Make CUDA/AOTI partitioner composable after another delegate

[shoumikhin]

Summary:
AotiPartitioner.partition tagged every call_function node, including executorch_call_delegate calls already lowered by an earlier partitioner. So when CudaPartitioner runs as a second partitioner — e.g. after a TensorRT partition in a stacked .pte where TensorRT lowers the ops it can and the CUDA backend handles the rest — it tried to re-delegate the foreign delegate node, producing a malformed nested delegate. This is the blocker to composing the two backends in one .pte.

Tag only the non-lowered nodes, reusing the existing get_non_lowered_nodes helper (which already excludes executorch_call_delegate calls and their output getitems), so the partitioner claims just the remaining ops and composes cleanly after another backend. In the single-partitioner case there are no delegate nodes, so get_non_lowered_nodes returns every call_function and behavior is unchanged.

The same composition gap existed for constants: the final loop tagged every untagged param/buffer/lifted constant with this partition's tag, including ones consumed only by the foreign delegate. Backend lowering rejected those, since it requires every user of a tagged constant to share that tag while the foreign delegate's call keeps the prior one. Now only genuinely unused constants are tagged here — tag_constant_data already claims the ones this partition uses, and a constant feeding only a prior delegate is left untagged. Mirrored in fbcode and xplat.

Differential Revision: D107690797

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20077

• 📄 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 3d9d130 with merge base 0d904b6 ():
💚 Looks good so far! There are no failures yet. 💚

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

[meta-codesync]

@shoumikhin has exported this pull request. If you are a Meta employee, you can view the originating Diff in D107690797.

[Copilot]

Pull request overview

This PR fixes composability of the CUDA (AOTInductor/AOTI) partitioner when it runs after another backend has already lowered part of the graph into executorch_call_delegate calls, preventing CUDA from re-tagging those already-delegated nodes and producing malformed nested delegates.

Changes:

• Update AotiPartitioner.partition() to tag only “non-lowered” call_function nodes by using get_non_lowered_nodes(), skipping existing executorch_call_delegate calls and their output getitems.
• Add a CUDA partitioner unit test that splices a fake already-lowered delegate into an exported graph and verifies CUDA does not re-tag it while still tagging remaining ATen ops.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
backends/aoti/aoti_partitioner.py	Uses get_non_lowered_nodes() to avoid tagging already-lowered delegate nodes, enabling stacked backend composition.
backends/cuda/tests/test_cuda_partitioner.py	Adds regression coverage ensuring CUDA partitioning does not re-tag pre-existing executorch_call_delegate nodes.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[shoumikhin]

@pytorchbot label "release notes: none"

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.