llama.cpp

35195689 - 2x faster (rms) norm cuda kernels (3.7% e2e improvement) (#2985)