Fix clip backward with broadcasted constant bounds

[Robertboy18]

What does this PR do?

Type of change: Bug fix

Hey! I was experimenting with formalizing the clipping backward rule in Lean using TorchLean, and noticed a small discrepancy here.

ClipFunction.forward supports broadcastable tensor bounds, but backward raises the scalar-only error even when those bounds are constants and do not need gradients.

This keeps the scalar restriction for learnable bounds, while allowing constant tensor bounds to broadcast normally. I also added a CPU regression test that compares both the output and input gradient against the equivalent PyTorch min/max operations.

I modeled the coordinatewise broadcasted input VJP in Lean as well and kernel-checked the result. The formalization is not included in this PR, but I'm happy to share it if useful :)

I used a bit of AI assistance during the investigation and debugging, then manually checked the code, tests, and Lean statement/results.

Usage

inputs = torch.randn(2, 3, requires_grad=True)
clip_min = torch.tensor([-1.0, 0.0, 1.0])
clip_max = torch.tensor([0.0, 1.0, 3.0])

outputs = clip(inputs, clip_min, clip_max)
outputs.sum().backward()

Testing

• Added a CPU regression test for vector bounds broadcast over a matrix.
• Compared forward output and input gradients with PyTorch.
• Ran focused pytest and pre-commit checks.

Before your PR is "Ready for review"

• Is this change backward compatible?: ✅
• If you copied code from any other sources or added a new PIP dependency, did you follow guidance in CONTRIBUTING.md: N/A
• Did you write any new necessary tests?: ✅
• Did you update Changelog?: N/A
• Did you get Claude approval on this PR?: N/A

Summary by CodeRabbit

• Bug Fixes

  • Fixed clipping operations with broadcastable, non-learnable minimum and maximum bounds.
  • Backward propagation now works correctly for tensor bounds without requiring scalar shapes.

• Tests

  • Added coverage verifying clipping outputs and input gradients against equivalent native PyTorch operations.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: d6d51867-01d3-42f3-b060-6703ac9b002c

📥 Commits

Reviewing files that changed from the base of the PR and between 089109d and d8fb74f.

📒 Files selected for processing (2)

• modelopt/torch/quantization/nn/functional.py
• tests/unit/torch/quantization/test_functional.py

---

📝 Walkthrough

Walkthrough

The clip backward validation now permits broadcastable non-learnable bounds. A CPU test verifies forward outputs and input gradients against native PyTorch operations.

Changes

Broadcastable Clip Bounds

Layer / File(s)	Summary
Clip validation and broadcast coverage modelopt/torch/quantization/nn/functional.py, tests/unit/torch/quantization/test_functional.py	Validation rejects non-scalar bounds only when they require gradients. The CPU test covers per-column broadcast bounds and compares outputs and input gradients with torch.minimum and torch.maximum.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested reviewers: sugunav14

🚥 Pre-merge checks | ✅ 6

✅ Passed checks (6 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the fix for clip backward propagation with broadcasted constant bounds.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Security Anti-Patterns	✅ Passed	The production change only adds requires_grad shape checks; no unsafe loads, remote-code flags, eval/exec, nosec comments, or dependency additions appear in the PR diff.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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