Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft: Feat/update plugins #84

Closed
wants to merge 2 commits into from

Conversation

blester125
Copy link
Collaborator

This PR is a draft of two things, first of using plugins for update types, the second is for using git to store and fetch updates that are based on deltas from previous values.

Points of Interest:

  1. Parameter metadata is in a new metadata file for the parameter within the parameter directory. This was critical for the model checkpoint clean filter running in contexts like git status
  2. The smudge filter isn't perfect for time-travel. When git checkout ${COMMIT} happens the version of the cleaned model checkpoint from that commit is passed via CLI, but the parameter values start with the ones on disk. For the disk only solution this is ok (as long as the parameters were checked out before the checked-in model checkpoint is smudged, which currently happens because .git_theta sorts first). For this solution you end up with a wrong checkpoint where the parameters are from the commit you are leaving instead of the commit you are entering. This can be fixed by running git reset --hard after the checkout happens. But any thoughts on how to fix this would be great.
  3. I converted some of the logging to file, because logging in git-theta-filter doesn't get shown on the screen. We should revert this or make logging more robust before merging.
  4. "Sparse" updates are very simplistic right now, they are stored as dense offsets just to get it done, not something smart like a CSR matrix.
  5. If we don't go with the git method, we can at least use the update plugins for a disk-based setting. Similarly, ideas like the recursive update can still be used with on disk updates.

Depends on #74, once that is merged the changes here will be easier to see.

@blester125 blester125 marked this pull request as draft November 21, 2022 17:44
It also looks at using the git history to store and apply updates.
@blester125 blester125 closed this Nov 28, 2022
@blester125 blester125 deleted the feat/update-plugins branch November 28, 2022 16:09
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.

None yet

1 participant