llama.cpp

ggml-webgpu: improve MTP inference by using mat-vec path for small batches

#24811

Merged