Support alternate 2D linear checkpoint layouts#166
Open
lesj0610 wants to merge 1 commit intoturboderp-org:devfrom
Open
Support alternate 2D linear checkpoint layouts#166lesj0610 wants to merge 1 commit intoturboderp-org:devfrom
lesj0610 wants to merge 1 commit intoturboderp-org:devfrom
Conversation
lesj0610
force-pushed
the
fix/checkpoint-layout-2d-linears
branch
from
d06d28a to
184f0a3
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This is a small compatibility improvement for architecture-preserving checkpoint format variation. The goal is not to support arbitrary model-specific rewrites, but to make linear weight loading more robust when modified checkpoints keep the same module graph while changing storage layout.
Some modified checkpoints keep the same architecture but store 2D linear weights in the opposite orientation. This can happen after resaves, merges, abliterations, repacks, or other tooling that preserves the module graph but rewrites tensor storage.
This keeps the configured orientation by default and only flips when the configured orientation does not fit the expected padded shape and the opposite orientation does.
The change is limited to 2D linear loads used by the existing FP16 path. Fused tensors and checkpoints that already match the current orientation keep the existing behavior.
Adds focused tests for the flip and no-flip cases.