[RFC] More robust revert functions for convergence tests

[Tcc0403]

🐛 Describe the bug

Many discussions show that the current revert functions have several limitations, including:

• incomplete revert: Fix: pailgemma type error #627 (comment) revert_liger_kernel_to_xxx can't revert LigerCrossEntropyLoss for transformers>=4.46.1 #542
• not automatically updating old reference: Fix flce not being patched after reverting in convergence test #385
• comparing wrong references: Fix: pailgemma type error #627 (comment)

Simply reloading the modules would generate new class/function objects different from the original ones, i.e. revert(patch(object)) != object, which can easily lead to bugs when doing comparisons (isinstance()), repatching, reusing configs and so on.

One way to address the issue is having a patch manager for testing to keep track of replaced objects, so we can easily revert back to the original objects after monkey patch. Or is there a way to achieve it with pytest?

cc @tyler-romero @austin362667

Reproduce

No response

Versions

none