[Test Coverage] Generation compatibility for `TransformerBridge` models lacking native `generate()`

[jlarson4]

Summary

TransformerBridge.generate() works for most autoregressive HF models, but some bridge models (e.g. encoder-only architectures, certain multimodal models) currently fall through to an AttributeError/RuntimeError. The compatibility test silently skips when this happens, hiding the gap.

Affected test

• tests/unit/model_bridge/compatibility/test_utils.py:104 has inline pytest.skip("Generation not supported for this TransformerBridge model") inside an except (AttributeError, RuntimeError):

What the test verifies

That model.generate(prompt, max_new_tokens=5) returns one of (str, list, torch.Tensor) for every parameterized model.

Acceptance criteria

• Models that legitimately don't support generation (BERT encoder-only) are filtered out of the parameterized set rather than handled with a runtime except + skip.
• The inline except + pytest.skip is removed.
• Remaining tests pass without a silent fallback.

Where to start

• Find the parameterized model set (@pytest.mark.parametrize or params= on the fixture).
• Categorize each by whether generation is supported. Filter the non-supporting ones out at parametrize time.
• Investigate and fix models that should support generation, but currently raise.