test: direct unit suite for TensorQuantizer with exact-value fake-quant checks (90 tests)

[arham766]

What does this PR do?

Type of change: new tests

Part of the unit-coverage initiative in #1902 (one PR per module; this one is self-contained). Covers: the TensorQuantizer class contract directly: 10 hand-computed INT8 exact-value forwards (narrow/unsigned/per-channel/int4/zero-amax), 4 exact FP8 E4M3 cases incl. RNE ties, block quant, fp16/bf16 round-trips, amax/pre_quant_scale get-set semantics, calibration choreography, modelopt_state+state_dict round-trip, STE backward variants, backend registry error paths. All rounding operands verified tie-safe for platform determinism.

Usage

N/A — tests only.

Testing

Suite is hermetic (CPU-only, no network, deterministic, <3s) and passes locally; the full collectable tests/unit tree passes with it (1999 passed). The suite was adversarially reviewed and verified to kill seeded mutations of the target module (details in #1902).

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 (external contributor)

Additional Information

Issue: #1902

Summary by CodeRabbit

• Tests
  • Added extensive CPU-focused unit coverage for tensor quantization, including default configuration, enable/disable behavior, scaling interactions, and robust handling of empty tensors.
  • Verified amax and pre-quant scale semantics (including state persistence and shape-change safeguards).
  • Added exact correctness checks for INT8/FP8 fake quantization and fixed-shape block quantization.
  • Covered dtype handling, calibration contract, state serialization/restoration, backward/STE behavior, backend registry validation, sequential quantizer delegation, and warning hygiene.

[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.

[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: 0b2f784f-eb8f-44ec-8a5b-a94c928547b2

📥 Commits

Reviewing files that changed from the base of the PR and between ad0ffd5 and 3703629.

📒 Files selected for processing (1)

• tests/unit/torch/quantization/test_tensor_quantizer.py

🚧 Files skipped from review as they are similar to previous changes (1)

• tests/unit/torch/quantization/test_tensor_quantizer.py

---

📝 Walkthrough

Walkthrough

This PR adds a new CPU-only unit test module for TensorQuantizer, covering construction, enable/disable behavior, amax and pre_quant_scale semantics, exact fake-quant math, block quantization, dtype handling, calibration, state serialization, repr/validation, backward gradients, hard-disabled behavior, backend registry behavior, SequentialQuantizer contracts, constant amax, and warning hygiene.

Changes

TensorQuantizer test suite

Layer / File(s)	Summary
Module setup and construction/enable defaults tests/unit/torch/quantization/test_tensor_quantizer.py	Adds docstring, imports, _tq helper, and tests for default construction, config propagation, and enable/disable forward behavior.
Amax and pre_quant_scale contracts tests/unit/torch/quantization/test_tensor_quantizer.py	Tests amax get/set, buffer persistence, reset, step-size behavior, and pre_quant_scale setter/context-manager behavior.
Exact fake-quant math and block quantization tests/unit/torch/quantization/test_tensor_quantizer.py	Hand-computed INT8/FP8 fake-quant forward tests, static block quantization, and dtype preservation tests.
Calibration and state serialization tests/unit/torch/quantization/test_tensor_quantizer.py	Tests calibration collection/load behavior and modelopt state serialization/restoration.
Repr, validation, and backward gradient tests tests/unit/torch/quantization/test_tensor_quantizer.py	Tests extra_repr stability, validate_attr/export_amax behavior, and STE gradient behavior.
Hard-disabled quantizer, backend registry, sequential and constant amax tests/unit/torch/quantization/test_tensor_quantizer.py	Tests HardDisabledTensorQuantizer errors, backend registry behavior, SequentialQuantizer delegation, constant amax override, and warning hygiene.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Possibly related issues

• [Coverage] Direct unit-test suites for five untested core modules (conversion, core_utils, tensor_quantizer, model_calib, export/postprocess) — plus bugs found along the way #1902: Adds direct unit tests for tensor_quantizer.py, matching the added coverage in this module.

🚥 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 accurately summarizes the new TensorQuantizer test suite and its exact-value fake-quant coverage.
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	Only a test file changed, and it contains none of the banned load/eval/nosec patterns or new deps.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[coderabbitai]

🧹 Nitpick comments (1)

tests/unit/torch/quantization/test_tensor_quantizer.py (1)

766-810: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Consider a fixture for backend registry register/unregister cleanup.

test_register_forward_unregister and test_reregister_warns each hand-roll try/finally registration cleanup for register_quant_backend/unregister_quant_backend. A small pytest fixture yielding a registered-name helper (with teardown always calling unregister_quant_backend) would remove the duplicated try/finally boilerplate across these tests.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/unit/torch/quantization/test_tensor_quantizer.py` around lines 766 -
810, The backend registry tests duplicate register/unregister cleanup logic in
test_register_forward_unregister and test_reregister_warns. Introduce a small
pytest fixture in TestQuantBackendRegistry that registers a backend name for
each test and always unregisters it in teardown, then update those tests to use
the fixture instead of hand-written try/finally blocks around
register_quant_backend and unregister_quant_backend.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@tests/unit/torch/quantization/test_tensor_quantizer.py`:
- Around line 766-810: The backend registry tests duplicate register/unregister
cleanup logic in test_register_forward_unregister and test_reregister_warns.
Introduce a small pytest fixture in TestQuantBackendRegistry that registers a
backend name for each test and always unregisters it in teardown, then update
those tests to use the fixture instead of hand-written try/finally blocks around
register_quant_backend and unregister_quant_backend.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 9bc62e74-ebe0-4ae7-a491-b8bfc7bf77f8

📥 Commits

Reviewing files that changed from the base of the PR and between b96a785 and ad0ffd5.

📒 Files selected for processing (1)

• tests/unit/torch/quantization/test_tensor_quantizer.py

[arham766]

Consolidated into #1927 per maintainer feedback in #1902 — the suite was trimmed to only the lines codecov reports uncovered, with parametrization clusters deduplicated. Closing in favor of that PR.