llama.cpp

f818065d - CUDA: batch out_prod broadcast (dps2>1) path with cublasSgemmBatched (#24426)