llama.cpp

CUDA: refactor topk-moe to enable more models (GLM 4.7, Nemotron etc.)

#19126

Merged