Avoid failure TFBlipModelTest::test_pipeline_image_to_text

[ydshieh]

What does this PR do?

Currently,

FAILED tests/models/blip/test_modeling_tf_blip.py::TFBlipModelTest::test_pipeline_image_to_text - NameError: name 'torch' is not defined

is failing on main after (#31342)

See https://app.circleci.com/pipelines/github/huggingface/transformers/97187/workflows/81b63940-fd42-4a6a-aafe-404301fafc1d/jobs/1285954

There is a slight difference between Pipeline.__init__ and pipeline: the later one has

    if isinstance(torch_dtype, str) and hasattr(torch, torch_dtype):
       torch_dtype = getattr(torch, torch_dtype)

which we should have a check on if torch is available. I will open another PR for that.

---

Also for 2 Failed: Timeout >120.0s: the fix for them is the same as their non _fp16 versions.

FAILED tests/models/vits/test_modeling_vits.py::VitsModelTest::test_pipeline_feature_extraction_fp16 - Failed: Timeout >120.0s

FAILED tests/models/efficientnet/test_modeling_efficientnet.py::EfficientNetModelTest::test_pipeline_image_feature_extraction_fp16 - Failed: Timeout >120.0s

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

Thanks for handling!