Refactor module access to use PyTorch get/set_submodule API

[scopophobic]

Summary

issue : #1362
This PR refactors module access and replacement to rely on PyTorch’s native
get_submodule / set_submodule APIs instead of maintaining custom traversal logic.

What changed

• Removed legacy get_attr / set_attr helpers
• Unified module access through get_module / set_module
  (thin wrappers over get_submodule / set_submodule)
• Verified all call sites only operate on modules (no attribute access)

Why

PyTorch has provided a robust module replace API since 1.x.
Using the native API improves correctness, readability, and long-term maintainability,
and avoids subtle issues with manual attribute traversal.

Behavior

No functional changes — this is a pure refactor.

[xin3he]

LGTM

[yiliu30]

LGTM, thanks!

[scopophobic]

Thank you for the review @xin3he @yiliu30 . why these two test-cases failing ?

[yiliu30]

Hi @scopophobic you can check the log here. https://dev.azure.com/lpot-inc/neural-compressor/_build/results?buildId=51719&view=logs&j=44c25250-aab3-5e31-d6d7-8ba2147b1266&t=262f41be-8379-5409-f492-e6c716395db9
The TestAutoRoundAct.test_act_config_MXFP4_saving failed.

[scopophobic]

I'm now seeing test failures that appear to be related to this change:

Root Cause

The new fail-fast behavior is breaking code in test/helpers.py (lines 98-100):

for key in model._tied_weights_keys:
    weight = get_attr(model, key)
    set_attr(model, key, copy.deepcopy(weight))

The issue: model._tied_weights_keys may contain keys that don't exist in all model configurations. The old
get_attr silently returned None, but the new version raises AttributeError.

Question: Which approach should I take?

Option 1: Keep silent None behavior

• Revert get_attr to return None for missing attributes
• Maintains backward compatibility but Loses the fail-fast benefit.

Option 2: Improve the test case

Update test/helpers.py to handle exceptions properly:

for key in model._tied_weights_keys:
    try:
        weight = get_attr(model, key)
        set_attr(model, key, copy.deepcopy(weight))
    except AttributeError:
        # Skip tied weights that don't exist in this model configuration
        pass

• Keeps the fail-fast behavior (better for catching bugs)
• Makes the test code more explicit about handling missing keys

My recommendation: Option 2 (improve the test case), since fail-fast with clear errors is better for debugging and prevents silent bugs from propagating.

What do you think?