llama.cpp

[CUDA] Increase number of output elements per-thread block if the K-dimension is small

#20635

Merged