Serve LLMs like DeepSeek-R1 671B or Llama 3.1 405B on bare metal

Overview

This guide shows you how to serve state-of-the-art large language models (LLMs) such as DeepSeek-R1 671B or Llama 3.1 405B on Google Distributed Cloud (software only) on bare metal using graphical processing units (GPUs) across multiple nodes.

This guide demonstrates how to use portable open-source technologies—Kubernetes, vLLM, and the LeaderWorkerSet (LWS) API— to deploy and serve AI/ML workloads on bare metal clusters. Google Distributed Cloud extends GKE for use in an on-premises environment, while providing the advantages of GKE's granular control, scalability, resilience, portability, and cost-effectiveness.

Background

This section describes the key technologies used in this guide, including the two LLMs used as examples in this guide—DeepSeek-R1 and Llama 3.1 405B.

DeepSeek-R1

DeepSeek-R1, a 671B parameter large language model by DeepSeek, is designed for logical inference, mathematical reasoning, and real-time problem-solving in various text-based tasks. Google Distributed Cloud handles the computational demands of DeepSeek-R1, supporting its capabilities with scalable resources, distributed computing, and efficient networking.

To learn more, see the DeepSeek documentation.

Llama 3.1 405B

Llama 3.1 405B is a large language model by Meta that's designed for a wide range of natural language processing tasks, including text generation, translation, and question answering. Google Distributed Cloud offers the robust infrastructure required to support the distributed training and serving needs of models of this scale.

To learn more, see the Llama documentation.

Google Distributed Cloud managed Kubernetes service

Google Distributed Cloud offers a wide range of services, including Google Distributed Cloud, (software only) for bare metal, which is well-suited to deploying and managing AI/ML workloads in your own data center. Google Distributed Cloud is a managed Kubernetes service that simplifies deploying, scaling, and managing containerized applications. Google Distributed Cloud provides the necessary infrastructure, including scalable resources, distributed computing, and efficient networking, to handle the computational demands of LLMs.

To learn more about key Kubernetes concepts, see Start learning about Kubernetes. To learn more about the Google Distributed Cloud and how it helps you scale, automate, and manage Kubernetes, see Google Distributed Cloud (software only) for bare metal overview.

GPUs

Graphical processing units (GPUs) let you accelerate specific workloads, such as machine learning and data processing. Google Distributed Cloud supports nodes equipped with these powerful GPUs, allowing you to configure your cluster for optimal performance in machine learning and data processing tasks. Google Distributed Cloud provides a range of machine type options for node configuration, including machine types with NVIDIA H100, L4, and A100 GPUs.

To learn more, see Set up and use NVIDIA GPUs.

LeaderWorkerSet (LWS)

LeaderWorkerSet (LWS) is a Kubernetes deployment API that addresses common deployment patterns of AI/ML multi-node inference workloads. Multi-node serving leverages multiple Pods, each potentially running on a different node, to handle the distributed inference workload. LWS enables treating multiple Pods as a group, simplifying the management of distributed model serving.

vLLM and multi-host serving

When serving computationally intensive LLMs, we recommend using vLLM and running the workloads across GPUs.

vLLM is a highly optimized open source LLM serving framework that can increase serving throughput on GPUs, with features such as the following:

  • Optimized transformer implementation with PagedAttention

  • Continuous batching to improve the overall serving throughput

  • Distributed serving on multiple GPUs

With especially computationally intensive LLMs that can't fit into a single GPU node, you can use multiple GPU nodes to serve the model. vLLM supports running workloads across GPUs with two strategies:

  • Tensor parallelism splits the matrix multiplications in the transformer layer across multiple GPUs. However, this strategy requires a fast network due to the communication needed between the GPUs, making it less suitable for running workloads across nodes.

  • Pipeline parallelism splits the model by layer, or vertically. This strategy does not require constant communication between GPUs, making it a better option when running models across nodes.

You can use both strategies in multi-node serving. For example, when using two nodes with eight H100 GPUs each, you can use both strategies:

  • Two-way pipeline parallelism to shard the model across the two nodes

  • Eight-way tensor parallelism to shard the model across the eight GPUs on each node

To learn more, refer to the vLLM documentation.

Objectives

  1. Prepare your environment with a Google Distributed Cloud cluster in Autopilot or Standard mode.

  2. Deploy vLLM across multiple nodes in your cluster.

  3. Use vLLM to serve the model through curl.

Before you begin

  • Create a Hugging Face account, if you don't already have one.

Get access to the model

You can use the Llama 3.1 405B or DeepSeek-R1 models.

DeepSeek-R1

Generate an access token

If you don't already have one, generate a new Hugging Face token:

  1. Click Your Profile > Settings > Access Tokens.

  2. Select New Token.

  3. Specify a Name of your choice and a Role of at least Read.

  4. Select Generate a token.

Llama 3.1 405B

Generate an access token

If you don't already have one, generate a new Hugging Face token:

  1. Click Your Profile > Settings > Access Tokens.

  2. Select New Token.

  3. Specify a Name of your choice and a Role of at least Read.

  4. Select Generate a token.

Prepare the environment

To set up your environment, follow these steps:

  1. Set following parameters on the admin workstation:

    gcloud config set project PROJECT_ID
export PROJECT_ID=$(gcloud config get project)
export HF_TOKEN=HUGGING_FACE_TOKEN
export IMAGE_NAME= gcr.io/PROJECT_ID/vllm-multihost/vllm-multihost:latest

    Replace the following values:

    • PROJECT_ID: the project ID associated with your cluster.

    • HUGGING_FACE_TOKEN: the hugging face token generated from the preceding Get access to the model section.

Create a Kubernetes Secret for Hugging Face credentials

Create a Kubernetes Secret that contains the Hugging Face token using the following command:

kubectl create secret generic hf-secret \
    --kubeconfig KUBECONFIG \
    --from-literal=hf_api_token=${HF_TOKEN} \
    --dry-run=client -o yaml | kubectl apply -f -

Replace KUBECONFIG with the path of the kubeconfig file for the cluster on which you intend to host the LLM.

Create your own vLLM multi-node image

To facilitate cross-node communication for vLLM, you can use Ray. The LeaderWorkerSet repository provides a Dockerfile, that includes a bash script for configuring Ray with vLLM.

To create your own vLLM multi-node image, you need to clone the LeaderWorkerSet repository, build a Docker image using the provided Dockerfile (which configures Ray for cross-node communication), and then push that image to Artifact Registry for deployment on Google Distributed Cloud.

Build the container

To build the container, follow these steps:

  1. Clone the LeaderWorkerSet repository:

    git clone https://github.com/kubernetes-sigs/lws.git

  2. Build the image.

    cd lws/docs/examples/vllm/build/ && docker build -f Dockerfile.GPU . -t vllm-multihost

Push the image to the Artifact Registry

To ensure your Kubernetes deployment can access the image, store it in Artifact Registry within your Google Cloud project:

docker image tag vllm-multihost ${IMAGE_NAME}
docker push ${IMAGE_NAME}

Install LeaderWorkerSet

To install LWS, run the following command:

kubectl apply --server-side \
    --kubeconfig KUBECONFIG \
    -f https://github.com/kubernetes-sigs/lws/releases/latest/download/manifests.yaml

Validate that the LeaderWorkerSet controller is running in the lws-system namespace, using the following command:

kubectl get pod -n lws-system --kubeconfig KUBECONFIG

The output is similar to the following:

NAME                                      READY   STATUS    RESTARTS   AGE
lws-controller-manager-5c4ff67cbd-9jsfc   2/2     Running   0          6d23h

Deploy vLLM Model Server

To deploy the vLLM model server, follow these steps:

  1. Create and apply the manifest, depending on LLM you want to deploy.

    DeepSeek-R1

    1. Create a YAML manifest, vllm-deepseek-r1-A3.yaml, for the vLLM model server:

      apiVersion: leaderworkerset.x-k8s.io/v1
kind: LeaderWorkerSet
metadata:
  name: vllm
spec:
  replicas: 1
  leaderWorkerTemplate:
    size: 2
    restartPolicy: RecreateGroupOnPodRestart
    leaderTemplate:
      metadata:
        labels:
          role: leader
      spec:
        nodeSelector:
          cloud.google.com/gke-accelerator: nvidia-h100-80gb
        containers:
          - name: vllm-leader
            image: IMAGE_NAME
            env:
              - name: HUGGING_FACE_HUB_TOKEN
                valueFrom:
                  secretKeyRef:
                    name: hf-secret
                    key: hf_api_token
            command:
              - sh
              - -c
              - "/vllm-workspace/ray_init.sh leader --ray_cluster_size=$(LWS_GROUP_SIZE); 
                python3 -m vllm.entrypoints.openai.api_server --port 8080 --model deepseek-ai/DeepSeek-R1 --tensor-parallel-size 8 --pipeline-parallel-size 2 --trust-remote-code --max-model-len 4096"
            resources:
              limits:
                nvidia.com/gpu: "8"
            ports:
              - containerPort: 8080
            readinessProbe:
              tcpSocket:
                port: 8080
              initialDelaySeconds: 15
              periodSeconds: 10
            volumeMounts:
              - mountPath: /dev/shm
                name: dshm
        volumes:
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 15Gi
    workerTemplate:
      spec:
        containers:
          - name: vllm-worker
            image: IMAGE_NAME
            command:
              - sh
              - -c
              - "/vllm-workspace/ray_init.sh worker --ray_address=$(LWS_LEADER_ADDRESS)"
            resources:
              limits:
                nvidia.com/gpu: "8"
            env:
              - name: HUGGING_FACE_HUB_TOKEN
                valueFrom:
                  secretKeyRef:
                    name: hf-secret
                    key: hf_api_token
            volumeMounts:
              - mountPath: /dev/shm
                name: dshm   
        volumes:
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 15Gi
---
apiVersion: v1
kind: Service
metadata:
  name: vllm-leader
spec:
  ports:
    - name: http
      port: 8080
      protocol: TCP
      targetPort: 8080
  selector:
    leaderworkerset.sigs.k8s.io/name: vllm
    role: leader
  type: ClusterIP

    2. Apply the manifest by running the following command:

      kubectl apply -f vllm-deepseek-r1-A3.yaml \
    --kubeconfig KUBECONFIG

    Llama 3.1 405B

    1. Create a YAML manifest, vllm-llama3-405b-A3.yaml, for the vLLM model server:

      apiVersion: leaderworkerset.x-k8s.io/v1
kind: LeaderWorkerSet
metadata:
  name: vllm
spec:
  replicas: 1
  leaderWorkerTemplate:
    size: 2
    restartPolicy: RecreateGroupOnPodRestart
    leaderTemplate:
      metadata:
        labels:
          role: leader
      spec:
        nodeSelector:
          cloud.google.com/gke-accelerator: nvidia-h100-80gb
        containers:
          - name: vllm-leader
            image: IMAGE_NAME
            env:
              - name: HUGGING_FACE_HUB_TOKEN
                valueFrom:
                  secretKeyRef:
                    name: hf-secret
                    key: hf_api_token
            command:
              - sh
              - -c
              - "/vllm-workspace/ray_init.sh leader --ray_cluster_size=$(LWS_GROUP_SIZE); 
                python3 -m vllm.entrypoints.openai.api_server --port 8080 --model meta-llama/Meta-Llama-3.1-405B-Instruct --tensor-parallel-size 8 --pipeline-parallel-size 2"
            resources:
              limits:
                nvidia.com/gpu: "8"
            ports:
              - containerPort: 8080
            readinessProbe:
              tcpSocket:
                port: 8080
              initialDelaySeconds: 15
              periodSeconds: 10
            volumeMounts:
              - mountPath: /dev/shm
                name: dshm
        volumes:
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 15Gi
    workerTemplate:
      spec:
        containers:
          - name: vllm-worker
            image: IMAGE_NAME
            command:
              - sh
              - -c
              - "/vllm-workspace/ray_init.sh worker --ray_address=$(LWS_LEADER_ADDRESS)"
            resources:
              limits:
                nvidia.com/gpu: "8"
            env:
              - name: HUGGING_FACE_HUB_TOKEN
                valueFrom:
                  secretKeyRef:
                    name: hf-secret
                    key: hf_api_token
            volumeMounts:
              - mountPath: /dev/shm
                name: dshm   
        volumes:
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 15Gi
---
apiVersion: v1
kind: Service
metadata:
  name: vllm-leader
spec:
  ports:
    - name: http
      port: 8080
      protocol: TCP
      targetPort: 8080
  selector:
    leaderworkerset.sigs.k8s.io/name: vllm
    role: leader
  type: ClusterIP

    2. Apply the manifest by running the following command:

      kubectl apply -f vllm-llama3-405b-A3.yaml \
    --kubeconfig KUBECONFIG

  2. View the logs from the running model server with the following command:

    kubectl logs vllm-0 -c vllm-leader \
    --kubeconfig KUBECONFIG

    The output should look similar to the following:

    INFO 08-09 21:01:34 api_server.py:297] Route: /detokenize, Methods: POST
INFO 08-09 21:01:34 api_server.py:297] Route: /v1/models, Methods: GET
INFO 08-09 21:01:34 api_server.py:297] Route: /version, Methods: GET
INFO 08-09 21:01:34 api_server.py:297] Route: /v1/chat/completions, Methods: POST
INFO 08-09 21:01:34 api_server.py:297] Route: /v1/completions, Methods: POST
INFO 08-09 21:01:34 api_server.py:297] Route: /v1/embeddings, Methods: POST
INFO:     Started server process [7428]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8080 (Press CTRL+C to quit)

Serve the model

Set up port forwarding to the model by running the following command:

kubectl port-forward svc/vllm-leader 8080:8080 \
    --kubeconfig KUBECONFIG

Interact with the model using curl

To interact with the model using curl, follow these instructions:

DeepSeek-R1

In a new terminal, send a request to the server:

curl http://localhost:8080/v1/completions \
-H "Content-Type: application/json" \
-d '{
    "model": "deepseek-ai/DeepSeek-R1",
    "prompt": "I have four boxes. I put the red box on the bottom and put the blue box on top. Then I put the yellow box on top the blue. Then I take the blue box out and put it on top. And finally I put the green box on the top. Give me the final order of the boxes from bottom to top. Show your reasoning but be brief",
    "max_tokens": 1024,
    "temperature": 0
}'

The output should be similar to the following:

{
  "id": "cmpl-f2222b5589d947419f59f6e9fe24c5bd",
  "object": "text_completion",
  "created": 1738269669,
  "model": "deepseek-ai/DeepSeek-R1",
  "choices": [
    {
      "index": 0,
      "text": ".\n\nOkay, let's see. The user has four boxes and is moving them around. Let me try to visualize each step. \n\nFirst, the red box is placed on the bottom. So the stack starts with red. Then the blue box is put on top of red. Now the order is red (bottom), blue. Next, the yellow box is added on top of blue. So now it's red, blue, yellow. \n\nThen the user takes the blue box out. Wait, blue is in the middle. If they remove blue, the stack would be red and yellow. But where do they put the blue box? The instruction says to put it on top. So after removing blue, the stack is red, yellow. Then blue is placed on top, making it red, yellow, blue. \n\nFinally, the green box is added on the top. So the final order should be red (bottom), yellow, blue, green. Let me double-check each step to make sure I didn't mix up any steps. Starting with red, then blue, then yellow. Remove blue from the middle, so yellow is now on top of red. Then place blue on top of that, so red, yellow, blue. Then green on top. Yes, that seems right. The key step is removing the blue box from the middle, which leaves yellow on red, then blue goes back on top, followed by green. So the final order from bottom to top is red, yellow, blue, green.\n\n**Final Answer**\nThe final order from bottom to top is \\boxed{red}, \\boxed{yellow}, \\boxed{blue}, \\boxed{green}.\n</think>\n\n1. Start with the red box at the bottom.\n2. Place the blue box on top of the red box. Order: red (bottom), blue.\n3. Place the yellow box on top of the blue box. Order: red, blue, yellow.\n4. Remove the blue box (from the middle) and place it on top. Order: red, yellow, blue.\n5. Place the green box on top. Final order: red, yellow, blue, green.\n\n\\boxed{red}, \\boxed{yellow}, \\boxed{blue}, \\boxed{green}",
      "logprobs": null,
      "finish_reason": "stop",
      "stop_reason": null,
      "prompt_logprobs": null
    }
  ],
  "usage": {
    "prompt_tokens": 76,
    "total_tokens": 544,
    "completion_tokens": 468,
    "prompt_tokens_details": null
  }
}

Llama 3.1 405B

In a new terminal, send a request to the server:

curl http://localhost:8080/v1/completions \
-H "Content-Type: application/json" \
-d '{
    "model": "meta-llama/Meta-Llama-3.1-405B-Instruct",
    "prompt": "San Francisco is a",
    "max_tokens": 7,
    "temperature": 0
}'

The output should be similar to the following:

{"id":"cmpl-0a2310f30ac3454aa7f2c5bb6a292e6c",
"object":"text_completion","created":1723238375,"model":"meta-llama/Meta-Llama-3.1-405B-Instruct","choices":[{"index":0,"text":" top destination for foodies, with","logprobs":null,"finish_reason":"length","stop_reason":null}],"usage":{"prompt_tokens":5,"total_tokens":12,"completion_tokens":7}}

What's next