fix(merger): handle non-sharded tensors in FSDP2 checkpoint merging#155
Merged
fix(merger): handle non-sharded tensors in FSDP2 checkpoint merging#155
Conversation
Non-sharded buffers (e.g. inv_freq) are stored as plain Tensors rather than DTensors, causing AttributeError on _local_tensor access. Now falls back to using the tensor directly, and deduplicates identical copies across ranks instead of concatenating them.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Non-sharded buffers (e.g.
time_embedding.inv_freq) are stored as plainTensorrather thanDTensorin FSDP2 checkpoints, causingAttributeError: 'Tensor' object has no attribute '_local_tensor'during merge._local_tensorattribute before accessing it; fall back to using the tensor directly for non-sharded buffers