Remove unused exported tensor constants after Cortex-M lowering by rascani · Pull Request #22862 · pytorch/executorch

rascani · 2026-09-15T23:04:01Z

Cortex-M layout conversion leaves original constant placeholders that FX dead-code elimination preserves. Add a separate RemoveUnusedConstantsPass to retire unused input specs and storage in linear time. Include it after constant lifting in both default pass lists, after padding fusion and scratch sizing.

Lift constants exactly once through the pass list before pruning, avoiding generated buffer-name collisions. Remove the unconditional final lifting call: explicit custom lists define the complete pipeline and can select no cleanup, lifting only, or lifting followed by pruning. The small lifting adapter makes the existing utility usable by the standard ExportedProgram pass runner.

Preserve user inputs, mutation and gradient targets, preserved module-call signatures, and state shared with other entry points. The existing EXIR pass and other backend defaults remain unchanged. Eleven focused Python tests passed, including buffer-name collision and all three custom-list behaviors.

Authored with AI assistance from Codex.

[ghstack-poisoned]

rascani · 2026-09-15T23:04:02Z

Stack from ghstack (oldest at bottom):

-> Remove unused exported tensor constants after Cortex-M lowering #22862

pytorch-bot · 2026-09-15T23:04:04Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22862

📄 Preview Python docs built from this PR

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

✅ No Failures

As of commit 6991dda with merge base 1cd22ed ():
💚 Looks good so far! There are no failures yet. 💚

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

[ghstack-poisoned]

Cortex-M layout conversion leaves original constant placeholders that FX dead-code elimination preserves. Add a separate RemoveUnusedConstantsPass to retire unused input specs and storage in linear time. Include it after constant lifting in both default pass lists, after padding fusion and scratch sizing. Lift constants exactly once through the pass list before pruning, avoiding generated buffer-name collisions. Remove the unconditional final lifting call: explicit custom lists define the complete pipeline and can select no cleanup, lifting only, or lifting followed by pruning. The small lifting adapter makes the existing utility usable by the standard ExportedProgram pass runner. Preserve user inputs, mutation and gradient targets, preserved module-call signatures, and state shared with other entry points. The existing EXIR pass and other backend defaults remain unchanged. Eleven focused Python tests passed, including buffer-name collision and all three custom-list behaviors. Authored with AI assistance from Codex. ghstack-source-id: d1c398d ghstack-comment-id: 5689274130 Pull-Request: #22862

[ghstack-poisoned]

Cortex-M layout conversion leaves original constant placeholders that FX dead-code elimination preserves. Add a separate RemoveUnusedConstantsPass to retire unused input specs and storage in linear time. Include it after constant lifting in both default pass lists, after padding fusion and scratch sizing. Lift constants exactly once through the pass list before pruning, avoiding generated buffer-name collisions. Remove the unconditional final lifting call: explicit custom lists define the complete pipeline and can select no cleanup, lifting only, or lifting followed by pruning. The small lifting adapter makes the existing utility usable by the standard ExportedProgram pass runner. Preserve user inputs, mutation and gradient targets, preserved module-call signatures, and state shared with other entry points. The existing EXIR pass and other backend defaults remain unchanged. Eleven focused Python tests passed, including buffer-name collision and all three custom-list behaviors. Authored with AI assistance from Codex. ghstack-source-id: e825659 ghstack-comment-id: 5689274130 Pull-Request: #22862

github-actions · 2026-09-18T18:27:33Z

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.

Detach the graph, signature, and state dictionary before the Cortex-M lifting adapter invokes the in-place lifting utility. This preserves the original program and prevents a generated buffer name from overwriting another entry point's shared state. Regression coverage exercises lift-only and lift-then-prune through both the Cortex-M manager and edge.transform(). The 19 focused tests and four MLPerf Tiny dialect tests pass. Authored with AI assistance from Codex. [ghstack-poisoned]

Cortex-M layout conversion leaves original constant placeholders that FX dead-code elimination preserves. Add a separate RemoveUnusedConstantsPass to retire unused input specs and storage in linear time. Include it after constant lifting in both default pass lists, after padding fusion and scratch sizing. Lift constants exactly once through the pass list before pruning, avoiding generated buffer-name collisions. Remove the unconditional final lifting call: explicit custom lists define the complete pipeline and can select no cleanup, lifting only, or lifting followed by pruning. The small lifting adapter makes the existing utility usable by the standard ExportedProgram pass runner. Preserve user inputs, mutation and gradient targets, preserved module-call signatures, and state shared with other entry points. The existing EXIR pass and other backend defaults remain unchanged. Eleven focused Python tests passed, including buffer-name collision and all three custom-list behaviors. Authored with AI assistance from Codex. ghstack-source-id: 3f25b1c ghstack-comment-id: 5689274130 Pull-Request: #22862

[ghstack-poisoned]

Cortex-M layout conversion leaves original constant placeholders that FX dead-code elimination preserves. Add a separate RemoveUnusedConstantsPass to retire unused input specs and storage in linear time. Include it after constant lifting in both default pass lists, after padding fusion and scratch sizing. Lift constants exactly once through the pass list before pruning, avoiding generated buffer-name collisions. Remove the unconditional final lifting call: explicit custom lists define the complete pipeline and can select no cleanup, lifting only, or lifting followed by pruning. The small lifting adapter makes the existing utility usable by the standard ExportedProgram pass runner. Preserve user inputs, mutation and gradient targets, preserved module-call signatures, and state shared with other entry points. The existing EXIR pass and other backend defaults remain unchanged. Eleven focused Python tests passed, including buffer-name collision and all three custom-list behaviors. Authored with AI assistance from Codex. ghstack-source-id: f68d6ae ghstack-comment-id: 5689274130 Pull-Request: #22862

rascani · 2026-09-23T19:57:59Z

Bump on this one @JakeStevens & @AdrianLundell

rascani added 4 commits September 15, 2026 16:03

Update

f420caa

[ghstack-poisoned]

Update

57305bb

[ghstack-poisoned]

Update

4859b8e

[ghstack-poisoned]

Update

7bc14d3

[ghstack-poisoned]

This was referenced Sep 15, 2026

Cortex-M: avoid redundant convolution and pooling work #22859

Merged

Cortex-M: size scratch buffers in a dedicated pass #22860

Merged

Cortex-M: fuse SAME padding into quantized convolutions #22861

Merged

rascani deployed to cadence September 15, 2026 23:04 — with GitHub Actions Active

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 Sep 15, 2026

rascani deployed to cadence September 15, 2026 23:23 — with GitHub Actions Active

Preserve the original program when pruning unused constants

a698091

[ghstack-poisoned]

rascani deployed to cadence September 16, 2026 17:47 — with GitHub Actions Active

rascani requested a review from AdrianLundell September 16, 2026 17:51

rascani marked this pull request as ready for review September 16, 2026 17:51

rascani requested review from digantdesai and psiddh September 16, 2026 17:51

rascani deployed to cadence September 16, 2026 18:09 — with GitHub Actions Active

rascani added 2 commits September 16, 2026 13:22

Rebase remaining Cortex-M cleanup PRs after the first two landed

c400265

[ghstack-poisoned]

Rebase remaining Cortex-M cleanup PRs after the first two landed

cc18f11

[ghstack-poisoned]

rascani deployed to cadence September 16, 2026 20:23 — with GitHub Actions Active

rascani deployed to cadence September 16, 2026 20:39 — with GitHub Actions Active

Base automatically changed from gh/rascani/44/head to main September 16, 2026 21:51

Merge branch 'main' into gh/rascani/45/head

ed50df6

rascani deployed to cadence September 18, 2026 18:27 — with GitHub Actions Active

rascani deployed to cadence September 18, 2026 18:44 — with GitHub Actions Active

JakeStevens reviewed Sep 18, 2026

View reviewed changes

Comment thread backends/cortex_m/passes/cortex_m_pass_manager.py

rascani deployed to cadence September 18, 2026 21:37 — with GitHub Actions Active

rascani deployed to cadence September 18, 2026 21:57 — with GitHub Actions Active

Rebase unused constants cleanup onto current main

dbcb05d

[ghstack-poisoned]

rascani deployed to cadence September 22, 2026 18:35 — with GitHub Actions Active

rascani deployed to cadence September 22, 2026 18:53 — with GitHub Actions Active

Merge branch 'main' into gh/rascani/45/head

6991dda

rascani deployed to cadence September 23, 2026 16:37 — with GitHub Actions Active

rascani deployed to cadence September 23, 2026 16:56 — with GitHub Actions Active

Navigation Menu

Remove unused exported tensor constants after Cortex-M lowering - #22862

Remove unused exported tensor constants after Cortex-M lowering#22862
rascani wants to merge 11 commits into
mainfrom
gh/rascani/45/head

rascani commented Sep 15, 2026

Uh oh!

rascani commented Sep 15, 2026 •

edited

Loading

Uh oh!

pytorch-bot Bot commented Sep 15, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Sep 18, 2026

Uh oh!

Uh oh!

rascani commented Sep 23, 2026

Uh oh!

This branch was successfully deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

rascani commented Sep 15, 2026

Uh oh!

rascani commented Sep 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot Bot commented Sep 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22862

✅ No Failures

Uh oh!

github-actions Bot commented Sep 18, 2026

This PR needs a release notes: label

Uh oh!

Uh oh!

rascani commented Sep 23, 2026

Uh oh!

This branch was successfully deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

rascani commented Sep 15, 2026 •

edited

Loading

pytorch-bot Bot commented Sep 15, 2026 •

edited

Loading

This PR needs a `release notes:` label