[5763448][ONNX][Autocast] Fix Resize input type mismatch error

[gcunhase]

What does this PR do?

Type of change: Bug fix

Overview: This PR fixes an input type mismatch in Resize layers when being converted to FP16.

Usage

$ python -m modelopt.onnx.autocast --onnx_path=$MODEL_NAME.onnx

Testing

Added unittest.

Before your PR is "Ready for review"

• Make sure you read and follow Contributor guidelines and your commits are signed.
• Is this change backward compatible?: Yes
• Did you write any new necessary tests?: Yes
• Did you add or update any necessary documentation?: No
• Did you update Changelog?: No

Additional Information

This issue is also fixed by using the standalone type inference logic from #719.

Summary by CodeRabbit

Release Notes

• Improvements

  • Enhanced the graph sanitization process to automatically duplicate shared constants during optimization, ensuring improved model handling and consistency.

• Tests

  • Added test coverage for mixed precision conversion of Conv-Resize model architectures.

_{✏️ Tip: You can customize this high-level summary in your review settings.}

[codecov]

Codecov Report

❌ Patch coverage is 85.71429% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 74.23%. Comparing base (18d9b1e) to head (7e9c6f4).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
modelopt/onnx/quantization/fp8.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #757   +/-   ##
=======================================
  Coverage   74.22%   74.23%           
=======================================
  Files         192      192           
  Lines       19027    19033    +6     
=======================================
+ Hits        14123    14129    +6     
  Misses       4904     4904           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[galagam]

LGTM

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

📝 Walkthrough

Walkthrough

The changes introduce a new constant duplication feature to GraphSanitizer that is invoked during graph sanitization, along with test coverage for Conv-Resize model handling in mixed-precision conversion workflows.

Changes

Cohort / File(s)	Summary
GraphSanitizer Enhancement modelopt/onnx/autocast/graphsanitizer.py	Added duplicate_shared_constants() method that duplicates shared constants via utility function and logs the operation. Method is invoked early in sanitize() right after graph name initialization.
Test Model Builder tests/_test_utils/onnx/lib_test_models.py	Added build_conv_resize_model() function that constructs ONNX graph with Conv followed by Resize operations, including initializers for weights and resize parameters. Note: Function appears defined twice with identical implementation.
Test Case Addition tests/unit/onnx/autocast/test_autocast.py	Added test_conv_resize_conversion() test that validates mixed-precision conversion of Conv-Resize models, following pattern of existing conv_isinf test. Import extended to include new test model builder.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 55.56% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main bug fix: addressing a Resize input type mismatch error during model autocast conversion to mixed precision (FP16).

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 0

🧹 Nitpick comments (2)

tests/_test_utils/onnx/lib_test_models.py (1)

958-972: Consider adding a brief comment explaining the shared constant pattern.

The intentional reuse of resize_roi_scales for both roi and scales inputs is the core of this test. A brief comment would help future maintainers understand this is deliberate, not accidental.

📝 Suggested clarifying comment

         helper.make_node(
             op_type="Resize",
+            # Note: resize_roi_scales is intentionally used for both roi and scales inputs
+            # to test the shared constant duplication fix (PR #757)
             inputs=[
                 "conv1_conv/Conv2D:0",
                 "resize_roi_scales",
                 "resize_roi_scales",
                 "resize_sizes",
             ],
             outputs=["output_0"],
             name="resize1_resize/Resize",
             coordinate_transformation_mode="asymmetric",
             cubic_coeff_a=-0.75,
             mode="nearest",
             nearest_mode="floor",
         ),

tests/unit/onnx/autocast/test_autocast.py (1)

193-210: Test validates the fix but could be more explicit about what it's testing.

The test successfully verifies that a Conv-Resize model can be converted to mixed precision without errors. However:

1. The comment on line 205 mentions "QDQ node placements" but the test doesn't check QDQ nodes—consider updating the comment to match the actual assertion.

2. Since this PR specifically fixes the Resize input type mismatch, consider adding an assertion that verifies the Resize node's data input is FP16 while its roi, scales, and sizes inputs remain their original types (the core bug was that shared constants caused type conflicts).

💡 Suggested improvements

     # Output model should be produced in the same tmp_path
     output_onnx_path = onnx_path.replace(".onnx", ".fp16.onnx")
     onnx.save(converted_model, output_onnx_path)

-    # Load the output model and check QDQ node placements
+    # Load the output model and verify conversion
     graph = gs.import_onnx(converted_model)

     # Check that Conv is converted
     conv_nodes = [n for n in graph.nodes if "Conv" in n.op]
     assert assert_input_precision(conv_nodes)

+    # Check that Resize node has correct input types:
+    # - Data input (index 0) should be FP16
+    # - roi/scales/sizes inputs should remain their original types
+    resize_nodes = [n for n in graph.nodes if n.op == "Resize"]
+    assert len(resize_nodes) == 1
+    resize_node = resize_nodes[0]
+    assert resize_node.inputs[0].dtype == np.float16, "Resize data input should be FP16"

📜 Review details

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 6ae96b5 and e8bc39e.

📒 Files selected for processing (3)

• modelopt/onnx/autocast/graphsanitizer.py
• tests/_test_utils/onnx/lib_test_models.py
• tests/unit/onnx/autocast/test_autocast.py

🧰 Additional context used

🧬 Code graph analysis (2)

tests/unit/onnx/autocast/test_autocast.py (2)

tests/_test_utils/onnx/lib_test_models.py (1)

• build_conv_resize_model (929-1009)

modelopt/onnx/autocast/convert.py (1)

• convert_to_mixed_precision (46-189)

tests/_test_utils/onnx/lib_test_models.py (1)

modelopt/onnx/utils.py (2)

• infer_shapes (731-744)
• check_model (555-567)

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (5)

• GitHub Check: linux
• GitHub Check: wait-checks / wait
• GitHub Check: wait-checks / wait
• GitHub Check: code-quality
• GitHub Check: build-docs

🔇 Additional comments (4)

modelopt/onnx/autocast/graphsanitizer.py (2)

70-70: Placement of duplicate_shared_constants() looks appropriate.

Calling this early in sanitize(), after ensuring the graph name but before naming nodes, is a sensible ordering. This ensures shared constants are duplicated before subsequent transformations that might depend on unique tensor names.

---

258-262: API contract verified — no issues found.

The utility function onnx_utils.duplicate_shared_constants exists in modelopt/onnx/utils.py at line 524 with the expected signature: def duplicate_shared_constants(onnx_model: onnx.ModelProto) -> tuple[onnx.ModelProto, bool]:. The import statement is correct, and the tuple unpacking in the method correctly handles the (model, bool) return type. The code is implemented correctly.

tests/_test_utils/onnx/lib_test_models.py (1)

929-1009: Test model correctly reproduces the shared constant scenario.

The build_conv_resize_model() function intentionally uses resize_roi_scales for both the roi and scales inputs of the Resize node (inputs at indices 1 and 2). This shared constant pattern is exactly what triggers the type mismatch error during FP16 conversion, making this an effective regression test for the bug fix.

The model structure and validation follow the established patterns in this file.

tests/unit/onnx/autocast/test_autocast.py (1)

23-23: LGTM!

Import correctly extended to include build_conv_resize_model.