Skip to content

Match softmax input storage and layout to its output (#24437) - #24437

Open
JakeStevens wants to merge 1 commit into
pytorch:mainfrom
JakeStevens:export-D129088170
Open

JakeStevens wants to merge 1 commit into
pytorch:mainfrom
JakeStevens:export-D129088170

Conversation

@JakeStevens

@JakeStevens JakeStevens commented Nov 2, 2026 •

Copy link
Copy Markdown
Contributor

Summary:

Ulterior Motive

Run native Swin attention softmaxes across buffer and texture storage plans.

Rationale

What: Repack the input of _softmax and _log_softmax when its storage or layout differs from the planned output.

Why: The kernel reads its input using the output's storage type and texture layout. A buffer bound to a texture shader can crash; a differently packed texture can return misplaced values.

Details

  • Adapt only the self input of these two operators.
  • Reuse the existing repack helper and leave matching representations unchanged.
  • Cover buffer-to-texture and width-to-channels-packed inputs with independent softmax references.

Authored with AI using Metamate on agentcloud.

Differential Revision: D129088170

@pytorch-bot

pytorch-bot Bot commented Nov 2, 2026 •

Copy link
Copy Markdown

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24437

Note: Links to docs will display an error until the docs builds have been completed.

❌ 9 Pending, 2 Unclassified Failures

As of commit 20e83ad with merge base cb3f24e (image):

UNCLASSIFIED FAILURES - DrCI could not classify the following jobs because the workflow did not run on the merge base. The failures may be pre-existing on trunk or introduced by this PR:

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Nov 2, 2026
@meta-codesync

meta-codesync Bot commented Nov 2, 2026

Copy link
Copy Markdown
Contributor

@JakeStevens has exported this pull request. If you are a Meta employee, you can view the originating Diff in D129088170.

@github-actions

github-actions Bot commented Nov 2, 2026

Copy link
Copy Markdown

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

Summary:

## Ulterior Motive

Run native Swin attention softmaxes across buffer and texture storage plans.

## Rationale

**What**: Repack the input of `_softmax` and `_log_softmax` when its storage or layout differs from the planned output.

**Why**: The kernel reads its input using the output's storage type and texture layout. A buffer bound to a texture shader can crash; a differently packed texture can return misplaced values.

## Details

- Adapt only the `self` input of these two operators.
- Reuse the existing repack helper and leave matching representations unchanged.
- Cover buffer-to-texture and width-to-channels-packed inputs with independent softmax references.

Authored with AI using Metamate on agentcloud.

Differential Revision: D129088170
@meta-codesync meta-codesync Bot changed the title Match softmax input storage and layout to its output Match softmax input storage and layout to its output (#24437) Nov 2, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. meta-exported

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant