llama.cpp

9ef7523e - cuda/hip: fix loop unrolling in ssm-conv (#20369)