Fullgraph graph capture with dynamo. #159749
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/159749
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 32 PendingAs of commit 321e7c1 with merge base ee9f8ba ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
This pull request was exported from Phabricator. Differential Revision: D79534608 |
| def _backend(gm, example_inputs): | ||
| from torch._guards import TracingContext | ||
| artifacts.backend_inputs[gm._backend_id] = BackendInput(gm, example_inputs, TracingContext.get().fake_mode) | ||
| return gm |
There was a problem hiding this comment.
I would prefer for us to NOT actually implement it this way (I want to directly call into convert_frame and bypass the eval frame handler entirely)
There was a problem hiding this comment.
Yeah I can imagine this may be shadowed by some global context if we dont bypass eval frame things, is this the main issue? how do we get the frame object to feed to convert_frame if we bypass eval_frame, maybe sys.settrace()?
There was a problem hiding this comment.
Discussed offline with @ezyang, it seems essentially we want to make this API only about fullgraph=True which makes sense to me, so I changed the API to fullgraph_capture. eval_frame is not preferred for aesthetic reason and not strictly blocking and can be changed later as impl detail.
Summary: Following up on Avik's doc https://docs.google.com/document/d/11RW0Bbkp1QwFbEu8rCNW5d7wUFaEkxbL0uLyqcc2jTk/edit?tab=t.0 We are experimenting with a new API which utilizes torch.compile(fullgraph=True) and intend to use it to replace the old dynamo.export() API. This PR adds a prototype for the API described in the doc. Test Plan: test_misc -- -k test_aot_capture Rollback Plan: Differential Revision: D79534608
dd2c00a to
e4f60fb
Compare
|
This pull request was exported from Phabricator. Differential Revision: D79534608 |
Summary: Following up on Avik's doc https://docs.google.com/document/d/11RW0Bbkp1QwFbEu8rCNW5d7wUFaEkxbL0uLyqcc2jTk/edit?tab=t.0 We are experimenting with a new API which utilizes torch.compile(fullgraph=True) and intend to use it to replace the old dynamo.export() API. This PR adds a prototype for the API described in the doc. Test Plan: test_misc -- -k test_aot_capture Rollback Plan: Differential Revision: D79534608
e4f60fb to
ff51ba0
Compare
|
This pull request was exported from Phabricator. Differential Revision: D79534608 |
Summary: Following up on Avik's doc https://docs.google.com/document/d/11RW0Bbkp1QwFbEu8rCNW5d7wUFaEkxbL0uLyqcc2jTk/edit?tab=t.0 We are experimenting with a new API which utilizes torch.compile(fullgraph=True) and intend to use it to replace the old dynamo.export() API. This PR adds a prototype for the API described in the doc. Test Plan: test_misc -- -k test_aot_capture Rollback Plan: Differential Revision: D79534608
ff51ba0 to
d83007e
Compare
|
This pull request was exported from Phabricator. Differential Revision: D79534608 |
|
To unblock @tugsbayasgalan from export experiments, I will just leave eval_frame there for now since imo our focus is on defining the core data structure for this PR. @ezyang mind reviewing this again? |
Summary: Following up on Avik's doc https://docs.google.com/document/d/11RW0Bbkp1QwFbEu8rCNW5d7wUFaEkxbL0uLyqcc2jTk/edit?tab=t.0 We are experimenting with a new API which utilizes torch.compile(fullgraph=True) and intend to use it to replace the old dynamo.export() API. This PR adds a prototype for the API described in the doc. Test Plan: test_misc -- -k test_aot_capture Rollback Plan: Differential Revision: D79534608
d83007e to
321e7c1
Compare
|
This pull request was exported from Phabricator. Differential Revision: D79534608 |
torch/_dynamo/eval_frame.py
Outdated
| from torch._guards import TracingContext | ||
|
|
||
| backend_inputs[gm._backend_id] = BackendInput( | ||
| gm, example_inputs, TracingContext.get().fake_mode |
There was a problem hiding this comment.
Should we just do shape env?
There was a problem hiding this comment.
@tugsbayasgalan I think fake mode can be useful for other consumers. Export as one consumer can easily get shape_env by fake_mode.shape_env
| has a standard API to return compilation artifacts. | ||
| """ | ||
|
|
||
| dynamo_artifacts: DynamoCaptureOutput |
There was a problem hiding this comment.
How does dynamo artifacts relate to backend_inputs? Is there common backend_id?
There was a problem hiding this comment.
@tugsbayasgalan If you read inside DynamoCaptureOutput, you can find common backend_id inside backends field.
Summary: Following up on Avik's doc https://docs.google.com/document/d/11RW0Bbkp1QwFbEu8rCNW5d7wUFaEkxbL0uLyqcc2jTk/edit?tab=t.0 We are experimenting with a new API which utilizes torch.compile(fullgraph=True) and intend to use it to replace the old dynamo.export() API. This PR adds a prototype for the API described in the doc. Test Plan: test_misc -- -k test_aot_capture Rollback Plan: Reviewed By: tugsbayasgalan Differential Revision: D79534608
321e7c1 to
2dd4e6a
Compare
|
This pull request was exported from Phabricator. Differential Revision: D79534608 |
Summary:
Following up on Avik's doc https://docs.google.com/document/d/11RW0Bbkp1QwFbEu8rCNW5d7wUFaEkxbL0uLyqcc2jTk/edit?tab=t.0
We are experimenting with a new API which utilizes torch.compile(fullgraph=True) and intend to use it to replace the old dynamo.export() API.
This PR adds a prototype for the API described in the doc.
Test Plan:
test_misc -- -k test_aot_capture
Rollback Plan:
Differential Revision: D79534608
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @Lucaskabela