[executorch][arm] Speed up lowering: cache constant names, fix quadratic permute chain collection, dedup producer queue, defer cleanup, and drop redundant recompiles

[apullin]

Summary:

This is a stack of 5 commits that speed up various parts of the model lowering process. I have been testing with ~1M param models, and lowering time has been 60+ minutes. This is an unfortunate reality from being implemented in Python, and ET will always struggle with networks of even moderate size BUT there is still plenty of performance left on the table.

The summary for each commit/diff are as follow:

---

D114790177:
Remove redundant GraphModule.recompile() at end of PropagateViewCopyPermutePass.call().

When pass modifies graph it already calls _retrace(), which eliminates dead code, lints, and runs ARM ExportPass retracing to return rebuilt graph module. Immediately following recompile() recompiles Python code not needed before pass manager continues, adding wall time.

Keep _retrace() and drop extra compile step.

---

D114790155:
Drop redundant recompile() before ARM pass retracing.
Remove redundant GraphModule.recompile() calls immediately before super().call(graph_module) in CanonicalizeViewCopyPermutePass and FuseDuplicateUsersPass. The following ExportPass retrace/interpreter does not consume compiled Python code object, so recompile adds wall time without changing graph.
Keep graph.lint() in modified paths for invariant checking. Behavior-preserving lowering speedup.

---

D114790128:
Skip redundant recompile in identical-input transform fusion.
Remove redundant GraphModule.recompile() immediately before super().call() in FuseIdenticalInputTransformsPass. Pass already eliminates dead code and lints before handing to ARM framework for retracing. Recompiling Python GraphModule at that point is unnecessary because ExportPass call reconstructs/interprets graph rather than relying on just-compiled code object. Behavior-preserving speedup.

---

D114224762:
Speed up permute propagation cleanup.
ARM lowering spends significant time in permute propagation. Reduce avoidable repeated work while keeping same optimizations:
Canonicalize view/permute chain collection now uses deque + membership set instead of list pop(0) and remove(), avoiding quadratic bookkeeping.
FuseDuplicateUsersPass deduplicates pending producer revisits while preserving same revisit behavior after fusions.
PropagateViewCopyPermutePass still retraces after each moved transform for metadata safety, but defers horizontal/vertical cleanup until full scan finds no more direct propagation moves. Preserves fixed-point behavior while avoiding expensive cleanup after every single moved transform.
Behavior-preserving speedup, all existing pass tests pass.

---

D114224782:
RemovePermutesAroundElementwiseTosaOps._is_constant called is_param_node() for every node reached during recursive visit() walk. Each is_param_node rebuilds immutable dict over full input_specs via uncached graph_signature properties. This made pass O(nodes * inputs). Precompute union of param/buffer/lifted-constant placeholder names once in init and use O(1) set membership. Semantics unchanged.

---

Differential Revision: D114224782 D114224762 D114790128 D114790155 D114790177

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21696

• 📄 Preview Python docs built from this PR

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

❌ 2 New Failures, 1 Unrelated Failure

As of commit 1c91383 with merge base 730b77a ():

NEW FAILURES - The following jobs have failed:

• Cadence Build & Test / hifi-build / hifi4 (gh)
  ##[error]Refusing to check out fork pull request code from a 'pull_request_target' workflow. This workflow runs with the base repository's GITHUB_TOKEN, secrets, default-branch cache scope, and runner access. Fetching and executing a fork's code in that trusted context commonly leads to "pwn request" vulnerabilities. To opt in, review the risks at https://gh.io/securely-using-pull_request_target and set 'allow-unsafe-pr-checkout: true' on the actions/checkout step.
• Cadence Build & Test / vision-build / vision (gh)
  ##[error]Refusing to check out fork pull request code from a 'pull_request_target' workflow. This workflow runs with the base repository's GITHUB_TOKEN, secrets, default-branch cache scope, and runner access. Fetching and executing a fork's code in that trusted context commonly leads to "pwn request" vulnerabilities. To opt in, review the risks at https://gh.io/securely-using-pull_request_target and set 'allow-unsafe-pr-checkout: true' on the actions/checkout step.

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• pull / test-qnn-testsuite-linux / test-backend-linux (qnn, models) / linux-job (gh) (matched linux rule in flaky-rules.json)
  The runner has received a shutdown signal. This can happen when the runner service is stopped, or a manually started runner is canceled.

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

[meta-codesync]

@apullin has exported this pull request. If you are a Meta employee, you can view the originating Diff in D114790177.

[github-actions]

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.