[dynamo] change error_on_graph_break/fullgraph semantics#161747
Closed
williamwen42 wants to merge 8 commits intogh/williamwen42/284/basefrom
Closed
[dynamo] change error_on_graph_break/fullgraph semantics#161747williamwen42 wants to merge 8 commits intogh/williamwen42/284/basefrom
williamwen42 wants to merge 8 commits intogh/williamwen42/284/basefrom
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/161747
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 7a755fe with merge base 734ce8e ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Contributor
|
Why does this need to be a kwarg to compile? couldn't this just be a custom decorator? I think this will be really confusing for users. |
Collaborator
|
|
Member
Author
|
@mlazos the kwarg is not strictly necessary, but it does allow for a user to specify the initial setting without an additional decorator (e.g. Here's a table summary of fullgraph vs. error_on_graph_break 
|
Lucaskabela
reviewed
Aug 29, 2025
| in the function that it will optimize. If True, then we require that the entire function be | ||
| capturable into a single graph. If this is not possible (that is, if there are graph breaks), | ||
| then this will raise an error. | ||
| error_on_graph_break (bool): If `fullgraph` is set, then this arg does nothing |
Contributor
There was a problem hiding this comment.
once we have some comprehensive documentation on this, might be worth linking to an example that shows when users would want to use this setting
Contributor
|
My thinking on this is that if this isn't a super common use case, it shouldn't be a kwarg due to the confusion with fullgraph (which I think comparatively is pretty common). If this is a power-user API I think a decorator makes more sense. If you're okay with this could we hold off on the kwarg until we see how often it would be needed (e.g. seeing people add the decorator to the top-level function) |
Member
Author
|
@mlazos The idea going forward is that |
This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - `torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - `torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben Lucaskabela mlazos [ghstack-poisoned]
Member
Author
|
@mlazos I removed |
This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - `torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben Lucaskabela mlazos [ghstack-poisoned]
This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - `torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben Lucaskabela mlazos [ghstack-poisoned]
This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - ~`torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks~ - `error_on_graph_break` is a new internal `torch.compile `setting that is lower-priority than `fullgraph`. It allows the user to toggle erroring/continuing on graph breaks. - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben Lucaskabela mlazos [ghstack-poisoned]
Collaborator
There was a problem hiding this comment.
Based on the source code, this one should fail due to support.gc_collect()
def test_highly_nested_subclass(self):
# Issues 25395 and 35983: test that the trashcan mechanism works
# correctly for OrderedDict: deleting a highly nested OrderDict
# should not crash Python.
OrderedDict = self.OrderedDict
deleted = []
with torch._dynamo.set_fullgraph(fullgraph=False):
class MyOD(OrderedDict):
def __del__(self):
deleted.append(self.i)
obj = None
for i in range(100):
obj = MyOD([(None, obj)])
obj.i = i
del obj
support.gc_collect()
self.assertEqual(deleted, list(reversed(range(100))))
mlazos
reviewed
Sep 3, 2025
| return func(*args, **kwargs) | ||
|
|
||
| # test e2e, with Inductor, as smoketest. | ||
| @torch.compile(fullgraph=True, backend="inductor") |
Contributor
There was a problem hiding this comment.
Why this test? I'm curious haha
Member
Author
There was a problem hiding this comment.
It was a test that was added after set_fullgraph was merged - in that change, fullgraph=True went through the same convert_frame.py path as fullgraph=False, which increments counters["frames"]["total"]. This PR reverted that change so that the counter is no longer incremented.
mlazos
approved these changes
Sep 3, 2025
This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - ~`torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks~ - `error_on_graph_break` is a new internal `torch.compile `setting that is lower-priority than `fullgraph`. It allows the user to toggle erroring/continuing on graph breaks. - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben Lucaskabela mlazos [ghstack-poisoned]
This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - ~`torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks~ - `error_on_graph_break` is a new internal `torch.compile `setting that is lower-priority than `fullgraph`. It allows the user to toggle erroring/continuing on graph breaks. - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben Lucaskabela mlazos [ghstack-poisoned]
Member
Author
|
@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 |
markc-614
pushed a commit
to markc-614/pytorch
that referenced
this pull request
Sep 17, 2025
) This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - ~`torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks~ - `error_on_graph_break` is a new internal `torch.compile `setting that is lower-priority than `fullgraph`. It allows the user to toggle erroring/continuing on graph breaks. - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time Pull Request resolved: pytorch#161747 Approved by: https://github.com/mlazos ghstack dependencies: pytorch#161739
mansiag05
pushed a commit
to mansiag05/pytorch
that referenced
this pull request
Sep 22, 2025
) This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - ~`torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks~ - `error_on_graph_break` is a new internal `torch.compile `setting that is lower-priority than `fullgraph`. It allows the user to toggle erroring/continuing on graph breaks. - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time Pull Request resolved: pytorch#161747 Approved by: https://github.com/mlazos ghstack dependencies: pytorch#161739
cleonard530
pushed a commit
to cleonard530/pytorch
that referenced
this pull request
Sep 22, 2025
) This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - ~`torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks~ - `error_on_graph_break` is a new internal `torch.compile `setting that is lower-priority than `fullgraph`. It allows the user to toggle erroring/continuing on graph breaks. - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time Pull Request resolved: pytorch#161747 Approved by: https://github.com/mlazos ghstack dependencies: pytorch#161739
dsashidh
pushed a commit
to dsashidh/pytorch
that referenced
this pull request
Sep 26, 2025
) This PR implements the semantics change to `torch._dynamo.error_on_graph_break`: - ~`torch.compile` now has a new `error_on_graph_break` kwarg that serves as a lower-priority toggle for erroring/continuing on graph breaks~ - `error_on_graph_break` is a new internal `torch.compile `setting that is lower-priority than `fullgraph`. It allows the user to toggle erroring/continuing on graph breaks. - `error_on_graph_break` does nothing when `fullgraph=True` - `error_on_graph_break` does NOT guarantee a single graph Followup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation: - `fullgraph=True`: enforce one graph, no graph breaks, cannot be toggled - `fullgraph=False, error_on_graph_break=True`: errors on graph breaks, latter can be toggled during compile time - `fullgraph=False, error_on_graph_break=False`: resumes tracing on graph breaks, latter can be toggled during compile time Pull Request resolved: pytorch#161747 Approved by: https://github.com/mlazos ghstack dependencies: pytorch#161739
Khanaksahu
pushed a commit
to Khanaksahu/pytorch-fork
that referenced
this pull request
Nov 17, 2025
ghstack-source-id: aa2f19f Pull Request resolved: pytorch/pytorch#161747
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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 (oldest at bottom):
This PR implements the semantics change to
torch._dynamo.error_on_graph_break:torch.compilenow has a newerror_on_graph_breakkwarg that serves as a lower-priority toggle for erroring/continuing on graph breakserror_on_graph_breakis a new internaltorch.compilesetting that is lower-priority thanfullgraph. It allows the user to toggle erroring/continuing on graph breaks.error_on_graph_breakdoes nothing whenfullgraph=Trueerror_on_graph_breakdoes NOT guarantee a single graphFollowup [DONE]: need to change the programming model docs to reflect the 3 graph break modes for compilation:
fullgraph=True: enforce one graph, no graph breaks, cannot be toggledfullgraph=False, error_on_graph_break=True: errors on graph breaks, latter can be toggled during compile timefullgraph=False, error_on_graph_break=False: resumes tracing on graph breaks, latter can be toggled during compile timecc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @Lucaskabela @mlazos