[cuda] Preserve operations order between vectorized and non-vectorized in ln grad input #111488
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/111488
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (3 Unrelated Failures)As of commit 2a3da21 with merge base ba2ba96 ( FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
@pytorchbot label "ciflow/periodic" |
pytorch-bot
bot
added
the
ciflow/periodic
malfet
approved these changes
Oct 18, 2023
|
@pytorchbot merge |
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 |
xuhancn
pushed a commit
to xuhancn/pytorch
that referenced
this pull request
Nov 7, 2023
…d in ln grad input (pytorch#111488) The vectorized implementation in pytorch#111021 changed the order of arithmetic instructions in `layer_norm_grad_input`, causing non bitwise identical results when compared to the non-vectorized implementation. At merging, all accuracy checks passed, including internal inductor ones. There are CI periodic inductor dynamo tests (e.g. `pit_b_224`) that run eager mode models several times and compare results. If the input buffers are aligned to the vector length, the vectorized implementation will be used. If not, the default one will be used. If the 2 eager runs end up having different buffer alignments, 2 implementations will be called and then the results would be very close but not bitwise identical. The tests check for bitwise identical results and in some cases they may fail. This fix makes sure that the operation order between non-vectorized and vectorized is the same and the 2 implementations **should** produce bitwise identical results. Pull Request resolved: pytorch#111488 Approved by: https://github.com/malfet
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.
The vectorized implementation in #111021 changed the order of arithmetic instructions in
layer_norm_grad_input, causing non bitwise identical results when compared to the non-vectorized implementation. At merging, all accuracy checks passed, including internal inductor ones.There are CI periodic inductor dynamo tests (e.g.
pit_b_224) that run eager mode models several times and compare results. If the input buffers are aligned to the vector length, the vectorized implementation will be used. If not, the default one will be used. If the 2 eager runs end up having different buffer alignments, 2 implementations will be called and then the results would be very close but not bitwise identical. The tests check for bitwise identical results and in some cases they may fail.This fix makes sure that the operation order between non-vectorized and vectorized is the same and the 2 implementations should produce bitwise identical results.