Skip to content

Conversation

Laurent2916
Copy link
Contributor

@Laurent2916 Laurent2916 commented Oct 9, 2024

That's a pretty fat PR:

  • Deleted all old conversion scripts.
  • Created a new conversion module which converts weight state_dicts based on simple mappings (most of the time), inspired by imaginAIry.
    • Mapping conversion is way more lighter/faster than using the old ModelConverter (this allows weights to be converted on the fly if necessary/wanted).
    • Each conversion is now declarative. The hash (and sometime git commit revision) of both the original weights and converted weights are now strictly checked (to avoid unwanted regressions).
  • Created new test fixture use_local_weights that allows to run the tests either with local converted weights, or the ones from the refiners huggingface org.
  • Modified all existing tests to use these new fixtures.
  • Added DINOv2 similarity when testing images in e2e tests.
  • Loosen some tests to be good on both a 3090 and an L4.

See the updated CONTRIBUTING.md for instructions.

@Laurent2916 Laurent2916 self-assigned this Oct 9, 2024
@Laurent2916 Laurent2916 force-pushed the pr/rework_weight_conversions branch from e88f156 to 58ccd3d Compare October 11, 2024 08:27
@Laurent2916 Laurent2916 force-pushed the pr/rework_weight_conversions branch from 58ccd3d to f45fa7e Compare October 11, 2024 08:29
catwell
catwell previously approved these changes Oct 14, 2024
@Laurent2916 Laurent2916 force-pushed the pr/rework_weight_conversions branch from 917a7e5 to 78a3336 Compare October 14, 2024 12:25
@Laurent2916 Laurent2916 requested a review from catwell October 14, 2024 12:26
@Laurent2916 Laurent2916 merged commit ef0a752 into main Oct 14, 2024
3 checks passed
@Laurent2916 Laurent2916 deleted the pr/rework_weight_conversions branch October 14, 2024 13:13
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.

2 participants