llama.cpp

convert : force f16 or f32 on step3-vl conv weights

#21646

Merged