fix: stop pinning live DataLoaders in quantized model state - #20

Summary

TinyMLQuantFxBaseModule stored the full qconfig_type dict — including the raw calibration/eval DataLoader objects auto-quantization embeds for its one-time bitwidth search — as self.qconfig_type, permanently reachable from the wrapped model.

Root cause

With persistent_workers=True and workers already spun up, copy.deepcopy(model) in export_model() hits a live _MultiProcessingDataLoaderIter and crashes:

NotImplementedError: ('{} cannot be pickled', '_MultiProcessingDataLoaderIter')

A second, related bug found while fixing the first: self.qconfig_type = qconfig_type is a reference assignment, so the wrapper module's dict and the caller's dict were the same object. The subsequent .pop('calibration_dataloader') / .pop('eval_dataloader') mutated whatever dict the caller passed in, not just the wrapper's own copy — a caller that constructs qconfig_type once and reuses or reads it afterward would see it silently emptied by a constructor call it doesn't own.

Fix

• Drop the DataLoader references from self.qconfig_type once the auto-quantization search has consumed them, since nothing reads them afterward.
• Copy qconfig_type before popping from it, so the caller's own dict is never mutated.
• Wire tinyml-modeloptimization/torchmodelopt into CI: the regression test for this was never actually executed — the workflow only triggered on tinyml-modelmaker/** paths and had no step to run torchmodelopt's own tests, so a regression here could merge without a red check.

Verification

Added test_quant_base_dataloader_leak.py, covering both the deepcopy-after-export crash and the caller-dict-mutation case.

Note: the CI workflow change here touches .github/workflows/test-modelmaker.yml, which two of my other open PRs (pr/test-ci-hygiene, pr/mps-eval-fixes) also touch — each adds a path-trigger/step at the same position. If merged after either, GitHub will show a trivial conflict; the correct resolution is keeping all the added lines (a union), not choosing one over the other.

🤖 Generated with Claude Code