llama.cpp
CUDA/HIP: refractor mmqv to unify the calculation of nwarps and rows per block between host and device code.
#12177
Merged

Loading