test: Implement Unit Tests for `StateDictAdapter` Class

[pankaj-bind]

Bug description

The StateDictAdapter class in torchtitan/protocols/state_dict_adapter.py, which is responsible for converting model state dictionaries between TorchTitan and Hugging Face formats, currently lacks unit tests. This absence of tests poses a risk of incorrect model weight conversions, which could impact model accuracy and make future code changes difficult to verify.

To reproduce:

1. Examine the tests/unit_tests/ directory.
2. Note the absence of a test file for torchtitan/protocols/state_dict_adapter.py.
3. The lack of tests means the conversion logic is unverified.

If allowed, I can discuss more on it, also I can write a solution for it.

[tianyu-l]

We have some high level guidance in https://github.com/pytorch/torchtitan/tree/main/scripts/checkpoint_conversion, but I felt it's not easy to do unit testing especially for large models. Happy to hear what your test strategy would be.

[fegin]

If we mock out the torch operations, we should be able to test the naming conversion and also check the generated shapes are expected.

[pankaj-bind]

@tianyu-l , My test strategy is to:

Use mocks for ModelArgs and placeholder tensors instead of loading full models. This keeps the tests fast and lightweight, as suggested by @fegin.
Focus on the conversion logic, verifying that the key names and tensor shapes are correctly mapped between TorchTitan and Hugging Face formats.
Implement round-trip tests(native -> HF -> native) to ensure consistency, and also test edge cases like empty dictionaries.

[tianyu-l]

I'm OK with such tests, although it sounds there's no hope to capture all numerical errors

which could impact model accuracy

E.g. due to #335

[fegin]

I'm OK with such tests, although it sounds there's no hope to capture all numerical errors
You couldn't expect a unittest to capture all numerical errors. The unittest can only detect certain logic errors and avoid some errors introduced by refactoring the existing code.