NXP backend: Fix bug when a partition output is also used by nodes inside the partition.

[MartinPavella]

Summary

[De]quantize nodes that form QDQ clusters are handled by turning float tensors to int8 and assigning to them the quantization parameters in the Neutron IR. But these operators cannot just disappear because the graph would be disconnected (there would be "holes" in it). Usually we can redirect the outputs of these operators to their inputs, and if this is not possible, we can convert the operators to identity operators, which are optimized out later.

Up until a recent PR, dequantize nodes were always being turned into identity ops (not skipped), which caused issues in optimization passes. This is because the QDQ nodes are converted before the compute nodes. So for example a dequantize that's near the end of the model would be converted to a noop Transpose and inserted into the operators list before the compute ops which topologically come before it. So the topological ordering of the graph was broken.

The recent PR made it so that only the dequantize nodes that consume model inputs are turned to identity, and the rest is skipped (fixing the topological order issue). That PR however failed to identify an edge case where if a node is both the output of a partition and it is also used as input to other nodes inside that partition, the dequantize node that comes after it cannot be skipped, and it must be turned into identity. Therefore, we cannot avoid breaking the topological order.

For this reason, after all edge operators are converted to Neutron IR (and before any IR optimizations), the operators of the graph need to be topologically sorted.

This PR fixes the bug with the intermediate output, and it introduces the topological operator sorting.

Test plan

Unit test provided.

cc @robert-kalmar @JakeStevens @digantdesai @rascani

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20423

• 📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• [ROCm] MI350 CI jobs will have longer queue times due to CI migration

❌ 1 New Failure, 1 Unrelated Failure

As of commit 0f02959 with merge base 1a78804 ():

NEW FAILURE - The following job has failed:

• pull / test-qnn-models-linux (mv2) / linux-job (gh)
  RuntimeError: Command docker exec -t f2ac853cb9a6c17a3e051238cec0b491f7dbd7aa4203c51b67bf35fd67752479 /exec failed with exit code 92

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

• pull / test-llama-runner-qnn-linux (fp32, qnn_16a16w, qnn) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)

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

[linux-foundation-easycla]

• ❌ - login: @MartinPavella / name: Martin Pavella. The commit (0f02959) is not authorized under a signed CLA. Please click here to be authorized. For further assistance with EasyCLA, please visit our EasyCLA portal and chat with our support bot.

[MartinPavella]

@novak-vaclav @irtrukhina please feel free to have a look.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.