Skip to content

Conversation

lstein
Copy link
Collaborator

@lstein lstein commented Aug 9, 2023

What type of PR is this? (check all applicable)

  • [X Feature
  • Documentation Update

Have you discussed this change with the InvokeAI team?

  • Yes

Have you updated all relevant documentation?

  • [X Yes

Description

Install torch+cu118 rather than cu117. Also updates documentation regarding installing xformers and the CUDA toolkit, which were referring to older versions.

Copy link
Member

@ebr ebr left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM! (Ubuntu 22.10 / python 3.11)

I did have to use the pip --force-reinstall --extra-index-url ... flags to coerce pip into getting torch2.0.1+cu118, even in a new virtual environment. Otherwise it was finding and using the cached cu117 version. This will not be a problem when using the installer, but may come up as a support question from users doing manual installs (even though this is already documented).

@psychedelicious psychedelicious self-requested a review August 10, 2023 00:01
Copy link
Collaborator

@psychedelicious psychedelicious left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just a datapoint: I get 13it/s on current torch with cu117, and 39it/s on cu118. I couldn't figure out why (xformers/torch sdp not working? dunno). It's possible this will resolve some comments about performance.

@hipsterusername hipsterusername enabled auto-merge (rebase) August 10, 2023 19:17
@hipsterusername hipsterusername merged commit 7bad9bc into main Aug 10, 2023
@hipsterusername hipsterusername deleted the feat/use-cu118 branch August 10, 2023 19:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants