mradermacher

/

model_requests

like 187

This model has a custom architecture that probably isn't supported by Llama.cpp.

It’s a pruned gemma 4 31b. vllm needs a custom model because the kv head count varies by layer, but llama.cpp can use the standard gemma 4 architecture.

Conversion uses the specified architecture which means that a source repository specifying the compatible architecture will be needed.

I see. In that case, I'll prepare a new repository.

Should I try queueing or wait for new repo? sorry for delay

Wait for a compatible version