Adds QAT ConvBN fuse pass to utils by JakeStevens · Pull Request #17599 · pytorch/executorch

JakeStevens · 2026-02-20T20:50:45Z

Summary:
Earlier PR adds support for a pass that quantizes the bias resulting from QAT ConvBN fusion without an initial bias.

This PR adds it to the NXP calibrate_and_quantize method.

Differential Revision: D93904683

cc @robert-kalmar @digantdesai

pytorch-bot · 2026-02-20T20:50:49Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17599

📄 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, 2 Unrelated Failures

As of commit 05b4b89 with merge base 19e8b68 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t f7a20546618fd530c116ff948fa4fa13eff2acf268fa5ba1c5b735cd3247d985 /exec failed with exit code 1
pull / test-samsung-quantmodels-linux / linux-job (gh)
RuntimeError: Command docker exec -t 6418c274d428f4ff1ac01c0db08c721c3a0aac8f8e29a6b89852a46ca5e73c3a /exec failed with exit code 1

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

pull / unittest-editable / macos / macos-job (gh) (similar failure)
AttributeError: '_OpNamespace' 'mkldnn' object has no attribute '_is_mkldnn_acl_supported'

BROKEN TRUNK - The following job failed but was present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

pull / unittest / macos / macos-job (gh) (trunk failure)
AttributeError: '_OpNamespace' 'mkldnn' object has no attribute '_is_mkldnn_acl_supported'

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

meta-codesync · 2026-02-20T20:50:52Z

@JakeStevens has exported this pull request. If you are a Meta employee, you can view the originating Diff in D93904683.

Summary: Earlier PR adds support for a pass that quantizes the bias resulting from QAT ConvBN fusion without an initial bias. This PR adds it to the NXP calibrate_and_quantize method. Differential Revision: D93904683

larryliu0820

Review automatically exported from Phabricator review in Meta.

Summary: Earlier PR adds support for a pass that quantizes the bias resulting from QAT ConvBN fusion without an initial bias. This PR adds it to the NXP calibrate_and_quantize method. Reviewed By: larryliu0820 Differential Revision: D93904683

Summary: Pull Request resolved: pytorch#17599 Earlier PR adds support for a pass that quantizes the bias resulting from QAT ConvBN fusion without an initial bias. This PR adds it to the NXP calibrate_and_quantize method. Reviewed By: larryliu0820 Differential Revision: D93904683

Summary: Earlier PR adds support for a pass that quantizes the bias resulting from QAT ConvBN fusion without an initial bias. This PR adds it to the NXP calibrate_and_quantize method. Reviewed By: larryliu0820 Differential Revision: D93904683

Summary: Pull Request resolved: pytorch#17599 Earlier PR adds support for a pass that quantizes the bias resulting from QAT ConvBN fusion without an initial bias. This PR adds it to the NXP calibrate_and_quantize method. Reviewed By: larryliu0820 Differential Revision: D93904683

Summary: Earlier PR adds support for a pass that quantizes the bias resulting from QAT ConvBN fusion without an initial bias. This PR adds it to the NXP calibrate_and_quantize method. Reviewed By: larryliu0820 Differential Revision: D93904683

roman-janik-nxp · 2026-03-02T15:26:18Z

@StrycekSimon please review.

robert-kalmar · 2026-03-03T16:32:25Z

Sorry for the wait time, I am currently investigating a failed test case in our internal tests related to this PR. Will notify when I have more info.

@StrycekSimon , the internal CI passed on this PR. What failure you refer to?

roman-janik-nxp · 2026-03-03T13:17:54Z

+        model, input_shape, use_qat=True, use_neutron_for_format_conversion=False
+    ).exported_program()
+
+    assert any("lowered_module" in node.name for node in edge_program.graph.nodes)


Please change to checking targets. We check for delegate calls and have a util for it.

Suggested change

assert any("lowered_module" in node.name for node in edge_program.graph.nodes)

assert graph_contains_any_of_ops(edge_program.graph, [torch.ops.higher_order.executorch_call_delegate])

JakeStevens · 2026-03-04T14:25:53Z

In the test test_biasless_convbn_fusion_qat, the added QuantizeFusedConvBnBiasAtenPass correctly quantize conv bias. This looks like fixing an error in the graph, shouldn't be the bias quantized when added to the graph? Do I understand it right @JakeStevens ?

Yes this is the intention, exactly. We need to "manually" quantize the bias after we fuse conv bn in QAT

Summary: Earlier PR adds support for a pass that quantizes the bias resulting from QAT ConvBN fusion without an initial bias. This PR adds it to the NXP calibrate_and_quantize method. Reviewed By: larryliu0820 Differential Revision: D93904683

robert-kalmar · 2026-03-04T16:06:31Z

@@ -23,6 +23,8 @@
    to_quantized_edge_program,
 )
 from executorch.backends.nxp.tests.executors import OverrideTargetSupportCheck


Suggested change

from executorch.backends.nxp.tests.executors import OverrideTargetSupportCheck

from executorch.backends.nxp.tests.executors import (

graph_contains_any_of_ops,

OverrideTargetSupportCheck,

)

.. to fix the nxp-unittest and linting error