fix(profiling): patch sub-component methods at class level to prevent instance-isolation bug

[Anai-Guo]

Fixes #13462

Root cause

annotate_pipeline called setattr(component, method_name, annotate(method, label)) where method is a bound method (retrieved via getattr(component, method_name)). The resulting wrapper is a plain function that closes over the original bound method — and thus retains a hard reference to the original component instance.

When a pipeline internally copies a sub-component (e.g. audio_scheduler = copy.copy(self.scheduler) in pipeline_ltx2.py), the copy's instance __dict__ inherits the wrapper. Calling audio_scheduler.step(...) then silently dispatches to self.scheduler.step(...), sharing internal state (_step_index) between the two schedulers and producing index-out-of-bounds errors.

Fix

Detect bound methods with inspect.ismethod and patch at the class level instead:

if inspect.ismethod(method):
    setattr(type(component), method_name, annotate(method.__func__, label))
else:
    setattr(component, method_name, annotate(method, label))

Because the wrapper is now set as a class attribute (an unbound function), Python's descriptor protocol re-binds it to whichever instance accesses it — including copies. Both the original scheduler and audio_scheduler will have properly isolated step calls.

Impact

• examples/profiling/profiling_utils.py only (no production pipeline code changed)
• Purely additive: adds import inspect and a 5-line if/else guard
• Pipelines without internal component copies are unaffected

[sayakpaul]

Class-level patches should be avoided as they're not transient.

[Anai-Guo]

Thanks for the feedback @sayakpaul! Updated the PR to make the patches explicitly transient:

1. annotate_pipeline() now returns a restore() callable that undoes all class-level (and instance-level) patches by restoring saved original attributes.

2. PipelineProfiler stores the restore callable as self._restore_annotations and calls it at the end of run() — before del pipe — so the class is restored to its original state after each profiling run.

Class-level patches are still needed to fix the instance-isolation bug (the copy issue where audio_scheduler = copy.copy(self.scheduler) would otherwise dispatch to the original scheduler's state), but they're now explicitly cleaned up rather than being permanent.