support vllm cpu in backendruntime

[googs1025]

What this PR does / why we need it

• support vllm cpu in backendruntime

Which issue(s) this PR fixes

Fixes #

Special notes for your reviewer

Does this PR introduce a user-facing change?

support vllm cpu in backendruntime

[googs1025]

To be honest: I think this is a bit redundant, because vllm cpu image is maintained separately (see: https://gallery.ecr.aws/q9t5s3a7/vllm-cpu-release-repo), and at this stage we seem to have to use another backend runtime to maintain it. 🤔

[googs1025]

root@VM-0-13-ubuntu:/home/ubuntu# kubectl get pods
NAME            READY   STATUS    RESTARTS   AGE
qwen3-0--6b-0   1/1     Running   0          24m
root@VM-0-13-ubuntu:/home/ubuntu# kubectl get pods -oyaml | grep image
      image: public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v0.8.5
      imagePullPolicy: IfNotPresent
      image: inftyai/model-loader:v0.0.10
      imagePullPolicy: IfNotPresent
      image: public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v0.8.5
      imageID: public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo@sha256:36444ac581f98dc4336a7044e4c9858d5b3997a7bda1e952c03af8b6917c8311
      image: docker.io/inftyai/model-loader:v0.0.10
      imageID: docker.io/inftyai/model-loader@sha256:b67a8bb3acbc496a62801b2110056b9774e52ddc029b379c7370113c7879c7d9

[googs1025]

root@VM-0-13-ubuntu:/home/ubuntu# kubectl port-forward svc/qwen3-0--6b-lb 8080:8080
Forwarding from 127.0.0.1:8080 -> 8080
Forwarding from [::1]:8080 -> 8080

[googs1025]

root@VM-0-13-ubuntu:/home/ubuntu# curl -X POST http://localhost:8080/v1/chat/completions   -H 'Content-Type: application/json'   -d '{
        "model": "qwen3-0--6b",
        "messages": [
          {
            "role": "user",
            "content": "Who are you?"
          }
        ]
      }' | jq
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  1436  100  1273  100   163    111     14  0:00:11  0:00:11 --:--:--   393
{
  "id": "chatcmpl-7f5a02b2bd964760832cdf7f5e0a104d",
  "object": "chat.completion",
  "created": 1751164548,
  "model": "qwen3-0--6b",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "reasoning_content": null,
        "content": "<think>\nOkay, the user is asking, \"Who are you?\" So, I need to respond appropriately. Let me start by recalling the previous interactions. The user mentioned they are asking about my identity, so I should confirm my name and provide a brief description.\n\nI should make sure to be friendly and offer further assistance. Maybe mention that I am a language model, but also highlight that I can help with various tasks. It's important to keep the response straightforward and conversational.\n\nWait, should I use a specific name? The user might be expecting a name, so I should include that. Also, avoid any technical jargon. Keep the tone natural and helpful. Let me put that together.\n</think>\n\nI am a language model developed by OpenAI. I can help with a wide range of tasks, from answering questions to providing information. How can I assist you today?",
        "tool_calls": []
      },
      "logprobs": null,
      "finish_reason": "stop",
      "stop_reason": null
    }
  ],
  "usage": {
    "prompt_tokens": 12,
    "total_tokens": 190,
    "completion_tokens": 178,
    "prompt_tokens_details": null
  },
  "prompt_logprobs": null
}
root@VM-0-13-ubuntu:/home/ubuntu#

[googs1025]

Seeing that other projects have vllm cpu type example, I think we also can integrate it into our backendruntime.

https://github.com/kubernetes-sigs/gateway-api-inference-extension/blob/main/config/manifests/vllm/cpu-deployment.yaml

[googs1025]

/kind feature

[carlory]

The image seems only support the x86 architecture. https://docs.vllm.ai/en/latest/getting_started/installation/cpu.html?h=cpu#pre-built-images

[googs1025]

The image seems only support the x86 architecture. https://docs.vllm.ai/en/latest/getting_started/installation/cpu.html?h=cpu#pre-built-images

Yes, according to the docs, currently vllm cpu only supports specific architectures 🤔

[cr7258]

I think we can move it to the examples instead of the part of the default template.

[googs1025]

This seems more reasonable, and I will change. In examples, we provide users another way to support vllm(in cpu).

[googs1025]

cc @kerthcet

[cr7258]

🐉 LGTM

[carlory]

/assign @kerthcet

[googs1025]

friendly ping @kerthcet
can you check this when you have time? 😄