PR #22437: Added frontend attribute handling to explicit_stream_annotation_async_wrapper #23021
+90
−103
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
![copybara-service[bot]](https://avatars.githubusercontent.com/in/44061?s=60&v=4){kind=small}
copybara-service
bot
force-pushed
the
test_730394376
branch
2 times, most recently
from
9d45ccd to
dcf6888
Compare
…ation_async_wrapper Imported from GitHub PR #22437 This is a small change that ensures the frontend attributes are correctly passed to both the `async-start` and `async-done` created pairs. This also clears the scheduling attributes that are directly on the call operation and inner ops. The specific goal of this change is to have stable support combining the scheduling group ids with stream annotation in JAX. ```python with set_xla_metadata(_scheduling_group_id=1): result = compute_on("gpu_stream:1")(jitted_func)(...) ``` Currently, the issue stems from the `set_xla_metadata` context manager, which will apply the frontend attribute to all operations, including the ones within our `jitted_func`. When the same scheduling annotations is found in two `HloComputation`s, an error is raised in `LegalizeSchedulingAnnotations`. This is intended to avoid hitting this check, and cleaning up the annotations on the wrapped streamed computation. Copybara import of the project: -- 994c2ee by chaser <[email protected]>: Added frontend attributed handling -- 9db58b2 by chaser <[email protected]>: Added clearing of scheduling annotations -- a83e32a by chaser <[email protected]>: Added HloInstruction.erase_frontend_attribute Merging this change closes #22437 Reverts 91ca1f1 FUTURE_COPYBARA_INTEGRATE_REVIEW=#22437 from chaserileyroberts:chase/frontend_forward_async_wrapper a83e32a PiperOrigin-RevId: 730394376
copybara-service
bot
force-pushed
the
test_730394376
branch
from
dcf6888 to
45984a1
Compare
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.
PR #22437: Added frontend attribute handling to explicit_stream_annotation_async_wrapper
Imported from GitHub PR #22437
This is a small change that ensures the frontend attributes are correctly passed to both the
async-startandasync-donecreated pairs. This also clears the scheduling attributes that are directly on the call operation and inner ops.The specific goal of this change is to have stable support combining the scheduling group ids with stream annotation in JAX.
Currently, the issue stems from the
set_xla_metadatacontext manager, which will apply the frontend attribute to all operations, including the ones within ourjitted_func. When the same scheduling annotations is found in twoHloComputations, an error is raised inLegalizeSchedulingAnnotations. This is intended to avoid hitting this check, and cleaning up the annotations on the wrapped streamed computation.Copybara import of the project:
--
994c2ee by chaser [email protected]:
Added frontend attributed handling
--
9db58b2 by chaser [email protected]:
Added clearing of scheduling annotations
--
a83e32a by chaser [email protected]:
Added HloInstruction.erase_frontend_attribute
Merging this change closes #22437
Reverts 91ca1f1
FUTURE_COPYBARA_INTEGRATE_REVIEW=#22437 from chaserileyroberts:chase/frontend_forward_async_wrapper a83e32a