llama.cpp

a3c28439 - cuda : fine-tune >= VOLTA params + use MMQ only for small batches