whisper.cpp

8bb85d96 - CUDA: prefer whole-tile FlashAttention scheduling for efficient two-stage kernels (llama/29435)