xm.save() should not set sync_xla_data=True when sync'ing. #8484
+25
−2
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
jeffhataws
requested review from
will-cromar,
jonb377,
tengyifei and
ManfeiBai
mcuiaws
force-pushed
the
fix_8422
branch
2 times, most recently
from
c2af6fa to
3d48f2a
Compare
jeffhataws
approved these changes
Dec 12, 2024
tengyifei
approved these changes
Dec 17, 2024
| @@ -143,6 +144,28 @@ def try_grad_accum(model, device, train_x, train_label, accum_steps): | |||
| alias_count == 1.0 | |||
| ), f"Expect 1 input-output alias pair for gradient accumulation, got {alias_count}" | |||
|
|
|||
| def test_xm_save_no_aliasing(self): | |||
There was a problem hiding this comment.
I think this test is added in https://github.com/pytorch/xla/pull/8467/files already
Setting sync_xla_data=True performs tensor graph sync as if it's a mark step, which triggers buffer aliasing to be performed. However, it's not safe to do so unless all live tensors are being sync'd. Also fix torch_xla.utils.serialization.save() which has the same issue. This fixes pytorch#8422
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Outdated suggestions cannot be applied.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.
Setting sync_xla_data=True performs tensor graph sync as if it's a mark step, which triggers buffer aliasing to be performed. However, according to the comments in https://github.com/pytorch/xla/blame/v2.5.1/torch_xla/csrc/xla_graph_executor.cpp#L1336, it's not safe to do so unless all live tensors are being sync'd.
This fixes #8422