NXP backend: Update aot_neutron_compile.py example pipeline #14142

StrycekSimon · 2025-09-10T08:58:13Z

Summary

Updates example in aot_neutron_compile.py to reflect current state of NXP backend.

Test plan

AOT examples should run automatically in CI. You can manually test it using backends/nxp/run_aot_example.sh

cc @digantdesai @JakeStevens @robert-kalmar @MartinPavella @roman-janik-nxp

pytorch-bot · 2025-09-10T08:58:16Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14142

📄 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

As of commit 815a3f7 with merge base fd9f176 ():

NEW FAILURES - The following jobs have failed:

pull / test-models-linux-basic (mv3, xnnpack-quantization-delegation, cmake, linux.arm64.2xlarge, execut... / linux-job (gh)
An error occurred trying to start process '/usr/bin/bash' with working directory '/home/ec2-user/actions-runner/_work/executorch/executorch/pytorch/executorch'. No such file or directory
pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t 549af6653269c562ae6c3241e8d0bd0a47b5f5169ee9a8e3038de3141d718e2e /exec failed with exit code 1

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

StrycekSimon · 2025-09-10T09:11:26Z

@pytorchbot label "release notes: nxp"

roman-janik-nxp · 2025-09-11T15:02:05Z

examples/nxp/aot_neutron_compile.py

+
+    edge_program_manager = NeutronEdgePassManager()(edge_program_manager)
+
+    if args.remove_quant_io_ops:


This step is done before delegation, but in executorch_pipeline.py it is done after delegation. In our local repo it is placed before delegation. Why are there differences?

I think it should be placed before the delegation. @MartinPavella, can you confirm it since you are the author of the updated executorch_pipeline.py?

In executorch_pipeline, the removal of I/O quantization after partitioning was up-streamed by you, with @skywall as the original author (cf2f170).
I did not add the IO removal, so I'm not sure which order is preferred. Perhaps Lukas can comment. (But I am leaning towards keeping it consistent with main-nxp.)

I agree, I will change it.

Edit: Done ✅

roman-janik-nxp · 2025-09-11T15:08:45Z

examples/nxp/aot_neutron_compile.py

+    if args.delegate:
+        logging.info("Executing Neutron Partitioner and Delegate")
+        edge_program_manager = edge_program_manager.to_backend(
            NeutronPartitioner(


Please unify edge_program_manager.to_backend(), compile_spec and partitioner creation with version in executorch_pipeline.py`.

digantdesai · 2025-09-15T16:55:21Z

backends/nxp/tests/executorch_pipeline.py


    if remove_quant_io_ops:
        edge_program_manager = edge_program_manager.transform(
            [RemoveIOQuantOpsPass(edge_program_manager=edge_program_manager)]


unrelated to this PR but why is this not part of the NeutronEdgePassManager?

Good point, I probably did put the RemoveIOQuantOpsPass call on a wrong place. I will double check it.

The RemoveIOQuantOps is intentionally not part of the NeutronPassManager default passes, as this changes the model API:
Normally when the quantization is performed and consequently the delegation to Neutron backend the model input and output type is untouched - e.g. FP32. There is a Quantize node doing the input quantization for the rest of the (optionally) Neutron compute.
This pass removes the Quantize and Dequantize nodes from the I/O tensors, and changes the datatype to int8 .

It can be applied by the NeutronPassManager, but with user choice option. Not by default.

As I said, we can add an optional arg to the PassManager for this pass. It feels like someone has to go out of their way to do this.

I moved the application to the PassManager itself. Now it is an opt-in boolean.

digantdesai · 2025-09-15T16:56:33Z

examples/nxp/aot_neutron_compile.py

    if args.remove_quant_io_ops:
-        edge_program = edge_program.transform(
-            [RemoveIOQuantOpsPass(edge_program_manager=edge_program)]
+        edge_program_manager = edge_program_manager.transform(


as I said before can we move the pass inside the passmanager and add a remove_quant_io as an arg to it?

StrycekSimon · 2025-09-22T15:12:30Z

I refactored the aot_neutron_compile.py file to use to_edge_transform_and_lower for simpicity of the example. The executorch_pipeline.py still requires converting to edge and than to backend separately for testing purposes.

examples/nxp/aot_neutron_compile.py

…pipeline.py

StrycekSimon · 2025-09-23T17:16:27Z

Sorry everyone for the tag, I somehow managed to mess up a rebase.

…EdgePassManager

…_lower

StrycekSimon · 2025-09-24T07:11:46Z

The fail with Samsung's tests is unrelated and appears in other PRs. The second red signal is some problem in the runner setup, and I don't see any relation to changes made in this PR.

robert-kalmar · 2025-09-25T07:59:26Z

@roman-janik-nxp please re-review if you agree with the implementation of the requested change.

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 10, 2025

pytorch-bot bot added the release notes: nxp Changes to the NXP Neutron backend delegate label Sep 10, 2025

roman-janik-nxp reviewed Sep 11, 2025

View reviewed changes

robert-kalmar added the module: nxp Issues related to NXP Neutron NPU delegation and code under backends/nxp/ label Sep 12, 2025

roman-janik-nxp requested changes Sep 15, 2025

View reviewed changes

StrycekSimon force-pushed the upstream/main-nxp/EIEX-533-upstream-addition-of-to_edge_with_preserved_ops-nxpedgepassmanage-to-aot_neutron_compile.py branch from 6d033df to ddf400c Compare September 15, 2025 09:20

roman-janik-nxp approved these changes Sep 15, 2025

View reviewed changes

digantdesai reviewed Sep 15, 2025

View reviewed changes

digantdesai approved these changes Sep 15, 2025

View reviewed changes

StrycekSimon force-pushed the upstream/main-nxp/EIEX-533-upstream-addition-of-to_edge_with_preserved_ops-nxpedgepassmanage-to-aot_neutron_compile.py branch 2 times, most recently from a038916 to 90c851c Compare September 22, 2025 15:08

roman-janik-nxp requested changes Sep 23, 2025

View reviewed changes

examples/nxp/aot_neutron_compile.py Show resolved Hide resolved

StrycekSimon force-pushed the upstream/main-nxp/EIEX-533-upstream-addition-of-to_edge_with_preserved_ops-nxpedgepassmanage-to-aot_neutron_compile.py branch from c4f2d68 to e3cdc16 Compare September 23, 2025 17:08

StrycekSimon requested review from larryliu0820, kirklandsign, lucylq, GregoryComer, swolchok, JacobSzwejbka, manuelcandales, shoumikhin, jackzhxng, mergennachin, kimishpatel, Gasoonjia and cccclai as code owners September 23, 2025 17:08

StrycekSimon requested a review from SS-JIA as a code owner September 23, 2025 17:08

roman-janik-nxp and others added 2 commits September 23, 2025 19:10

NXP backend: Update aot_neutron_compile.py example pipeline

06a9834

NXP backend: Unify pipeines in aot_neutron_compile.py and executorch_…

2fbd13a

…pipeline.py

StrycekSimon force-pushed the upstream/main-nxp/EIEX-533-upstream-addition-of-to_edge_with_preserved_ops-nxpedgepassmanage-to-aot_neutron_compile.py branch from e3cdc16 to bd17839 Compare September 23, 2025 17:14

StrycekSimon removed request for swolchok, shoumikhin, mergennachin, SS-JIA, larryliu0820, GregoryComer, cccclai, kimishpatel, lucylq, JacobSzwejbka, Gasoonjia, jackzhxng, manuelcandales and kirklandsign September 23, 2025 17:15

StrycekSimon force-pushed the upstream/main-nxp/EIEX-533-upstream-addition-of-to_edge_with_preserved_ops-nxpedgepassmanage-to-aot_neutron_compile.py branch from bd17839 to bdf54da Compare September 23, 2025 17:59

StrycekSimon added 2 commits September 23, 2025 20:24

NXP backend: Automatic application of RemoveIOQuantOpsPass in Neutron…

992f1b6

…EdgePassManager

NXP backend: Replace use of export_to_edge with to_edge_transform_and…

815a3f7

…_lower

StrycekSimon force-pushed the upstream/main-nxp/EIEX-533-upstream-addition-of-to_edge_with_preserved_ops-nxpedgepassmanage-to-aot_neutron_compile.py branch from bdf54da to 815a3f7 Compare September 23, 2025 18:31

StrycekSimon requested a review from roman-janik-nxp September 24, 2025 07:12

robert-kalmar approved these changes Sep 25, 2025

View reviewed changes

roman-janik-nxp approved these changes Sep 25, 2025

View reviewed changes

roman-janik-nxp merged commit 25b3d63 into pytorch:main Sep 25, 2025
128 of 130 checks passed

roman-janik-nxp deleted the upstream/main-nxp/EIEX-533-upstream-addition-of-to_edge_with_preserved_ops-nxpedgepassmanage-to-aot_neutron_compile.py branch September 25, 2025 09:18


		edge_program_manager = NeutronEdgePassManager()(edge_program_manager)

		if args.remove_quant_io_ops:

NXP backend: Update aot_neutron_compile.py example pipeline #14142

NXP backend: Update aot_neutron_compile.py example pipeline #14142

Uh oh!

Conversation

StrycekSimon commented Sep 10, 2025 • edited by pytorch-bot bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

pytorch-bot bot commented Sep 10, 2025 • 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/14142

❌ 2 New Failures

Uh oh!

StrycekSimon commented Sep 10, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

StrycekSimon Sep 15, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

StrycekSimon commented Sep 22, 2025

Uh oh!

Uh oh!

StrycekSimon commented Sep 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

StrycekSimon commented Sep 24, 2025

Uh oh!

robert-kalmar commented Sep 25, 2025

Uh oh!

Uh oh!

Uh oh!

StrycekSimon commented Sep 10, 2025 •

edited by pytorch-bot bot

Loading

pytorch-bot bot commented Sep 10, 2025 •

edited

Loading

StrycekSimon Sep 15, 2025 •

edited

Loading

StrycekSimon commented Sep 23, 2025 •

edited

Loading