Path to upstream vLLM and HF support for granite_switch architecture

[yairallouche]

Summary

Granite Switch currently works as a vLLM plugin (out-of-tree model) and a standalone HF model. Should we pursue making GraniteSwitchForCausalLM a supported model architecture in upstream vLLM and/or HuggingFace transformers?

Considerations

• Pros: Eliminates plugin maintenance burden; broader visibility and testing; benefits from upstream CI and release process.
• Cons: Upstream acceptance criteria may require changes; ongoing maintenance commitment; slower iteration cycle.

Next steps

• Gauge vLLM community interest and understand acceptance criteria
• Attempt HF model classes upstream via PR
• This is a tracking issue — concrete work depends on community feedback

Consolidates original issues #21 (vLLM upstream), #50 (HF upstream), and #61 (vLLM PR).

Links

https://docs.vllm.ai/en/latest/contributing/model/registration/

[lastras]

I think this is something we should pursue but I think there is some
Concrete work we need to do here to create more native kernels for the granite switch architecture.