Test FSDP with submodule non-reentrant checkpointing #89781
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
[ghstack-poisoned]
mrshenli
requested review from
H-Huang,
awgu,
kwen2501,
pritamdamania87,
rohan-varma and
zhaojuanmao
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/89781
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 73c7752: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
[ghstack-poisoned]
[ghstack-poisoned]
rohan-varma
reviewed
Nov 28, 2022
awgu
reviewed
Nov 28, 2022
Collaborator
awgu
left a comment
awgu
left a comment
There was a problem hiding this comment.
Thanks for adding these tests!
At a meta level, I am wondering how we should approach testing more interleavings in a systematic and complete way.
Also, I am not sure if you want to wait for the assert relaxation PR to land and then update test_checkpoint_submodule_reentrant() or not.
rohan-varma
approved these changes
Nov 28, 2022
Contributor
rohan-varma
left a comment
rohan-varma
left a comment
There was a problem hiding this comment.
LGTM, stamping to unblock. Will file a follow up issue to debug why this FSDP + AC structure does not work.
With combining FSDP with reentrant checkpointing, the post backward hook might run twice, and then hit [this error](https://github.com/pytorch/pytorch/blob/e20ec44544c17d6d3d411f88b870e05043bda731/torch/distributed/fsdp/_runtime_utils.py#L487). This is because reentrant backward uses nested autograd GraphTasks. The inner GraphTask is not aware of the outer one and therefore will flush pending `AccumulateGrad` invocations on exit, which in turn triggers the post backward hooks registered by FSDP. Later, the outer GraphTask will trigger that again, leading to the above error. PR #89791 relaxes the FSDP training state check, but we still run into grad value check failures occasionally. Therefore, this PR only lands the test for non-reentrant test, and we can enable the reentrant test when the accuracy issues are addressed. [ghstack-poisoned]
mrshenli
added a commit
that referenced
this pull request
Nov 29, 2022
With combining FSDP with reentrant checkpointing, the post backward hook might run twice, and then hit [this error](https://github.com/pytorch/pytorch/blob/e20ec44544c17d6d3d411f88b870e05043bda731/torch/distributed/fsdp/_runtime_utils.py#L487). This is because reentrant backward uses nested autograd GraphTasks. The inner GraphTask is not aware of the outer one and therefore will flush pending `AccumulateGrad` invocations on exit, which in turn triggers the post backward hooks registered by FSDP. Later, the outer GraphTask will trigger that again, leading to the above error. PR #89791 relaxes the FSDP training state check, but we still run into grad value check failures occasionally. Therefore, this PR only lands the test for non-reentrant test, and we can enable the reentrant test when the accuracy issues are addressed. ghstack-source-id: 8848c4c Pull Request resolved: #89781
Contributor
Author
Updated the PR summary to include that. Due to the grad value issue, the new tests are not testing that code path at the moment. |
Contributor
Author
|
@pytorchbot merge -g |
Collaborator
Merge startedYour change will be merged once all checks on your PR pass since you used the green (-g) flag (ETA: 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
kulinseth
pushed a commit
to kulinseth/pytorch
that referenced
this pull request
Dec 10, 2022
With combining FSDP with reentrant checkpointing, the post backward hook might run twice, and then hit [this error](https://github.com/pytorch/pytorch/blob/e20ec44544c17d6d3d411f88b870e05043bda731/torch/distributed/fsdp/_runtime_utils.py#L487). This is because reentrant backward uses nested autograd GraphTasks. The inner GraphTask is not aware of the outer one and therefore will flush pending `AccumulateGrad` invocations on exit, which in turn triggers the post backward hooks registered by FSDP. Later, the outer GraphTask will trigger that again, leading to the above error. PR pytorch#89791 relaxes the FSDP training state check, but we still run into grad value check failures occasionally. Therefore, this PR only lands the test for non-reentrant test, and we can enable the reentrant test when the accuracy issues are addressed. Pull Request resolved: pytorch#89781 Approved by: https://github.com/rohan-varma
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Stack from ghstack (oldest at bottom):
With combining FSDP with reentrant checkpointing, the post backward
hook might run twice, and then hit this
error.
This is because reentrant backward uses nested autograd GraphTasks.
The inner GraphTask is not aware of the outer one and therefore
will flush pending
AccumulateGradinvocations on exit, which inturn triggers the post backward hooks registered by FSDP. Later,
the outer GraphTask will trigger that again, leading to the above
error.
PR #89791 relaxes the FSDP training state check, but we still run
into grad value check failures occasionally. Therefore, this PR only
lands the test for non-reentrant test, and we can enable the
reentrant test when the accuracy issues are addressed.