test: filter generation compatibility models

[axelray-dev]

Fixes #1328

What changed

Replaced the runtime try/except + pytest.skip pattern in test_generation_compatibility with a dedicated generating_model fixture that only includes generation-capable models.

• Added GENERATING_MODELS class constant (currently ["gpt2"])
• Added generating_model_name and generating_model fixtures
• Changed test_generation_compatibility to use the new generating_model fixture
• Removed the inline except (AttributeError, RuntimeError) + pytest.skip

Why filtering is better than runtime skip

The previous approach silently skipped the generation test when AttributeError or RuntimeError was raised, hiding the fact that some models don't support generation. By filtering at the fixture level:

• Generation support is explicit in the test parameterization
• Tests that should run actually run (no silent skips)
• Adding a non-generating model to the main model_name fixture won't accidentally suppress the generation test
• Future contributors can see which models are expected to support generation by looking at GENERATING_MODELS

Checks run

uv run pytest tests/unit/model_bridge/compatibility/test_utils.py -v -x
# 14 passed in 77.38s (including test_generation_compatibility[gpt2])

Skipped checks

• Full test suite not run (would take significantly longer and download many models)
• No production code changed, so only the targeted test file was verified

[jlarson4]

Hi @axelray-dev!

A few requests:

1. Could you rebase this to origin/dev to clean up the diff? There are old commits that are leaking in because your code was based on an outdated version of origin/main
2. I'd like to adjust the approach here. The GENERATING_MODELS allowlist solution here is functional, but it creates another static list of models that will need to be maintained and has the potential to be fragile if forgotten. Could we instead update the ArchitectureAdapter base class to add a supports_generation flag that defaults to true. We can then set it to false false on models that don't support generation. The test suite can check the flag for the skip, rather than managing an allowlist. It also allows us access to the supports_generation information if we need it for future features.

[axelray-dev]

Updated. Here is what changed:

1. Rebased onto origin/dev. The diff is now clean (4 files, +12/-8).
2. Added supports_generation: bool = True as a class attribute on ArchitectureAdapter, next to the existing applicable_phases flag.
3. Set supports_generation = False on BertArchitectureAdapter and HubertArchitectureAdapter (encoder-only models).
4. Replaced the GENERATING_MODELS allowlist and separate fixtures in the test with a runtime pytest.skip check on model.adapter.supports_generation.

Tests run:

uv run pytest tests/unit/model_bridge/compatibility/test_utils.py -v

All 14 tests passed. The diff is additions-only for the adapter and test files (no unrelated changes).

[jlarson4]

Excellent work, thank you @axelray-dev