test: direct unit suite for the QuantModule machinery (35 tests)#1925
test: direct unit suite for the QuantModule machinery (35 tests)#1925arham766 wants to merge 1 commit into
Conversation
Targets the module-conversion plumbing itself, complementing the large
indirect surface: registry class-swap with dynamic-class caching,
_setup quantizer wiring and default descriptors, forward wrapping with
exact-value equality vs manual quantizer application, the external
monkey-patched-forward branch (quantized input + export restore -
covered nowhere else), quantize_weight/fold_weight semantics,
modelopt_post_restore device handling, ParallelState, and export class
restoration. Adversarially reviewed: 3/3 seeded mutations killed.
Documents a genuine crash: fold_weight derives the weight name with
name[:-10], stripping '_quantizer' instead of '_weight_quantizer' as
its comment claims - custom per-weight quantizers
(quantizer_attr_names('gate_up_proj')) trip the hasattr assertion, and
neither _QuantLlama4TextExperts nor _QuantGptOssExperts overrides
fold_weight, so public mtq.fold_weight() on a quantized Llama-4 or
GPT-OSS model crashes (reproduced; any fix must also respect the
transposed quantization of those weights).
Part of the coverage initiative in NVIDIA#1902.
Signed-off-by: arham766 <arhamislam766@yahoo.com>
|
Warning Review limit reached
Next review available in: 2 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/quantization/nn/modules/quant_module.py — the module-conversion machinery exercised indirectly everywhere but never directly (non-overlap map vs test_forward_patching / test_module_registry / test_quantize_replace / test_quant_linear verified by review). Covers registry class-swap + dynamic-class caching, _setup quantizer wiring and default descriptors, forward wrapping with exact-value equality, the external monkey-patched-forward branch (receives quantized input; export restores the patch — covered nowhere else), quantize_weight/fold_weight semantics incl. exception-safety, modelopt_post_restore device handling, ParallelState, export class restoration. Adversarial review: 3/3 mutations killed. Genuine crash documented and independently reproduced: fold_weight derives the weight attr as name[:-10] (strips '_quantizer', not the '_weight_quantizer' its comment claims), so custom per-weight quantizers from quantizer_attr_names() trip the hasattr assertion — and neither _QuantLlama4TextExperts nor _QuantGptOssExperts overrides fold_weight, so public mtq.fold_weight() on a quantized Llama-4/GPT-OSS model crashes. Any fix must also respect those weights' transposed quantization; happy to send it as a follow-up.
Usage
N/A — tests only.
Testing
Hermetic, CPU-only, deterministic, <1s. Full tests/unit/torch/quantization dir green alongside (789 passed, pre-existing skips only). Adversarially reviewed with independent re-derivation and mutation testing as described.
Before your PR is "Ready for review"
Additional Information
Issue: #1902