llama.cpp

[CUDA] Reduce the number of stream-k blocks to reduce the overhead of the flash_attn_stream_k_fixup kernel

#21086

Closed