Skip to content

CPU offloading / low-VRAM support #137

@philpax

Description

@philpax

As contributed in #97. That was not merged at the time due to a lack of agreement over what this would entail, but I think we have a clearer picture of it now:

  • Wherever possible, when loading with World Engine or Quark, avoid materialising intermediate results within the GPU.
  • Scene Authoring's models should be resident on the CPU wherever possible. I think the primary beneficiaries of this would be the VLM and the text encoder for Flux.

This should make it possible to use WP-1.5-1B with Scene Authoring fairly reliably on a 8GB VRAM card.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions