test: direct unit suite for DistillationModel (37 tests)#1923
Conversation
CPU-only tests with tiny hand-weighted Linear stacks so every activation and loss is exact in fp32 - the existing distill tests build on timm models and assert scalar-ness, not values. Covers convert-path construction, per-layer-pair capture hooks (values, grad_fn asymmetry, clearing), compute_kd_loss with exact two-pair balanced totals and skip_balancer contents, teacher eval re-forcing, fwd-only context managers, state-dict key equality and value restoration, export hook cleanup, and error paths. Adversarially reviewed: 4/4 seeded mutations killed (incl. a pair-ordering probe). Also surfaces that export() leaves stale _intermediate_output attributes pinning the last captured activations - reported for follow-up. Part of the coverage initiative in NVIDIA#1902. Signed-off-by: arham766 <arhamislam766@yahoo.com>
|
Warning Review limit reached
Next review available in: 10 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Enterprise Run ID: 📒 Files selected for processing (1)
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
1 participant
What does this PR do?
Type of change: new tests
Part of the unit-coverage initiative in #1902. Direct unit suite for modelopt/torch/distill/distillation_model.py (+registry.py) using tiny hand-weighted Linear stacks so every activation and loss is exact in fp32 — the existing distill tests (timm-based, currently uncollectable without timm) assert scalar-ness and key presence, not values. Covers: convert-path construction (teacher instance vs ModelLike class), per-layer-pair capture hooks (exact tensor values, student-grad/teacher-no-grad asymmetry, overwrite semantics, eval-mode warning asymmetry), compute_kd_loss (exact two-pair balanced total 0.25·1+0.25·4+0.5·2=2.25; skip_balancer dict contents; captures cleared), teacher eval re-forcing and frozen params, only_teacher/only_student contexts incl. return values, state-dict key equality with a plain student + value restoration, export() class restore and hook cleanup, and error paths. Adversarial review: 14 derivations verified, 4/4 mutations killed including a pair-ordering probe. Follow-up finding: export() removes hooks but leaves stale _intermediate_output attributes pinning the last captured activations (incl. an autograd graph) on both student and teacher layers.
Usage
N/A — tests only.
Testing
Hermetic, CPU-only, deterministic, <2s. Combined run with the sibling calib/distill suites green; full tests/unit/torch/quantization dir unaffected. Adversarially reviewed with mutation testing as described above.
Before your PR is "Ready for review"
Additional Information
Issue: #1902