Got an old slow low vram GPU laying around? Might be worth it to use for Just Vision mmproj llama.cpp

For many, Vram is precious, I see many people recommend using --no-mmproj-offload to save gpu vram but it is painfully slow. Especially if you are using it with agentic coding. If possible, add that secondary gpu just for mmproj with --mmdev CUDA1(your gpu). It will be a magnitude faster than --no-mmproj-offload without effecting your inference speed.

--mmdev