Skip to content

[Feature Request] Expose llama.cpp --no-mmproj-offload option when loading vision models #1873

Description

@wikihua

Feature request

Please consider adding an option in LM Studio to control whether the multimodal projector (mmproj) is offloaded to GPU when loading vision-capable GGUF models.

In llama.cpp, vision models can be launched with:

--mmproj /path/to/mmproj.gguf
--no-mmproj-offload

This allows the main language model to use GPU offload while keeping the multimodal projector on CPU.

Why this would be useful

For large vision-capable models, the mmproj file can consume additional VRAM. In some cases, users may want to:

keep most or all LLM layers on GPU for text generation speed
keep mmproj on CPU to reduce VRAM usage
still keep vision input available when needed
avoid lowering the overall GPU offload ratio just to make room for mmproj

Currently, LM Studio exposes general GPU offload settings, but I could not find a separate option for controlling mmproj offload behavior.

Suggested UI / API behavior

Possible implementations:

Add an advanced model loading option:
Multimodal projector offload: Auto / GPU / CPU
Or expose a boolean option:
Offload mmproj to GPU: on/off

A dedicated mmproj offload setting would make LM Studio much more flexible for large multimodal GGUF models.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions