llama.cpp

d7f794ea - convert : avoid dequantizing mxfp4 for GPT-OSS