llama.cpp

[CUDA ] Write an optimized flash_attn_stream_k_fixup kernel

#21159

Merged