I made a llama.cpp Metal optimization that gives a nice decode speedup for IQ3_XXS models on Apple Silicon: https://github.com/ggml-org/llama.cpp/pull/28086 Basically on On my test workload I saw decode go from about 65.6 → 73.9 tok/s for tiel coder 35B A3B Would love if you guys could try it and sh
