[CI] Refactor Bria Transformer Tests

[DN6]

What does this PR do?

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[sayakpaul]

@claude review this PR.

[claude]

Claude finished @sayakpaul's task in 2m 53s —— View job

---

PR Review: [CI] Refactor Bria Transformer Tests

• Gather context and read files
• Read review rules
• Analyze changed files
• Post review feedback

---

Overall this is a clean refactor to the new BaseModelTesterConfig / mixin style, adding meaningful new test coverage (IPAdapterTesterMixin, LoraTesterMixin, LoraHotSwappingForModelTesterMixin, TorchCompileTesterMixin for Fibo). No correctness bugs in the model code itself — these are test-only changes. A few minor issues worth considering:

---

Issues

1. generator property called once per tensor — each gets an independent seed-0 generator

In both BriaTransformerTesterConfig and BriaFiboTransformerTesterConfig, get_dummy_inputs accesses self.generator once per randn_tensor call. Since generator is a @property that constructs a torch.Generator("cpu").manual_seed(0) fresh each access, each tensor independently starts from seed 0. For identically-shaped tensors this produces identical values; for differently-shaped ones the overlap is partial. The effect is subtle but non-conventional and could mask model bugs (e.g., if the model collapses when certain inputs are equal).

The idiomatic approach is to create the generator once per get_dummy_inputs() call:

# tests/models/transformers/test_models_transformer_bria.py, get_dummy_inputs
def get_dummy_inputs(self, batch_size: int = 1) -> dict[str, torch.Tensor]:
    generator = torch.Generator("cpu").manual_seed(0)  # one generator, shared across all calls
    ...
    return {
        "hidden_states": randn_tensor(..., generator=generator, ...),
        "encoder_hidden_states": randn_tensor(..., generator=generator, ...),
        ...
    }

Same applies to BriaFiboTransformerTesterConfig.get_dummy_inputs() (test_models_transformer_bria_fibo.py:69). Fix this →

2. TestBriaTransformerLoRAHotSwap.get_dummy_inputs doesn't seed its tensors

The override at the bottom of test_models_transformer_bria.py constructs tensors with bare randn_tensor(..., device=torch_device) — no generator. The LoRA mixin calls torch.manual_seed(0) in some places, but the inputs themselves are non-reproducible, which could cause flakiness. Consider adding a generator here too for consistency.

---

Nits

• The input_shape / output_shape property order is swapped in BriaTransformerTesterConfig vs the original — both still return (16, 4), so no functional change, but a small readability inconsistency.
• Removing the inline comments (# "ip_adapter" (cross-attention weights), # "image_proj" (ImageProjection layer weights)) from create_bria_ip_adapter_state_dict reduces readability slightly for a reader not familiar with IP-adapter structure.

---

The uses_custom_attn_processor = True removal is intentional and correct — the new ModelTesterMixin (in testing_utils/common.py) doesn't have test_set_attn_processor_for_determinism, so this flag isn't needed.

[sayakpaul]

I guess we can remove the expensive test suites from here provided the popularity of the model?

[sayakpaul]

Same comment as above.