feature: multiple runtime classes

[kfox1111]

Describe the feature

https://github.com/vllm-project/production-stack/blob/main/helm/templates/deployment-vllm-multi.yaml#L450

only looks at a global for runtimeClassName. It would be nice to be able to specify it per instance too. So, if it is specified, use it, otherwise fall back to the global one.

This can be useful when you have different types of gpu nodes and want to target models to pools of hardware with different capabilities.

Why do you need this feature?

No response

Additional context

No response

[HanFa]

Hi @kfox1111, I have the proposed interface change ready at #755, could you possibly review it and see if it satisfies your use case?

[kfox1111]

Looks good to me. thanks! :)