NXP backend: Add recipes for Neutron backend lowering. by MartinPavella · Pull Request #21516 · pytorch/executorch

MartinPavella · 2026-07-31T13:39:37Z

Summary

This PR introduces a declarative recipe-based lowering for the NXP Neutron backend. The previous solution was implemented as executorch_pipeline.py:to_quantized_executorch_program(). The new solution provides the same functionailty and produces the same results. The benefit of the recipe-based approach is compatibility with other backends (recipe fusing) and adhering to ExecuTorch standards.

Once this is merged, the old Neutron lowering pipeline can be removed completely.

Test plan

pytest backends/nxp/tests/generic_tests/test_recipe_export.py

cc @robert-kalmar @JakeStevens @digantdesai @rascani

pytorch-bot · 2026-07-31T13:39:42Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21516

📄 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 358c0d4 with merge base adaaa8b ():
💚 Looks good so far! There are no failures yet. 💚

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

linux-foundation-easycla · 2026-07-31T13:39:44Z

❌ - login: @MartinPavella / name: Martin Pavella. The commit (084c302, 358c0d4) 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.

The original implementation used an optional callback which if provided would override the logic in the QuantizeStage. This was done for simplicity. However, that approach made recipe fusing impossible and sort of went against the idea of recipes. Therefore, this commit introduces full QAT support into the QuantizeStage.

Copilot

Pull request overview

This PR adds a recipe-driven export/lowering path for the NXP Neutron backend by extending the generic export pipeline to support additional lowering hooks (pre-partition callback, post-partition transforms) and a richer quantization recipe flow (PTQ/QAT phases + pass hooks). It also introduces NXP-specific recipe types/provider implementations and a new test suite to validate recipe behavior and combination.

Changes:

Extend EdgeTransformAndLowerStage to run an optional pre_partitioning_callback and apply post_partitioning_transforms from the lowering recipe.
Extend QuantizeStage / QuantizationRecipe to support QAT (prepare_qat_pt2e), optional training, custom calibration inputs, and post-phase GraphModule pass hooks.
Add NXP Neutron recipe types/provider and a new test suite validating PTQ/QAT/no-delegate flows and recipe-combination behavior.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
export/stages.py	Adds lowering pre/post partition hooks and expands quantization flow to support QAT/training/calibration hooks.
export/recipe.py	Extends recipe dataclasses (quantization + lowering) and updates recipe-combination logic for new hooks.
backends/nxp/tests/generic_tests/test_recipe_export.py	New tests covering Neutron recipe export behavior, config flags, and recipe combination.
backends/nxp/tests/executorch_pipeline.py	Deprecates imperative lowering helpers in favor of recipes.
backends/nxp/recipes/nxp_recipe_types.py	Introduces NXP-specific `RecipeType` enum for Neutron exports.
backends/nxp/recipes/nxp_recipe_provider.py	Implements the NXP recipe provider, recipe config, and lowering/quantization recipe builders.
backends/nxp/edge_passes/neutron_edge_pass_manager.py	Switches PassManager import to ExecuTorch’s pass manager implementation.

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

novak-vaclav · 2026-08-04T15:40:39Z

+        if (
+            hasattr(er := artifact.context["export_recipe"], "lowering_recipe")
+            and er.lowering_recipe is not None
+            and (callback := er.lowering_recipe.pre_partitioning_callback) is not None
+        ):
+            callback(self._partitioners, artifact.data)


Applicable, please use get method, similarily to line 272, or a shared local variable.

novak-vaclav · 2026-08-04T15:57:53Z

+    def __call__(self, s: str, epm: EdgeProgramManager) -> NeutronEdgePassManager:
+        return self.neutron_edge_pass_manager


I think this is applicable (the Copilot's comment)

novak-vaclav · 2026-08-04T15:59:10Z

+    input_spec: Iterable[ModelInputSpec] | tuple[int, ...] | list[tuple[int, ...]]
+    target: str = "imxrt700"
+    operators_not_to_delegate: list[str] = None
+    intermediates_dir: str | None = None


Applicable (the Copilot's comment).

robert-kalmar · 2026-08-04T19:52:47Z

+    "NXP backend: The imperative lowering functions will be removed in the future. "
+    "Please use the recipe-based approach instead."
+)
 def to_quantized_edge_program(


The recipes are still experimental:
export/recipe.py

@experimental( "This API and all of its related functionality such as ExportSession and ExportRecipe are experimental." ) @dataclass class ExportRecipe:

I don't think we should deprecate our functions until recipes are production.

novak-vaclav

Below I have a few questions, comments and suggestions, otherwise very good job!!
I really like the idea of recipes, I believe it will make everything much more understandable for our users 👍👍👍😄

btw. I'm attaching an extended test suite my AI agents generated, feel free to run it to check the correctness after making modifications to the code or to draw inspiration from 😊

test_recipe_export_extended.py

novak-vaclav · 2026-08-03T13:15:52Z

+        input_spec: Model input description. Accepts a single shape tuple, a list of
+                    shape tuples (one per input), or a list of ModelInputSpec objects.
+        target: Neutron hardware target string. Default: "imxrt700".
+        operators_not_to_delegate: Optional list of op names excluded from NPU delegation.


nit: I would add an example of entry in the operators_not_to_delegate list, like "aten::add".

novak-vaclav · 2026-08-03T13:40:12Z

+)
+from executorch.backends.nxp.recipes.nxp_recipe_types import NXP_BACKEND, NXPRecipeType
+from executorch.backends.nxp.tests.executorch_pipeline import (
+    _get_default_quantizer,


nit: From my coding experience in Python, methods with _ prefix are meant to be "private" and used only in the class/module they are defined or implemented in. Thus I think it would be better to rename the method in executorch_pipeline to get_default_quantizer without the prefix.

novak-vaclav · 2026-08-03T14:15:30Z

+            compile_spec,
+            neutron_target_spec,
+            rc.custom_delegation_options,
+            preserve_ops=[torch.ops.aten.prelu.default],


This comment might not be directly related to this PR, but I think it's time to refactor the preserve_ops list we now use in multiple different calls (afaik it's here and in executorch_pipeline).
I'm saying this because the logic of preserving ops became more complicated and conditional in the aten.pad PR, and also Roman added aten.hardswish in his PR.
I suggest modifying NeutronPartitioner to set preserve_ops to [prelu, hardswish, pad] as default or extracting [prelu, hardswish, pad] to some global variable.
There is also a core_aten_ops_exception_list variable in our backend, which seems to do something similar to preserve_ops.
If you don't want to solve it in this PR, just add aten.pad and aten.harswish here and we will tackle the refactoring in another issue.

It should be unified as I commented too. We have it also in aot_neutron_example.py.

novak-vaclav · 2026-08-04T13:36:46Z

+        nodes = list(graph.nodes)
+
+        real_nodes = [n for n in nodes if n.op not in ("placeholder", "output")]
+        assert real_nodes[0].target != quantized_decomposed.quantize_per_tensor.out


nit: I think it would be more robust to check that the model input's child is the quant node and the model output's predecessor is dequant node.
There might be an edge case where there could be a model like this: x -> add(x, 2) -> y and the list of nodes could be ordered like this [x, get_attr(2), quant, add, dequant, y], since get_attr(2) and quant are on the same "level" in the model's graph.
Same goes for checking the dtype.

novak-vaclav · 2026-08-04T13:39:42Z

+        assert placeholder_nodes[0].meta["val"].dtype == torch.int8
+
+    def test_use_quant_state_dict_false(self):
+        """use_quant_state_dict=False: NeutronPartitioner gets None state_dict, no crash."""


I'd check that post-quantization parameter values were not passed to the partitioner.

novak-vaclav · 2026-08-04T15:59:10Z

+    input_spec: Iterable[ModelInputSpec] | tuple[int, ...] | list[tuple[int, ...]]
+    target: str = "imxrt700"
+    operators_not_to_delegate: list[str] = None
+    intermediates_dir: str | None = None


Applicable (the Copilot's comment).

novak-vaclav · 2026-08-04T16:06:02Z

+        post_prepare_passes=[_post_prepare],
+        # Applied after training (or after prepare when no train_fn).
+        post_train_passes=[_remove_simulated_bn_and_fuse],
+        # Applied after calibration (only reached when train_fn is None).


Are you sure post_calibration_passes are applied only when train_fn is None?
in stages.py on line 512, I see post_calibration_passes are always run.

novak-vaclav · 2026-08-05T08:27:40Z

+                m = move_exported_model_to_train(m)
+                qr.train_fn(m)
+
+            if qr.is_qat:


I would restructure these conditions containing qr.is_qat and qr.train_fn, since train_fn is not None implies qr.is_qat (or vice versa). Because of move_exported_model_to_train being idempotent, I would simply do:

if qr.train_fn is not None: m = move_exported_model_to_train(m) qr.train_fn(m) m = move_exported_model_to_eval(m)

novak-vaclav · 2026-08-05T09:17:55Z

+
+    input_spec: Iterable[ModelInputSpec] | tuple[int, ...] | list[tuple[int, ...]]
+    target: str = "imxrt700"
+    operators_not_to_delegate: list[str] = None


nit: the type of operators_not_to_delegate should be list[str] | None

novak-vaclav · 2026-08-05T09:25:14Z

+                f"NXP backend: create_recipe() requires `{NEUTRON_RECIPE_CONFIG_KEY}=<NeutronRecipeConfig>`."
+            )
+
+        if rc.custom_delegation_options is None:


I'm not a big fan of modifying the config that is passed to the recipe creator, I view it as an anti-pattern. Additionally, it might have unwanted consequences if someone decides to reuse the config somewhere else. Same goes for line 177, where the issue is even more severe.
If you decide to keep it anyway, please add a doc string to this method with a mention about this behavior.

roman-janik-nxp

Great job Martin 👍 . Implementation is solid in my opinion. I found mostly text errors or bugs in tests. I studied the code and the classes thoroughly to try to understand it.

roman-janik-nxp · 2026-08-04T14:32:35Z

+            case NXPRecipeType.INT8_PTQ_NEUTRON:
+                return self._build_recipe(recipe_type, rc, is_qat=False, delegate=True)
+            case NXPRecipeType.INT8_QAT_NEUTRON:
+                if rc.train_fn is None:


Is the train_fn really needed to be required? In calibrate_and_quantize() we don't require it. In _build_quantization_recipe() you reference it "or after prepare when no train_fn".

roman-janik-nxp · 2026-08-04T14:53:19Z

+            compile_spec,
+            neutron_target_spec,
+            rc.custom_delegation_options,
+            preserve_ops=[torch.ops.aten.prelu.default],


We should place the preserve_ops list in one place, I suggest a global variable in executor_pipeline.py.

roman-janik-nxp · 2026-08-04T15:06:38Z

+    def is_cnn_op(n):
+        return any(op in n.name.lower() for op in ["conv", "relu", "view", "add"])
+
+    assert not graph_contains_any(graph, is_cnn_op)


I wouldn't use ops names instead of targets like we do everywhere else. Applies everywhere.

roman-janik-nxp · 2026-08-04T15:26:22Z

+        model = SimpleCNN()
+        rc = NeutronRecipeConfig(INPUT_SHAPE, get_quantizer_fn=my_quantizer_fn)
+        sess = _run_export(model, rc)
+        assert custom_quantizer_called, "Custom quantizer factory was not called."


Suggested change

assert custom_quantizer_called, "Custom quantizer factory was not called."

assert custom_quantizer_called, "Custom quantizer factory was called."

roman-janik-nxp · 2026-08-04T15:27:41Z

+            INPUT_SHAPE, get_calibration_inputs_fn=my_calibration_fn
+        )
+        sess = _run_export(model, rc)
+        assert calibration_called, "Custom calibration function was not called."


Suggested change

assert calibration_called, "Custom calibration function was not called."

assert calibration_called, "Custom calibration function was called."

roman-janik-nxp · 2026-08-05T12:22:33Z

+        assert sess.get_edge_program_manager() is not None
+
+    def test_custom_delegation_options_explicit(self):
+        """Explicitly provided CustomDelegationOptions is forwarded correctly."""


Suggested change

"""Explicitly provided CustomDelegationOptions is forwarded correctly."""

"""Explicitly provided CustomDelegationOptions are forwarded correctly."""

roman-janik-nxp · 2026-08-05T12:44:51Z

+
+        example_inputs = [(torch.randn(INPUT_SHAPE),)]
+        export(model, example_inputs=example_inputs, export_recipe=recipe)
+        assert transform_called, "post_partitioning_transforms were not executed."


Suggested change

assert transform_called, "post_partitioning_transforms were not executed."

assert transform_called, "post_partitioning_transforms were executed."

roman-janik-nxp · 2026-08-05T12:57:00Z

+    "NXP backend: The imperative lowering functions will be removed in the future. "
+    "Please use the recipe-based approach instead."
+)
 def to_quantized_edge_program(


The recipes are still experimental:
export/recipe.py

@experimental( "This API and all of its related functionality such as ExportSession and ExportRecipe are experimental." ) @dataclass class ExportRecipe:

I don't think we should deprecate our functions until recipes are production.

MartinPavella requested review from irtrukhina, jirioc, novak-vaclav and roman-janik-nxp July 31, 2026 13:39

MartinPavella self-assigned this Jul 31, 2026

MartinPavella requested a review from robert-kalmar as a code owner July 31, 2026 13:39

MartinPavella added module: nxp Issues related to NXP Neutron NPU delegation and code under backends/nxp/ release notes: nxp Changes to the NXP Neutron backend delegate labels Jul 31, 2026

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 Jul 31, 2026

MartinPavella added 2 commits August 3, 2026 12:40

NXB backend: Add recipes for Neutron backend lowering.

084c302

MartinPavella force-pushed the nxg01483/EIEX-999-implement-recipe-for-neutron-backend branch from 4c575c7 to 358c0d4 Compare August 3, 2026 10:40

rascani requested a review from JacobSzwejbka August 3, 2026 12:48

novak-vaclav requested a review from Copilot August 3, 2026 15:38

Copilot started reviewing on behalf of novak-vaclav August 3, 2026 15:38 View session

Copilot AI reviewed Aug 3, 2026

View reviewed changes

robert-kalmar changed the title ~~NXB backend: Add recipes for Neutron backend lowering.~~ NXP backend: Add recipes for Neutron backend lowering. Aug 4, 2026

robert-kalmar reviewed Aug 4, 2026

View reviewed changes

novak-vaclav reviewed Aug 5, 2026

View reviewed changes

roman-janik-nxp reviewed Aug 5, 2026

View reviewed changes

		def __call__(self, s: str, epm: EdgeProgramManager) -> NeutronEdgePassManager:
		return self.neutron_edge_pass_manager

	assert custom_quantizer_called, "Custom quantizer factory was not called."
	assert custom_quantizer_called, "Custom quantizer factory was called."

	assert calibration_called, "Custom calibration function was not called."
	assert calibration_called, "Custom calibration function was called."

	"""Explicitly provided CustomDelegationOptions is forwarded correctly."""
	"""Explicitly provided CustomDelegationOptions are forwarded correctly."""

	assert transform_called, "post_partitioning_transforms were not executed."
	assert transform_called, "post_partitioning_transforms were executed."

Uh oh!

Conversation

MartinPavella commented Jul 31, 2026 • 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 Jul 31, 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/21516

✅ No Failures

Uh oh!

linux-foundation-easycla Bot commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Choose a reason for hiding this comment

Uh oh!

novak-vaclav Aug 4, 2026 • 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!

Uh oh!

novak-vaclav Aug 4, 2026 • 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!

novak-vaclav left a comment

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!

novak-vaclav Aug 4, 2026 • 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!

roman-janik-nxp left a comment

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!

MartinPavella commented Jul 31, 2026 •

edited by pytorch-bot Bot

Loading

pytorch-bot Bot commented Jul 31, 2026 •

edited

Loading

linux-foundation-easycla Bot commented Jul 31, 2026 •

edited

Loading

novak-vaclav Aug 4, 2026 •

edited

Loading

novak-vaclav Aug 4, 2026 •

edited

Loading

novak-vaclav Aug 4, 2026 •

edited

Loading