Qualcomm AI Engine Direct - accept an ExportedProgram in to_edge_tran… by psiddh · Pull Request #23025 · pytorch/executorch

psiddh · 2026-09-22T21:25:21Z

…sform_and_lower_to_qnn

to_edge_transform_and_lower_to_qnn only took a module and always re-exported it internally, so a caller that has already captured a program - with its own dynamic shapes, fx fixes or graph surgery - had no way to hand that program in and had to reimplement the transform/partition/lower sequence against backends.qualcomm._passes directly.

Accept ExportedProgram (and Dict[str, ExportedProgram]) and skip the internal re-export for those entries. Module inputs are unchanged.

cc @cbilgin

pytorch-bot · 2026-09-22T21:25:25Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23025

📄 Preview Python docs built from this PR

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

✅ You can merge normally! (1 Unrelated Failure)

As of commit 57f1267 with merge base 273cb33 ():

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

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

pull / test-arm-backend-no-driver (test_pytest_models_tosa) / linux-job (gh) (trunk failure)
RuntimeError: Command bash /__w/_temp/exec_script failed with exit code 1

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

linux-foundation-easycla · 2026-09-22T21:25:28Z

The committers listed above are authorized under a signed CLA.

✅ login: psiddh / name: Siddartha Pothapragada (cae5d49)

github-actions · 2026-09-22T21:26:03Z

This PR needs a `release notes:` label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

Copilot

Copilot review overview

🟡 Changes recommended

Update the inputs type annotation and add coverage for dictionary ExportedProgram inputs.

Review effort: Lite
Findings: 1

Open (1)

Annotate inputs as Optional to match documented None support · New

What changed in this PR

Adds support for passing caller-supplied ExportedProgram instances to Qualcomm QNN lowering without re-exporting.

Changes:

Supports single and dictionary ExportedProgram inputs.
Preserves caller-defined graph transformations and dynamic shapes.
Adds regression coverage for supplied programs.

File	Summary	Review findings
`backends/qualcomm/utils/utils.py`	Extends the lowering API for `ExportedProgram` inputs.	Moderate: annotate `inputs=None` correctly (4 votes). Nit: add dictionary-input regression coverage (1 vote).
`backends/qualcomm/tests/test_passes.py`	Tests lowering a supplied `ExportedProgram`.	Coverage does not yet exercise the dictionary path.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

mergennachin

Review assisted by Codex.

I also agree with the existing inputs annotation comment: the public signature should allow the documented inputs=None usage, through Optional or an overload.

I found no runtime correctness issue in the dispatch change. QNN CI checks passed; full QNN lowering was not run locally on macOS.

shewu-quic

Thank you for your contribution. Could you help to add the test in our new test framework?

diff --git a/backends/qualcomm/tests/rework/src/utils.py b/backends/qualcomm/tests/rework/src/utils.py
index 9202007b18..95004f6e69 100644
--- a/backends/qualcomm/tests/rework/src/utils.py
+++ b/backends/qualcomm/tests/rework/src/utils.py
@@ -633,3 +633,20 @@ class QAT:
             assert len(weight_fqs) > 0, "no int8-range weight FQ found for fp16a8w"
             # should complete without error
             __class__._get_converted_module(module, prepared, inputs)
+
+class LoweringWithExportedProgram:
+    @staticmethod
+    def test(compile_spec):
+        from unittest import mock
+        module = _UtilsModel()
+        inputs = (torch.randn(1, 4, 8, 8), torch.randn(1, 4, 8, 8))
+        exported = torch.export.export(module, inputs, strict=True)
+        with mock.patch(
+            "torch.export.export",
+            wraps=torch.export.export,
+        ) as export_spy:
+            _ = to_edge_transform_and_lower_to_qnn(
+                exported, inputs, compile_spec
+            )
+            export_spy.assert_not_called()
+    
diff --git a/backends/qualcomm/tests/rework/utils/test.py b/backends/qualcomm/tests/rework/utils/test.py
index 8330f79ab8..f412714702 100644
--- a/backends/qualcomm/tests/rework/utils/test.py
+++ b/backends/qualcomm/tests/rework/utils/test.py
@@ -39,3 +39,6 @@ def test_skip_node_quantizer(subtests, quantizer, compile_spec):
 
 def test_qat(subtests):
     QAT.test(subtests)  # noqa: F405
+    
+def test_lowering_with_exported_program(compile_spec):
+    LoweringWithExportedProgram.test(compile_spec)  # noqa: F405

psiddh · 2026-09-29T21:12:20Z

@mergennachin , you were right that the shape assertion couldn't detect a re-export. The test now spies on torch.export.export and asserts it is never called during lowering. I checked it against your counterexample (accept the ExportedProgram, then re-export m.module()): it fails with Expected 'export' to not have been called. Called 1 times.

@shewu-quic added LoweringWithExportedProgram to the rework framework using the same spy, plus the validation you asked for: supplying inputs or dynamic_shapes alongside an ExportedProgram now warns naming which arguments were ignored, with a test for it.

…sform_and_lower_to_qnn to_edge_transform_and_lower_to_qnn only took a module and always re-exported it internally, so a caller that has already captured a program - with its own dynamic shapes, fx fixes or graph surgery - had no way to hand that program in and had to reimplement the transform/partition/lower sequence against backends.qualcomm._passes directly. Accept ExportedProgram (and Dict[str, ExportedProgram]) and skip the internal re-export for those entries. Module inputs are unchanged. inputs is now Optional, matching the documented inputs=None usage, and supplying inputs or dynamic_shapes alongside an already-captured program warns rather than silently dropping them. A shape assertion alone cannot detect a re-export, since re-exporting exported.module() reproduces the original graph. The tests therefore spy on torch.export.export and require that lowering never calls it; both the unittest and the rework-framework test fail against a variant that accepts the program and re-exports it.

Copilot

Copilot review overview

🟡 Changes recommended

SDK setup may discard supplied programs through re-exec, and dictionary-input regression coverage is missing.

Review effort: Lite
Findings: 1

Open (1)

Avoid SDK re-exec after supplied ExportedProgram capture · New

Resolved since last review (1)

Annotate inputs as Optional to match documented None support

Copilot

Copilot review overview

🔵 Needs a closer look

SDK setup may discard supplied programs via re-exec, and dictionary-form inputs lack dedicated regression coverage.

Review effort: Lite
Findings: 1

Open (1)

Avoid SDK re-exec after supplied ExportedProgram capture

shewu-quic

LGTM. Thanks!

Copilot

Copilot review overview

🟢 Approval recommended

No unresolved issues were identified, and all readiness assessments support approval.

Review effort: Lite
Findings: None

Resolved since last review (1)

Avoid SDK re-exec after supplied ExportedProgram capture

digantdesai

stamping to unblock

psiddh · 2026-09-30T05:11:50Z

Thank you for your contribution. Could you help to add the test in our new test framework?

diff --git a/backends/qualcomm/tests/rework/src/utils.py b/backends/qualcomm/tests/rework/src/utils.py
index 9202007b18..95004f6e69 100644
--- a/backends/qualcomm/tests/rework/src/utils.py
+++ b/backends/qualcomm/tests/rework/src/utils.py
@@ -633,3 +633,20 @@ class QAT:
             assert len(weight_fqs) > 0, "no int8-range weight FQ found for fp16a8w"
             # should complete without error
             __class__._get_converted_module(module, prepared, inputs)
+
+class LoweringWithExportedProgram:
+    @staticmethod
+    def test(compile_spec):
+        from unittest import mock
+        module = _UtilsModel()
+        inputs = (torch.randn(1, 4, 8, 8), torch.randn(1, 4, 8, 8))
+        exported = torch.export.export(module, inputs, strict=True)
+        with mock.patch(
+            "torch.export.export",
+            wraps=torch.export.export,
+        ) as export_spy:
+            _ = to_edge_transform_and_lower_to_qnn(
+                exported, inputs, compile_spec
+            )
+            export_spy.assert_not_called()
+    
diff --git a/backends/qualcomm/tests/rework/utils/test.py b/backends/qualcomm/tests/rework/utils/test.py
index 8330f79ab8..f412714702 100644
--- a/backends/qualcomm/tests/rework/utils/test.py
+++ b/backends/qualcomm/tests/rework/utils/test.py
@@ -39,3 +39,6 @@ def test_skip_node_quantizer(subtests, quantizer, compile_spec):
 
 def test_qat(subtests):
     QAT.test(subtests)  # noqa: F405
+    
+def test_lowering_with_exported_program(compile_spec):
+    LoweringWithExportedProgram.test(compile_spec)  # noqa: F405

Done

psiddh · 2026-09-30T05:20:44Z

Addressed all the feedback

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 22, 2026

psiddh marked this pull request as ready for review September 28, 2026 06:18

psiddh requested a review from abhinaykukkadapu as a code owner September 28, 2026 06:18

Copilot AI lite review requested due to automatic review settings September 28, 2026 06:18

psiddh added the module: qnn Issues related to Qualcomm's QNN delegate and code under backends/qualcomm/ label Sep 28, 2026

Copilot started reviewing on behalf of psiddh September 28, 2026 06:19 View session

Copilot AI reviewed Sep 28, 2026

View reviewed changes

Comment thread backends/qualcomm/utils/utils.py Outdated

mergennachin reviewed Sep 28, 2026

View reviewed changes

Comment thread backends/qualcomm/tests/test_passes.py Outdated

shewu-quic reviewed Sep 29, 2026

View reviewed changes

Comment thread backends/qualcomm/utils/utils.py

Copilot AI lite review requested due to automatic review settings September 29, 2026 21:11

psiddh force-pushed the qnn-lower-accepts-exported-program branch from 8e58297 to cae5d49 Compare September 29, 2026 21:11

Copilot started reviewing on behalf of psiddh September 29, 2026 21:11 View session

psiddh force-pushed the qnn-lower-accepts-exported-program branch from cae5d49 to 2e2e12c Compare September 29, 2026 21:14

Copilot AI reviewed Sep 29, 2026

View reviewed changes

Comment thread backends/qualcomm/utils/utils.py

Copilot AI lite review requested due to automatic review settings September 29, 2026 21:14

Copilot started reviewing on behalf of psiddh September 29, 2026 21:14 View session

Copilot AI reviewed Sep 29, 2026

View reviewed changes

shewu-quic approved these changes Sep 29, 2026

View reviewed changes

Merge branch 'main' into qnn-lower-accepts-exported-program

57f1267

Copilot AI lite review requested due to automatic review settings September 30, 2026 00:34

Copilot started reviewing on behalf of psiddh September 30, 2026 00:35 View session

Copilot AI reviewed Sep 30, 2026

View reviewed changes

digantdesai approved these changes Sep 30, 2026

View reviewed changes

psiddh merged commit fb60708 into pytorch:main Sep 30, 2026
208 of 209 checks passed

Conversation

psiddh commented Sep 22, 2026 • edited by pytorch-bot Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot Bot commented Sep 22, 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/23025

✅ You can merge normally! (1 Unrelated Failure)

Uh oh!

linux-foundation-easycla Bot commented Sep 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Sep 22, 2026

This PR needs a release notes: label

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Uh oh!

mergennachin left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

shewu-quic left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

psiddh commented Sep 29, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🔵 Needs a closer look

Uh oh!

shewu-quic left a comment

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟢 Approval recommended

Uh oh!

digantdesai left a comment

Choose a reason for hiding this comment

Uh oh!

psiddh commented Sep 30, 2026

Uh oh!

psiddh commented Sep 30, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

psiddh commented Sep 22, 2026 •

edited by pytorch-bot Bot

Loading

pytorch-bot Bot commented Sep 22, 2026 •

edited

Loading

linux-foundation-easycla Bot commented Sep 22, 2026 •

edited

Loading

This PR needs a `release notes:` label