Could we request support for a smallish (~4-5B param) modern vision LLM? LLava-1.6 or Nanollava?

[kinchahoy]

🚀 The feature, motivation and pitch

Having good basic pytorch support for inferencing LLMs is key to continued success of pytorch. Vision LLM models tend to have uneven support on mainstream inferencing engines like Llama.cpp due to the need to reimplement CLIP/SIGLIP etc. Pytorch could natively support performant vision LLMs with quantization on ARM devices, which would make a big difference in usability.

Alternatives

No response

Additional context

No response

RFC (Optional)

No response