Skip to content

Installer must be more smart about Nvidia 2xxx and lower install #125

Description

@BieHDC

I installed InvokeAI today on a pc with a GTX 980TI for testing and i did select said option, but it installed versions of torch and torchvision that did not support the old CUDA version of this specific card.
I fixed it by manually going into the developer console and tying:
uv pip uninstall torch torchvision xformers
uv pip install torch==2.5.0 torchvision==0.20.0 xformers --index-url https://download.pytorch.org/whl/cu124
After that invoke works at least for SDXL since i did not test more scenarios.
Works as in 8-9sec per iteration. Uses around all of the 6gb vram without lowvram-toggle and around 4gb with it enabled.

The command was sourced from here https://github.com/llmlaba/llmlaba.github.io/blob/main/_articles/cuda-pytorch-compatibility.md if anyone reading this has another gpu that falls into this issue and here https://developer.nvidia.com/cuda/gpus/legacy is nvidias list for which gpu does which cuda version.

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