Fix tests broken by a local GPU

[brendanlong]

Description

Two of the tests fail if you have a local GPU.

test_apertus failed because it creates CPU tensors and then compares them to tensors created on the default device. I fixed this by updating the model config to use the CPU as well.

test_hooked_encoder has a CUDA-only test (test_cuda) which expects an undefined mlm_tokens fixture. Presumably it passes in CI because CUDA isn't available. I updated this to use tokens instead.

Before:

> poetry run pytest tests/unit/pretrained_weight_conversions/test_apertus.py tests/acceptance/test_hooked_encoder.py 
========================================================= test session starts =========================================================
platform linux -- Python 3.12.9, pytest-8.3.5, pluggy-1.5.0
rootdir: /home/brendanlong/workspace/TransformerLens
configfile: pyproject.toml
plugins: jaxtyping-0.2.19, doctestplus-1.3.0, nbval-0.10.0, cov-5.0.0, anyio-4.5.2, typeguard-4.4.0
collected 29 items                                                                                                                    

tests/unit/pretrained_weight_conversions/test_apertus.py ........F...                                                           [ 41%]
tests/acceptance/test_hooked_encoder.py ................E                                                                       [100%]

=============================================================== ERRORS ================================================================
_____________________________________________________ ERROR at setup of test_cuda _____________________________________________________
file /home/brendanlong/workspace/TransformerLens/tests/acceptance/test_hooked_encoder.py, line 224
  @pytest.mark.skipif(not torch.cuda.is_available(), reason="Requires a CUDA device")
  def test_cuda(mlm_tokens):
E       fixture 'mlm_tokens' not found
>       available fixtures: anyio_backend, anyio_backend_name, anyio_backend_options, cache, capfd, capfdbinary, caplog, capsys, capsysbinary, cov, doctest_namespace, huggingface_bert, monkeypatch, no_cover, our_bert, pytestconfig, record_property, record_testsuite_property, record_xml_attribute, recwarn, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory, tokenizer, tokens
>       use 'pytest --fixtures [testpath]' for help on them.

/home/brendanlong/workspace/TransformerLens/tests/acceptance/test_hooked_encoder.py:224
============================================================== FAILURES ===============================================================
___________________________________ TestApertusWeightConversion.test_xielu_params_fallback_defaults ___________________________________

self = <test_apertus.TestApertusWeightConversion object at 0x753ad8da8080>

    def test_xielu_params_fallback_defaults(self):
        """When activation params aren't found, defaults should be used."""
        cfg = make_cfg()
        model = make_mock_model(cfg, has_act_fn=False)
        # Also remove alternate attribute paths
        layer = model.model.layers[0]
        if hasattr(layer.mlp, "act"):
            del layer.mlp.act
        if hasattr(layer.mlp, "alpha_p"):
            del layer.mlp.alpha_p
    
        sd = convert_apertus_weights(model, cfg)
>       torch.testing.assert_close(
            sd["blocks.0.mlp.act_fn.alpha_p"], torch.tensor(0.8, dtype=cfg.dtype)
        )
E       AssertionError: The values for attribute 'device' do not match: cuda:0 != cpu.

tests/unit/pretrained_weight_conversions/test_apertus.py:157: AssertionError
---------------------------------------------------------- Captured log call ----------------------------------------------------------
WARNING  transformer_lens.pretrained.weight_conversions.apertus:apertus.py:117 XIeLU activation parameters not found in layer 0, using defaults
========================================================== warnings summary ===========================================================
../../.cache/pypoetry/virtualenvs/transformer-lens-MODGY9P9-py3.12/lib/python3.12/site-packages/jaxtyping/import_hook.py:106: 52 warnings
  /home/brendanlong/.cache/pypoetry/virtualenvs/transformer-lens-MODGY9P9-py3.12/lib/python3.12/site-packages/jaxtyping/import_hook.py:106: DeprecationWarning: ast.Str is deprecated and will be removed in Python 3.14; use ast.Constant instead
    elif isinstance(child, ast.Expr) and isinstance(child.value, ast.Str):

transformer_lens/SVDInterpreter.py:26
  /home/brendanlong/workspace/TransformerLens/transformer_lens/SVDInterpreter.py:26: InstrumentationWarning: @typechecked only supports instrumenting functions wrapped with @classmethod, @staticmethod or @property -- not typechecking transformer_lens.SVDInterpreter.SVDInterpreter.get_singular_vectors
    @typechecked

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
======================================================= short test summary info =======================================================
FAILED tests/unit/pretrained_weight_conversions/test_apertus.py::TestApertusWeightConversion::test_xielu_params_fallback_defaults - AssertionError: The values for attribute 'device' do not match: cuda:0 != cpu.
ERROR tests/acceptance/test_hooked_encoder.py::test_cuda
========================================= 1 failed, 27 passed, 53 warnings, 1 error in 15.02s =========================================

After:

> poetry run pytest tests/unit/pretrained_weight_conversions/test_apertus.py tests/acceptance/test_hooked_encoder.py 
========================================================= test session starts =========================================================
platform linux -- Python 3.12.9, pytest-8.3.5, pluggy-1.5.0
rootdir: /home/brendanlong/workspace/TransformerLens
configfile: pyproject.toml
plugins: jaxtyping-0.2.19, doctestplus-1.3.0, nbval-0.10.0, cov-5.0.0, anyio-4.5.2, typeguard-4.4.0
collected 29 items                                                                                                                    

tests/unit/pretrained_weight_conversions/test_apertus.py ............                                                           [ 41%]
tests/acceptance/test_hooked_encoder.py .................                                                                       [100%]

========================================================== warnings summary ===========================================================
../../.cache/pypoetry/virtualenvs/transformer-lens-MODGY9P9-py3.12/lib/python3.12/site-packages/jaxtyping/import_hook.py:106: 52 warnings
  /home/brendanlong/.cache/pypoetry/virtualenvs/transformer-lens-MODGY9P9-py3.12/lib/python3.12/site-packages/jaxtyping/import_hook.py:106: DeprecationWarning: ast.Str is deprecated and will be removed in Python 3.14; use ast.Constant instead
    elif isinstance(child, ast.Expr) and isinstance(child.value, ast.Str):

transformer_lens/SVDInterpreter.py:26
  /home/brendanlong/workspace/TransformerLens/transformer_lens/SVDInterpreter.py:26: InstrumentationWarning: @typechecked only supports instrumenting functions wrapped with @classmethod, @staticmethod or @property -- not typechecking transformer_lens.SVDInterpreter.SVDInterpreter.get_singular_vectors
    @typechecked

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
================================================== 29 passed, 53 warnings in 15.71s ===================================================

I'm also able to run the full CI suite:

> poetry run pytest tests/unit tests/integration tests/acceptance
[...]
94 passed, 30 skipped, 59 warnings in 327.31s (0:05:27)

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[brendanlong]

The CI test failure seems to be caused by HuggingFace rate limiting, which is confusing since I thought you had HF_TOKEN set. Either way it's not caused by my changes and should work if you re-run them.

[brendanlong]

I had to remove .type since cfg.device is a str now. There's a pre-existing type issue here since cfg.device has type str | None so the old code shouldn't have compiled.

[brendanlong]

See #1230