llama.cpp

ggml-cuda: ds_read_b128 for q4_0 and q4_1 mmq kernels

#21168

Merged