Fix fuse_pad_into_conv incorrectly fusing when Pad contains negative values

[Copilot]

The fuse_pad_into_conv optimization would fuse a Pad node into Conv even when pads contained negative values (valid for Pad as crop operations, but invalid for Conv's pads attribute), producing an invalid model.

Changes

• _fuse_pad_into_conv.py: Added a guard in _FuseConvPadBase.check() to bail out when any computed pad value is negative, preventing fusion that would produce an invalid Conv node:

  if any(p < 0 for p in self._pads_list):
      return check_result.fail(f"{pads.name} must not contain negative values.")

• _fuse_pad_into_conv_test.py: Added a test case to test_unsupported_fuse_pad_into_conv verifying that fusion is rejected when spatial pads are negative.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>fuse_pad_into_conv optimization incorrectly fuses when 'pads' contain negative values</issue_title>
<issue_description>converting the attached ONNX model with the following script:

from onnxscript import ir
from onnxscript.optimizer import optimize

ir_model = ir.load("input.onnx")
optimized_model = optimize(ir_model)
ir.save(optimized_model, "output.onnx")

results in the following output model:

The Conv layer has negative pad values, which is not allowed per spec. This optimization should check for negative pad values.

input.onnx.zip

versions of relevant packages:

ml_dtypes         0.5.4
onnx              1.20.1
onnx-ir           0.1.16
onnxscript        0.6.2
```</issue_description>

## Comments on the Issue (you are @copilot in this section)

<comments>
</comments>

• Fixes fuse_pad_into_conv optimization incorrectly fuses when 'pads' contain negative values #2838

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[bas-aarts]

What is the reason the PR cannot be merged? It's a rathe simple fix.

[Copilot]

Pull request overview

This PR fixes an invalid optimization in the rewriter where Pad was being fused into Conv even when Pad.pads contained negative values (crop semantics), which would produce an invalid Conv node (negative pads are not allowed for Conv).

Changes:

• Add a guard in _FuseConvPadBase.check() to reject fusion when any computed pad value is negative.
• Add a regression test ensuring fusion is rejected when spatial pad values are negative.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxscript/rewriter/rules/common/_fuse_pad_into_conv.py	Prevents Pad→Conv fusion when computed pads contain negative values, avoiding invalid Conv attributes.
onnxscript/rewriter/rules/common/_fuse_pad_into_conv_test.py	Adds coverage for the negative-padding (crop) case to ensure the fusion is correctly rejected.

[codecov]

❌ 1 Tests Failed:

Tests completed	Failed	Passed	Skipped
11842	1	11841	2255

View the top 1 failed test(s) by shortest run time

onnxscript.rewriter.ort_fusions.fused_matmul_rule_sets_test.TestFusedMatmulRules::test_fused_matmul_div_models_2_matmul_div_div

Stack Traces | 0.024s run time

.nox\test\Lib\site-packages\parameterized\parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
onnxscript\rewriter\ort_fusions\fused_matmul_rule_sets_test.py:323: in test_fused_matmul_div_models
    self._check_model(model_proto, rewritten_model, atol=1e-6)
onnxscript\rewriter\ort_fusions\fused_matmul_rule_sets_test.py:289: in _check_model
    np.testing.assert_allclose(a, b, atol=atol, rtol=rtol)
E   AssertionError: 
E   Not equal to tolerance rtol=1e-07, atol=1e-06
E   
E   Mismatched elements: 1 / 6 (16.7%)
E   Mismatch at index:
E    [0, 1]: -27.57023048400879 (ACTUAL), -27.570234298706055 (DESIRED)
E   Max absolute difference among violations: 3.8146973e-06
E   Max relative difference among violations: 1.3836289e-07
E    ACTUAL: array([[-10.351939, -27.57023 ,  -5.415182],
E          [-20.712341,  -1.450797,   0.456769]], dtype=float32)
E    DESIRED: array([[-10.35194 , -27.570234,  -5.415183],
E          [-20.712343,  -1.450797,   0.45677 ]], dtype=float32)

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}