack --enable-ray

Note

• Replace /path/to/your/model with the actual path.
• Set NCCL_SOCKET_IFNAME and GLOO_SOCKET_IFNAME to the network interface used for inter-node communication. We use eth2 as an example.
• The --enable-ray flag enables Ray for distributed inference, which is required by vLLM.

After GPUStack server is up and running, run the following commands to get the initial admin password and the token for worker registration:

docker exec gpustack cat /var/lib/gpustack/initial_admin_password
docker exec gpustack cat /var/lib/gpustack/token

Step 2: Install GPUStack Workers

On each worker node, run the following command to start a GPUStack worker:

docker run -d --name gpustack \
    --restart=unless-stopped \
    --gpus all \
    --network=host \
    --ipc=host \
    -v gpustack-data:/var/lib/gpustack \
    -v /path/to/your/model:/path/to/your/model \
    -e NCCL_SOCKET_IFNAME=eth2 \
    -e GLOO_SOCKET_IFNAME=eth2 \
    gpustack/gpustack \
    --server-url http://your_gpustack_server_ip_or_hostname \
    --token your_gpustack_token \
    --enable-ray

Note

• Replace the placeholder paths, IP address/hostname, and token accordingly.
• Ensure the model path matches that of the server and is valid on all worker nodes.

Step 3: Access GPUStack UI

Once the server and all workers are running, access the GPUStack UI via your browser:

http://your_gpustack_server_ip_or_hostname

Log in using the admin username and the password obtained in Step 1. Navigate to the Resources page to verify that all workers are in the Ready state and their GPUs are listed.

Step 4: Deploy the DeepSeek R1 Model

1. Go to the Models page.
2. Click Deploy Model.
3. Select Local Path as your source.
4. Enter a name (e.g., DeepSeek-R1) in the Name field.
5. Specify the Model Path as the directory that contains the DeepSeek R1 model files on each worker node.
6. Ensure the Backend is set to vLLM.
7. After passing the compatibility check, click Save to deploy.

Step 5: Monitor Deployment

You can monitor the deployment status on the Models page. Hover over distributed across workers to view GPU and worker usage. Click View Logs to see real-time logs showing model loading progress. It may take a few minutes to load the model.

After the model is running, revisit the Resources tab to check GPU utilization. By default, vLLM uses 90% of GPU memory. You may adjust this in the model configuration settings.

Step 6: Run Inference via Playground

Once the model is deployed and running, you can test it using the GPUStack Playground.

1. Navigate to the Playground -> Chat.
2. If only one model is deployed, it will be selected by default. Otherwise, use the dropdown menu to choose DeepSeek-R1.
3. Enter prompts and interact with the model.

You can also use the Compare tab to test conccurrent inference scenarios.

You have now successfully deployed and run DeepSeek R1 671B using Distributed vLLM on a GPUStack cluster. Explore the model’s performance and capabilities in your own applications.

For further assistance, feel free to reach out to the GPUStack community or support team.