llama.cpp

8060c074 - ggml: add tensor scaling (per-tensor, per-channel, per-expert, per-expert-per-channel) support