whisper.cpp

77bf6b5f - CUDA: mul_mat_vec_q max. batch size 8 -> 4 (llama/5370)