fix: Hessian auto-quant crash and qconfig_dict mutation bugs

[musicalplatypus]

Summary

Two bugs in Hessian-based auto-quantization (auto_quantization.py/qconfig_types.py), both real, reproducible on the code as it stands.

Bare TypeError instead of graceful degradation

run_auto_quantization()'s optimal_bitwidth stays None whenever no calibration_dataloader is supplied, or compute_hessian_sensitivity() can't compute sensitivities (it returns ({}, {}) when sample inputs/targets/criterion aren't available) — a path callers like audio_classification/train.py already anticipate and try to degrade gracefully from, logging "Could not obtain sample data... Proceeding without it". That graceful-degradation intent was defeated a few lines later:

total_bit_budget = optimal_bitwidth * total_params

TypeError: unsupported operand type(s) for *: 'NoneType' and 'int'.

Fixed by returning the caller's already-built qconfig_mapping unchanged when optimal_bitwidth is None — this is the mapping the caller constructed from the manually-specified uniform bitwidth before auto-quantization was attempted, making it the semantically correct fallback (not a forced all-float32 fallback, and not a crash).

apply_mixed_precision() mutates the caller's live config dict

qconfig_dict['weight']['bitwidth'] = bit_width
qconfig_dict['activation']['bitwidth'] = bit_width

Two of its four call sites — qconfig_types.py's manual mixed-precision path, and auto_quantization.py's default Hessian auto-quantization path (i.e. essentially every successful quantized run) — pass the original, un-copied qconfig_dict, which is frequently the same dict object retained as self.qconfig_type on the live model wrapper (quant_base.py). Any downstream logging, checkpoint metadata, or reuse of that config silently observed whichever bitwidth tier this function processed last, not the model's actual/intended default. Two other call sites already build a throwaway copy before calling in — confirming the mutation was an oversight, not intended behavior.

Fixed in the function itself, not just the two unsafe call sites: builds a local copy per bit-width tier instead of mutating the caller's dict, so every caller is protected regardless of whether it already defended itself.

Verification

3 of 4 added tests fail on the unmodified code — both crash tests reproduce the exact TypeError, and the mutation test observes the caller's dict actually change from 8 to 4 — and pass after the fix. The 4th test confirms the mutation fix doesn't break the actual mixed-precision assignment. Also verified this merges cleanly alongside pr/dataloader-pinning-fix (#20), which touches a sibling file (quant_base.py) in the same directory.

🤖 Generated with Claude Code