llama.cpp

CUDA: Improve NVFP4 W4A4 activation quantization

#25730

Merged