[DDP] Test inference works with eval() and no_grad() #59666
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
Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/) [ghstack-poisoned]
rohan-varma
requested review from
H-Huang,
cbalioglu,
mrshenli,
pritamdamania87,
wayi1 and
zhaojuanmao
as code owners
💊 CI failures summary and remediationsAs of commit 3364263 (more details on the Dr. CI page and at hud.pytorch.org/pr/59666):
1 failure not recognized by patterns:
ci.pytorch.org: 1 failedThis comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions to the (internal) Dr. CI Users group. |
facebook-github-bot
added
oncall: distributed
rohan-varma
added a commit
that referenced
this pull request
Jun 8, 2021
Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/) ghstack-source-id: 130878421 Pull Request resolved: #59666
|
cc @mrshenli |
This was referenced Jun 8, 2021
Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Jun 11, 2021
Pull Request resolved: #59666 Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. ghstack-source-id: 131262081 Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/)
wayi1
reviewed
Jun 12, 2021
| if self.rank == 0: | ||
| with torch.no_grad(): | ||
| for _ in range(6): | ||
| ddp_out = model(inp) |
There was a problem hiding this comment.
Nit: If you rename model as ddp_model, then you can just use one line here:
self.assertEqual(ddp_model(inp), local_model(inp))
Just more concise. It's optional.
wayi1
reviewed
Jun 12, 2021
| # or eval setting and there is no hang. | ||
| rank = self.rank | ||
| torch.cuda.set_device(rank) | ||
| model = Net().cuda() |
There was a problem hiding this comment.
Nit: I believe you can just create a for loop over two tuples of <model, input>, where the models are Net().cuda() and nn.SyncBatchNorm. This can save some duplicate code and improve the readability.
wayi1
reviewed
Jun 12, 2021
| self.assertEqual(ddp_out, local_out) | ||
| torch.cuda.synchronize() | ||
|
|
||
| self._barrier(timeout=30) |
There was a problem hiding this comment.
Are all the 3 barriers here necessary, even after cuda.sync? Why do they need a non-default higher timeout here?
There was a problem hiding this comment.
I don't think the calls to synchronize are necessary actually, will remove those.
The calls to barrier are there because we're only running inference on rank 0, if the inference unexpectedly takes too log, the default barrier which I believe has a timeout of 10s can timeout, leading to a false positive failure.
In this test the inference does not take nearly the full 30s but wanted to have plenty of buffer to avoid flakiness.
Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Jun 16, 2021
Pull Request resolved: #59666 Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. ghstack-source-id: 131561892 Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/)
Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Jun 17, 2021
Pull Request resolved: #59666 Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. ghstack-source-id: 131723578 Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/)
Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Jun 17, 2021
Pull Request resolved: #59666 Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. ghstack-source-id: 131749203 Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/)
wayi1
reviewed
Jun 18, 2021
| device_ids=[rank] | ||
| ) | ||
| inp = torch.randn(10, 2, device=rank) | ||
| inp_syncbn = torch.randn(10, 2, 4, 4, device=rank) |
There was a problem hiding this comment.
What's the abbreviation of "inp"?
There was a problem hiding this comment.
it is short for "input"
wayi1
approved these changes
Jun 18, 2021
Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Jun 19, 2021
Pull Request resolved: #59666 Tests that inference with DDP model won't hang when user sets eval() or no_grad(). Note that if the model has a syncBN layer, they need both eval() and no_grad() as eval() makes SyncBN work like a regular BN layer. ghstack-source-id: 131906625 Differential Revision: [D28974146](https://our.internmc.facebook.com/intern/diff/D28974146/)
|
This pull request has been merged in 0131a59. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
Stack from ghstack:
Tests that inference with DDP model won't hang when user sets eval()
or no_grad(). Note that if the model has a syncBN layer, they need both eval()
and no_grad() as eval() makes SyncBN work like a regular BN layer.
Differential Revision: D28974146