Add regression test for ITensor with [BackwardDerivative]

[jhelferty-nv]

Add regression test for slangpy#639. Tests forward and backward passes
for a function using ITensor interface parameters with
[BackwardDerivative], the pattern from the original issue report. The
underlying compiler crash was fixed by the auto-diff refactor in
shader-slang/slang#9808.

Fixes #639

Made with Cursor

Summary by CodeRabbit

• Tests
  • Added regression test coverage for custom backward derivative support in differentiable tensor operations, validating forward computation correctness and backward gradient computation execution.

[coderabbitai]

📝 Walkthrough

Walkthrough

A regression test module is added to validate issue #639, introducing parameterized tests for forward and backward passes of a Slang shader function with ITensor arguments and a custom backward derivative implementation.

Changes

Cohort / File(s)	Summary
Regression Test for ITensor Custom Backward slangpy/tests/slangpy_tests/test_itensor_custom_backward.py	Introduces two parameterized pytest tests: one verifying forward computation correctness against NumPy output, and another confirming backward pass execution without crashes. Tests target custom [BackwardDerivative] handling with ITensor<float, 1> parameters.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

Possibly related PRs

• PR #882 — Modifies autograd/backward gradient-handling internals that these new tests exercise in the forward and backward passes.

Suggested reviewers

• bmillsNV
• tdavidovicNV

Poem

🐰 A shader that tensors once feared,
Now tests make the path crystal-cleared,
With backwards so true, and ITensors too,
The segfault's defeated—regression adieu! 🎉

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title 'Add regression test for ITensor with [BackwardDerivative]' directly and accurately reflects the main change: a new regression test for ITensor with custom backwards derivatives targeting issue #639.
Linked Issues check	✅ Passed	The PR adds a regression test that exercises both forward and backward passes for ITensor with [BackwardDerivative], addressing the core requirements from issue #639: verifying the forward pass executes without segfault and the backward pass runs successfully.
Out of Scope Changes check	✅ Passed	All changes are scoped to adding regression tests for issue #639; the new test module targets ITensor behavior with [BackwardDerivative] and contains no unrelated modifications.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

Warning

Review ran into problems

🔥 Problems

Timed out fetching pipeline failures after 30000ms

---

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]

🧹 Nitpick comments (1)

slangpy/tests/slangpy_tests/test_itensor_custom_backward.py (1)

59-63: Use .grad_in for semantic clarity in backward pass setup.

The pattern in test_difftensor_load_variants.py explicitly uses result.grad_in.storage.copy_from_numpy() rather than result.grad.storage, and also asserts result.grad_in is not None. While .grad is a valid convenience property that resolves to .grad_in when .grad_out is unavailable, using .grad_in directly is more semantically explicit for seeding the input gradient to the backward pass. For consistency with the codebase pattern and clarity, use:

result.grad_in.storage.copy_from_numpy(np.ones(N, dtype=np.float32))

The regression test purpose is already documented in the test docstring.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@slangpy/tests/slangpy_tests/test_itensor_custom_backward.py` around lines 59
- 63, Replace the backward-pass seeding to use the explicit input-gradient
field: ensure the test sets the seeded gradient into result.grad_in rather than
the convenience alias result.grad; call
result.grad_in.storage.copy_from_numpy(np.ones(N, dtype=np.float32)) and assert
result.grad_in is not None before copying so the backward invocation
func.bwds(coord=spy.grid((N,)), x=xt, _result=result) receives the semantically
correct input gradient.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Nitpick comments:
In `@slangpy/tests/slangpy_tests/test_itensor_custom_backward.py`:
- Around line 59-63: Replace the backward-pass seeding to use the explicit
input-gradient field: ensure the test sets the seeded gradient into
result.grad_in rather than the convenience alias result.grad; call
result.grad_in.storage.copy_from_numpy(np.ones(N, dtype=np.float32)) and assert
result.grad_in is not None before copying so the backward invocation
func.bwds(coord=spy.grid((N,)), x=xt, _result=result) receives the semantically
correct input gradient.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 1d523617-177d-45f5-8fad-e501b2098117

📥 Commits

Reviewing files that changed from the base of the PR and between bcc393c and 1773ad5.

📒 Files selected for processing (1)

• slangpy/tests/slangpy_tests/test_itensor_custom_backward.py

[ccummingsNV]

Looks good to me - I'm guessing it has to wait for fixed slang

[jhelferty-nv]

Correct, we don't have a public Slang release with shader-slang/slang#9808 in it yet.

[jhelferty-nv]

Moving this back to draft because it's taking a while to get a new slang update into slangpy with shader-slang/slang#9808.