llama.cpp

ggml : optimize cuda ssm_scan using warp-level reduction

#18505

Merged