Use fp32 accumulation in portable/optimized norms #16727
Merged
+83
−14
Conversation
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/16727
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (6 Unrelated Failures)As of commit 9386079 with merge base 8e7d761 ( BROKEN TRUNK - The following jobs 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. |
Contributor
|
@GregoryComer has exported this pull request. If you are a Meta employee, you can view the originating Diff in D91096147. |
meta-cla
bot
added
the
CLA Signed
GregoryComer
added
the
release notes: ops & kernels
GregoryComer
force-pushed
the
export-D91096147
branch
from
1a89352 to
dfd8f9a
Compare
GregoryComer
added a commit
to GregoryComer/executorch
that referenced
this pull request
Jan 21, 2026
Summary: When running in fp16, our layer norm implementation diverges from ATen, especially for larger variances. This is because ATen internally accumulates mean and variance in fp32 for fp16/bf16, whereas we accumulate in the input/output dtype. To resolve this, I've matched the ATen behavior by using acc_t (similar to OpMathType from ATen) in optimized. In portable, I've just removed the cast back to CTYPE, as the accumulation logic was already in float. I've added an additional test case with high variance. In combination with the existing coverage, this should show that the behavior is unchanged for existing cases and no longer overflows or diverges for larger inputs. Differential Revision: D91096147
Summary: When running in fp16, our layer norm implementation diverges from ATen, especially for larger variances. This is because ATen internally accumulates mean and variance in fp32 for fp16/bf16, whereas we accumulate in the input/output dtype. To resolve this, I've matched the ATen behavior by using acc_t (similar to OpMathType from ATen) in optimized. In portable, I've just removed the cast back to CTYPE, as the accumulation logic was already in float. There was also a straight up bug where the element count was being cast to CTYPE in group norm, which tends to overflow or lose precision. I've added an additional test case with high variance. In combination with the existing coverage, this should show that the behavior is unchanged for existing cases and no longer overflows or diverges for larger inputs. Thanks Jason Zhu for spotting this! Differential Revision: D91096147
GregoryComer
force-pushed
the
export-D91096147
branch
from
dfd8f9a to
9386079
Compare
GregoryComer
commented
Jan 21, 2026
| struct ComputeDTypeTraits<int8_t> { | ||
| using type = int32_t; | ||
| }; | ||
| // For 16 bit float types, ops should perform internal math in float32. |
Member
Author
There was a problem hiding this comment.
This matches the ATen convention from OpMathType.h.
GregoryComer
commented
Jan 21, 2026
| // compute E[X] and Var[x] = E[x^2] - E[x]^2 | ||
| CTYPE sum = reduce_add(x, static_cast<CTYPE>(inner_size)); | ||
| CTYPE sq_sum = vec_powerf(x, static_cast<CTYPE>(inner_size)); | ||
| float sum = reduce_add(x, inner_size); |
Member
Author
There was a problem hiding this comment.
Note: We should ideally use float64 if CTYPE is double, but we're already forcing float32 in the various utility functions - reduce_add and vec_powerf return floats and use floats internally. I'm just removing the implicit cast to CTYPE on the return value.
I'm inclined to leave this for a follow-up on an as-needed basis.
GregoryComer
changed the title
Member
Author
|
Note that unit tests are failing on trunk. Checking the job logs on this PR shows it's an unrelated failure. |
larryliu0820
approved these changes
Jan 21, 2026
mergennachin
approved these changes
Jan 21, 2026
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
Summary:
When running in fp16, our layer norm implementation diverges from ATen, especially for larger variances. This is because ATen internally accumulates mean and variance in fp32 for fp16/bf16, whereas we accumulate in the input/output dtype.
To resolve this, I've matched the ATen behavior by using acc_t (similar to OpMathType from ATen) in optimized. In portable, I've just removed the cast back to CTYPE, as the accumulation logic was already in float.
There was also a straight up bug where the element count was being cast to CTYPE in group norm, which tends to overflow or lose precision.
I've added an additional test case with high variance. In combination with the existing coverage, this should show that the behavior is unchanged for existing cases and no longer overflows or diverges for larger inputs.
Thanks Jason Zhu for spotting this!
Differential Revision: D91096147