Support launching with --no-mmproj-offload or without mmproj on llama.cpp

[hun]

Feature Description

When a model contains an mmproj, it will be put in VRAM, even if i don't intend to use it. Using --no-mmproj-offload would work, but it is forbidden by lemonade. Just allowing the option would be fine for me.

Use Case / Motivation

Have more VRAM for useful things

Platform Relevance

All platforms

Additional Context

No response

[boecks]

I second that, because currently, I have to define user_models just so i can get a non-vision version of e.g. qwen3.6 models for agentic coding. My goal is the same as OP: do not allocate vram for an unused feature.