Skip to content

Running DeepSeek R1 671B with Distributed Ascend MindIE

This tutorial guides you through the process of configuring and running the original DeepSeek R1 671B using Distributed Ascend MindIE 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 Ascend MindIE, 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 Ascend Atlas AI processors. For example:
Server(NPU) Number of Nodes
Atlas 800T A2 (910B3 x 8) 4
  • All nodes need to support Huawei Collective Communication adaptive Protocol (HCCP), installed Huawei Collective Communication Library (HCCL), and enabled Huawei Cache Coherence System (HCCS).
  • 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 910B3 NPUs and connected via 200G Huawei Cache Conherence Network (HCCN).
  • Altas NPUs do not support the FP8 precision originally used by DeepSeek R1. Hence, we use the BF16 version from Unsloth.

Step 1: Install GPUStack Server

According to the Ascend Installation, you can use the following command to start the GPUStack server with the built-in worker:

sudo docker run -d --name gpustack \
    --restart unless-stopped \
    --privileged \
    --env "ASCEND_VISIBLE_DEVICES=$(sudo ls /dev/davinci* | head -1 | grep -o '[0-9]\+' || echo "0")" \
    --network host \
    --volume /var/run/docker.sock:/var/run/docker.sock \
    --volume gpustack-data:/var/lib/gpustack \
    --volume /path/to/your/model:/path/to/your/model \
    --runtime ascend \
    gpustack/gpustack

Note

  • Replace /path/to/your/model with the actual path on your system where the DeepSeek R1 model files are stored.
  • Ensure the hccn_tool tool is installed and configured correctly on your system. This is required for discvoring the HCCN network communication.

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

sudo docker exec gpustack \
    cat /var/lib/gpustack/initial_admin_password

Step 2: Access GPUStack UI

Login to the GPUStack UI using the admin user and the obtained password.

http://your_gpustack_server_ip_or_hostname

Step 3: Install GPUStack Workers

Navigate to the Workers page in the GPUStack UI, click Add Worker button to get the command for adding workers.

And then on each worker node, run the worker adding command to start a GPUStack worker:

sudo docker run -d --name gpustack \
    --restart unless-stopped \
    --privileged \
    --env "ASCEND_VISIBLE_DEVICES=$(sudo ls /dev/davinci* | head -1 | grep -o '[0-9]\+' || echo "0")" \
    --network host \
    --volume /var/run/docker.sock:/var/run/docker.sock \
    --volume gpustack-data:/var/lib/gpustack \
    --volume /path/to/your/model:/path/to/your/model \
    --runtime ascend \
    gpustack/gpustack \
    --server-url http://your_gpustack_server_ip_or_hostname \
    --token your_gpustack_cluster_token

Note

  • Replace the placeholder paths, IP address/hostname, and cluster token accordingly.
  • Replace /path/to/your/model with the actual path on your system where the DeepSeek R1 model files are stored.
  • Ensure the hccn_tool tool is installed and configured correctly on your system. This is required for discvoring the HCCN network communication.

After all workers are added, return to the GPUStack UI.

Navigate to the Workers page to verify that all workers are in the Ready state and their GPUs are listed.

initial-resources

Step 4: Deploy the DeepSeek R1 Model

  1. Go to the Deployments 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 Ascend MindIE.
  7. Expand Advanced, append following parameters to Backend Parameters: - --data-parallel-size=4 - --tensor-parallel-size=8 - --moe-tensor-parallel-size=1 - --moe-expert-parallel-size=32 - --npu-memory-fraction=0.95, since we are using Data Parallelism, the memory fraction should be set to 0.95 to ensure efficient memory usage across all NPUs.
  8. After passing the compatibility check, click Save to deploy.

deploy-model-1 deploy-model-2

Step 5: Monitor Deployment

You can monitor the deployment status on the Deployments 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.

model-info

After the model is running, navigate to the Workers page to check GPU utilization. Ascend MindIE uses 95% of NPU memory with above settings.

resources-loaded

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.

playground-chat

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

playground-compare

You have now successfully deployed and run DeepSeek R1 671B using Distributed Ascend MindIE 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.