[ddp] parameter verification #74113
Closed
[ddp] parameter verification #74113
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
Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 Differential Revision: [D34772067](https://our.internmc.facebook.com/intern/diff/D34772067/) [ghstack-poisoned]
rohan-varma
requested review from
H-Huang,
mingzhe09088,
mrshenli,
pritamdamania87 and
zhaojuanmao
as code owners
CI Flow Status⚛️ CI FlowRuleset - Version:
|
Contributor
🔗 Helpful links
💊 CI failures summary and remediationsAs of commit 62f0b92 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Please report bugs/suggestions to the (internal) Dr. CI Users group. |
facebook-github-bot
added
the
oncall: distributed
Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 Differential Revision: [D34772067](https://our.internmc.facebook.com/intern/diff/D34772067/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Mar 11, 2022
Pull Request resolved: #74113 Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 ghstack-source-id: 151159056 Differential Revision: [D34772067](https://our.internmc.facebook.com/intern/diff/D34772067/)
Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 Differential Revision: [D34772067](https://our.internmc.facebook.com/intern/diff/D34772067/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Mar 11, 2022
Pull Request resolved: #74113 Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 ghstack-source-id: 151191152 Differential Revision: [D34772067](https://our.internmc.facebook.com/intern/diff/D34772067/)
Closed
Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 Differential Revision: [D34772067](https://our.internmc.facebook.com/intern/diff/D34772067/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Mar 14, 2022
Pull Request resolved: #74113 Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 ghstack-source-id: 151275647 Differential Revision: [D34772067](https://our.internmc.facebook.com/intern/diff/D34772067/)
mrshenli
approved these changes
Mar 14, 2022
|
|
||
| // Broadcast and verify parameter size. | ||
| std::vector<at::Tensor> param_size_vec{param_size_tensor}; | ||
| process_group->broadcast(param_size_vec)->wait(); |
Contributor
There was a problem hiding this comment.
nit (can be done as future BE work item): DDP ctor will broadcast param values anyway. Is it possible to implement these two broadcasts in the same location using two async ops? I recall the param value broadcast was done in the distributed.py file?
Contributor
Author
There was a problem hiding this comment.
Yes, there's actually 3 collectives that go on now:
- parameter size verification
- parameter shape verification
- parameter broadcast
We could make these all async, and wait on all three, and register then() callbacks on the first two to raise appropriate errors. Filed #74185
| std::vector<at::Tensor> param_size_vec{param_size_tensor}; | ||
| process_group->broadcast(param_size_vec)->wait(); | ||
| auto res = param_size_tensor[0].item<int>(); | ||
| TORCH_CHECK( |
Contributor
There was a problem hiding this comment.
curious, does this mean some DDP process will crash, while other won't if their size matches? If that's the case, does it mean allgather might be better, as all processes can make the same decision here and throw the same error?
Contributor
Author
There was a problem hiding this comment.
This is a good point, it would be a lot better if all processes make the same decision. Next update will fix this
| group_gloo = dist.new_group( | ||
| timeout=timedelta(seconds=60), backend=dist.Backend.GLOO | ||
| ) | ||
| # Set NCCL_BLOCKING_WAIT and use a new NCCL group to improve test |
Contributor
There was a problem hiding this comment.
curious, why this can help improve determinism?
Contributor
Author
There was a problem hiding this comment.
We'd like NCCL_BLOCKING_WAIT so that rank 0 doesn't need to be taken down by async error handling which is non-deterministic when it kicks in and runs, and it also takes down the process which is hard to accomodate for in unittest.
As a result, we also need a new NCCL group as the default NCCL group is initialized before this test executes as a part of setup. I'll add some clarifying comments.
Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 Differential Revision: [D34772067](https://our.internmc.facebook.com/intern/diff/D34772067/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Mar 14, 2022
Pull Request resolved: #74113 Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 ghstack-source-id: 151319259 Differential Revision: [D34772067](https://our.internmc.facebook.com/intern/diff/D34772067/)
facebook-github-bot
pushed a commit
that referenced
this pull request
Mar 15, 2022
Summary: Pull Request resolved: #74113 Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks. Closes #73547 ghstack-source-id: 151319259 Test Plan: UT Reviewed By: mrshenli Differential Revision: D34772067 fbshipit-source-id: 456933111e9996823f1a220b474998e17fb74210
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):
Check mismatch in # of parameters by broadcasting and verifying from rank 0. As a result, non-zero ranks raise an error when # of parameters are mismatched across ranks.
Closes #73547
Differential Revision: D34772067