llama.cpp

CUDA/HIP: fix race in flash_attn_ext_f16_process_tile when nbatch_combine != DKQ/2

#30103

Merged