[dynamo] fix dynamo + DTensor to work with 2d #108329
Closed
Conversation
This file contains 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
This PR fixes the dynamo + DTensor integration so that the current graph break FSDP can work with tensor parallel by moving the torch.compile after FSDP wrapping [ghstack-poisoned]
wanchaol
requested review from
mrshenli,
zhaojuanmao,
rohan-varma,
H-Huang,
awgu,
kwen2501,
fegin,
fduwjj and
wz337
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/108329
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 2452522 with merge base e68b3ad ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
wanchaol
added a commit
that referenced
this pull request
Aug 31, 2023
This PR fixes the dynamo + DTensor integration so that the current graph break FSDP can work with tensor parallel by moving the torch.compile after FSDP wrapping ghstack-source-id: 1922a173c9b148a8a6bfe19e820bbebd531435dd Pull Request resolved: #108329
pair debugged with wconstab and we found some issue in both dynamo and the TP's fsdp extension side. This PR fixes the dynamo + DTensor integration so that the current graph break FSDP can work with tensor parallel by moving the torch.compile after FSDP wrapping. [ghstack-poisoned]
wanchaol
added a commit
that referenced
this pull request
Aug 31, 2023
pair debugged with @wconstab and we found some issue in both dynamo and the TP's fsdp extension side. This PR fixes the dynamo + DTensor integration so that the current graph break FSDP can work with tensor parallel by moving the torch.compile after FSDP wrapping. ghstack-source-id: 1922a173c9b148a8a6bfe19e820bbebd531435dd Pull Request resolved: #108329
wanchaol
added
release notes: distributed (dtensor)
pair debugged with wconstab and we found some issue in both dynamo and the TP's fsdp extension side. This PR fixes the dynamo + DTensor integration so that the current graph break FSDP can work with tensor parallel by moving the torch.compile after FSDP wrapping. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx chenyang78 aakhundov [ghstack-poisoned]
wanchaol
added a commit
that referenced
this pull request
Aug 31, 2023
pair debugged with @wconstab and we found some issue in both dynamo and the TP's fsdp extension side. This PR fixes the dynamo + DTensor integration so that the current graph break FSDP can work with tensor parallel by moving the torch.compile after FSDP wrapping. ghstack-source-id: 4c26e2721a36a92d0988044f4c4fdc7491dc6dfd Pull Request resolved: #108329
Skylion007
approved these changes
Aug 31, 2023
wconstab
approved these changes
Aug 31, 2023
| process_group=fsdp_pg, | ||
| device_id=self.rank, | ||
| use_orig_params=True, | ||
| ) | ||
|
|
||
| # TODO: once aot autograd support is ready we can just use default backend | ||
| compiled_2d = torch.compile(fsdp_2d, backend="eager") |
There was a problem hiding this comment.
nit: we might also want to test inductor backend, as it does break sometimes more than eager, due to customized logic for some parts of aot-autograd
torch/_dynamo/variables/torch.py
Outdated
| @@ -572,6 +572,7 @@ def get_state_from_generator(): | |||
| elif is_from_local(self.value): | |||
| # rewrite non-primitive args/kwargs to be included in the on-the-fly prim function | |||
| # and rewrite args to have only proxyable args, then insert call_function | |||
| # TODO: support cases where device_mesh + placements specified as kwargs | |||
There was a problem hiding this comment.
mby good to assert kwargs is empty (or that it just contains the 1 bool you expect, etc)
There was a problem hiding this comment.
Actually it's not quite hard to directly support kwargs so I added that directly
pair debugged with wconstab and we found some issue in both dynamo and the TP's fsdp extension side. This PR fixes the dynamo + DTensor integration so that the current graph break FSDP can work with tensor parallel by moving the torch.compile after FSDP wrapping. cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx chenyang78 aakhundov [ghstack-poisoned]
wanchaol
added a commit
that referenced
this pull request
Aug 31, 2023
pair debugged with @wconstab and we found some issue in both dynamo and the TP's fsdp extension side. This PR fixes the dynamo + DTensor integration so that the current graph break FSDP can work with tensor parallel by moving the torch.compile after FSDP wrapping. ghstack-source-id: 4f48e003224e0d48e32b3db57923859d91b50e0e Pull Request resolved: #108329
|
@pytorchbot merge |
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 |
|
Nice! |
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):
pair debugged with @wconstab and we found some issue in both dynamo and
the TP's fsdp extension side. This PR fixes the dynamo + DTensor integration
so that the current graph break FSDP can work with tensor parallel by moving
the torch.compile after FSDP wrapping.
cc @voznesenskym @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @chenyang78 @aakhundov