llama.cpp

sycl : port multi-column MMVQ from CUDA backend (~45% speculative decoding speedup on Intel Arc)

#21845

Merged