[dtensor] fix simplefsdp mixed-precision training bugs #154975
Closed
Conversation
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/154975
Note: Links to docs will display an error until the docs builds have been completed. ⏳ 1 Pending, 1 Unrelated FailureAs of commit c172c07 with merge base a7e496a ( UNSTABLE - The following job is marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
the
oncall: distributed
tianyu-l
added
release notes: distributed (dtensor)
Contributor
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
iupaikov-amd
pushed a commit
to ROCm/pytorch
that referenced
this pull request
Jun 4, 2025
This is a follow-up on the previous dtensor redistribute PR: pytorch#150740, which enables SimpleFSDP's mixed-precision training. In the most recent integration in TorchTitan: pytorch/torchtitan#1250, we found some discrepancies between SimpleFSDP's `fully_shard` and `replicate` modes when MPT is enabled. After debugging, I found the problem is in dtensor redistribute --`local_tensor` is taken out again from the original `input`. Thus, the dtensor used for communication has its original precision instead of using `forward_dtype`. This PR fixes this issue and corrects previously added test cases. After fixing the bug, the loss curves of `fully_shard` and `replicate` mode match perfectly.  Pull Request resolved: pytorch#154975 Approved by: https://github.com/tianyu-l
angelayi
pushed a commit
to angelayi/pytorch
that referenced
this pull request
Jun 5, 2025
This is a follow-up on the previous dtensor redistribute PR: pytorch#150740, which enables SimpleFSDP's mixed-precision training. In the most recent integration in TorchTitan: pytorch/torchtitan#1250, we found some discrepancies between SimpleFSDP's `fully_shard` and `replicate` modes when MPT is enabled. After debugging, I found the problem is in dtensor redistribute --`local_tensor` is taken out again from the original `input`. Thus, the dtensor used for communication has its original precision instead of using `forward_dtype`. This PR fixes this issue and corrects previously added test cases. After fixing the bug, the loss curves of `fully_shard` and `replicate` mode match perfectly.  Pull Request resolved: pytorch#154975 Approved by: https://github.com/tianyu-l
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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 follow-up on the previous dtensor redistribute PR: #150740, which enables SimpleFSDP's mixed-precision training.
In the most recent integration in TorchTitan: pytorch/torchtitan#1250, we found some discrepancies between SimpleFSDP's
fully_shardandreplicatemodes when MPT is enabled. After debugging, I found the problem is in dtensor redistribute --local_tensoris taken out again from the originalinput. Thus, the dtensor used for communication has its original precision instead of usingforward_dtype.This PR fixes this issue and corrects previously added test cases.
After fixing the bug, the loss curves of
fully_shardandreplicatemode match perfectly.cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k