Skip to content

[Feature] Fine-grained model loading controls (VRAM/RAM offloading, profiles) #6530

Description

@mr-pepe69

I started using Unsloth Studio and I can't help but compare it with LM Studio. There are some features that I miss, one of them is the ability to tweak how the model is loaded. Currently it seems that it tries to fill the VRAM and offload everything else to RAM. I use Unsloth Studio directly on my main PC.
I would like more fine control so I can use models while I play games or watch videos. I like the control over how many experts are offloaded to RAM.
If you decided to add this feature I recommend that you also add Profiles. For example:

Profile 1 loads more layers into the GPU and decreases the context window.
Profile 2 has quantized KV cache and a larger context window.

This could be managed just before loading the model as a pop-up window with some tabs on top to choose the profile.
In regard to what sliders and toggles would have, I think it would be nice to cover LM Studio features. More users would feel familiar with it and have an easier adoption.
I did see the sidebar that lets you tweak the model, I just feel that there are things missing for me. this stuff could be implement directly there.
Beside that the app is pretty good, running it in docker is convenient.
I apologize if this isn't as technical as it needs to be — I am just a user that saw that something could improve.

Metadata

Metadata

Assignees

Labels

feature requestFeature request pending on roadmap

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions