[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

未关闭
  • 哪里不对？
#1,902 5 条评论 0 个 reaction 已指派 1 人 在 GitHub 查看

维护者通常 2 天内回复

@kevalmorabia97 已经在做这个了。

开始于 2026年7月21日。

评估

这个 Issue 还没有评估数据。

描述

Summary

While auditing test coverage, we found that several load-bearing modules have no direct unit-test file — they are exercised only incidentally through higher-level flows. We wrote hermetic, CPU-only unit suites (per CONTRIBUTING's tests/unit rules: fast, no network, no GPU) for five of them, adversarially reviewed and mutation-checked each suite, and will attach one PR per module so each is reviewable in isolation:

Module Suite Tests
torch/quantization/conversion.py test_conversion.py 63
torch/quantization/utils/core_utils.py test_core_utils.py 128
torch/quantization/nn/modules/tensor_quantizer.py test_tensor_quantizer.py 90 (incl. hand-computed INT8/FP8/block fake-quant values)
torch/quantization/model_calib.py test_model_calib.py 42
torch/export/postprocess.py test_postprocess.py 74 (TP/PP split layouts, padding math)

397 tests, ~4.2k lines, total runtime ~5s. Each suite was verified to kill seeded mutations of its target (wrong scale direction, flipped split axes, broken running-max, etc.), so this is regression-detection capacity, not line-count.

Bugs found while writing these (documented in-test with NOTE: documents current behavior comments; happy to file/fix separately as preferred)

  1. _normalize_fused_experts_quantizer_name collides with SequentialQuantizer child names (introduced in #1340: the singular weight_quantizer.<N> pattern matches sequential children). Two consequences, both reproduced: (a) applying the same list-of-configs twice via set_quantizer_attributes_full nests SequentialQuantizers inside sub-slots (non-idempotent, corrupt state); (b) set_quantizer_attributes_partial with a list and a *weight_quantizer wildcard now raises ValueError mid-iteration for targets that are already sequential — a regression against its own docstring. Suggested fix: skip normalization when the matched quantizer's parent is a SequentialQuantizer.
  2. awq() silently no-ops on an unknown algorithm string (public __all__ API; only the mtq.quantize config layer validates the literal). One-line raise ValueError in an else would match the module's fail-fast style.
  3. TensorQuantizer.extra_repr dead code: the disabled branch builds a detailed string, then returns the literal "disabled", discarding it.
  4. update_lm_head_quantization warns "Enable lm_head quantization" even when quantizers are already disabled, and its sole call site (model_config_export.py:323) appears to pass inference_pipeline_parallel into the inference_tensor_parallel parameter.
  5. Minor exception-safety inconsistencies: replace_function/calibrate_with_adapters/enable_fake_quant yield without try/finally (unlike export_torch_mode et al.), so an exception mid-context leaks the patched state.

The behavior-documenting tests are written to fail-and-force-update when these are fixed.

PRs

Will be linked below, one per module, each self-contained.

主要语言
Python
星标
5.5k
派生
785
平均合并
3 天 22 小时
30 天内合并 PR
277

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

NVIDIA/Model-Optimizer 的其他 Issue

查看 NVIDIA/Model-Optimizer 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。