llama.cpp
Force NVFP4 W4A8 path for NVFP4_W4A16 layers on Blackwell, where NVFP4 normally uses the native W4A4 path.
#24364
Open

Loading