Revert "CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#24546)" - #28551

Revert "CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#24546)"#28551

pwilkin merged 1 commit into

masterggml-org/llama.cpp:masterfrom

revert-mmq-moe-ncolsggml-org/llama.cpp:revert-mmq-moe-ncolsCopy head branch name to clipboard