_tf_available for customized built tensorflow

[kevint324]

System Info

n/a

Who can help?

@Rocketknight1

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

  File "virtualenv_mlu/lib/python3.8/site-packages/transformers/pipelines/base.py", line 212, in infer_framework_load_model
    raise RuntimeError(
RuntimeError: At least one of TensorFlow 2.0 or PyTorch should be installed. To install TensorFlow 2.0, read the instructions at https://www.tensorflow.org/install/ To install PyTorch, read the instructions at https://pytorch.or

Expected behavior

https://github.com/huggingface/transformers/blob/02b176c4ce14340d26d42825523f406959c6c202/src/transformers/utils/import_utils.py#L63L75

I built a tensorflow-xxu for our in house accelerator and tried to run the transformer example.
I got the RuntimeError indicates tf is not available.
Currently the _tf_available has a hard-coded candidates list.
I'm not suire if extending the existing list is a good idea.
Maybe adding some runtime flexibility would be better?

Thanks
Kevin

[ydshieh]

Hi @kevint324 I think it (extending the list ) is fine if you would work with a specific transformers version. But it would be a bit tedious if you want to use newer versions constantly.

cc @Rocketknight1 for the idea regarding adding some runtime flexibility.

[ydshieh]

Changed the tag to Feature request instead :-)

[Rocketknight1]

Hi @kevint324, I filed a PR that might resolve this, but I want to check with other maintainers that it's okay before I merge it. In the meantime, can you try it out? Just run pip install git+https://github.com/huggingface/transformers.git@allow_force_tf_availability, then set the environment variable FORCE_TF_AVAILABLE=1 before running your code, and it should skip those checks now.

[kevint324]

Yes, it works.
Thanks for the quick fix.