Fix some TF GPT-J CI testings

[ydshieh]

What does this PR do?

Fix some TF-GPT-J CI testing (scheduled)

• test_mixed_precision: require some casting
• test_saved_model_creation and test_saved_model_creation_extended: require shape_list instead of shape
• test_model_from_pretrained: skip for now otherwise GPU OOM

With the changes this PR, only the following test fails: test_gptj_sample_max_time: for example

transformers/tests/gptj/test_modeling_tf_gptj.py

Line 413 in c85547a

model.generate(input_ids, do_sample=False, max_time=MAX_TIME, max_length=256)

the PT gives a quite short generation sequence (say 19), while TF gives a sequence of length 256, and it takes much more time and therefore fails the tests.

I feel this remaining issue is better to be addressed in another PR.

[ydshieh]

I need to check why

@unittest.skipIf(len(tf.config.list_physical_devices("GPU")) > 0, "skip testing on GPU for now to avoid GPU OOM.")

causes problems in other tests (torch, pipeline etc ...)

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Thanks for fixing!

[gante]

Thank for looking at these, @ydshieh 🙏

[gante]

Regarding the max_time/generate test -- when I first reviewed the GPT-J PR, I had little understanding of the current state of TF generate. Now I can tell that this test makes no sense :D Contrarily to PT's generate, TF generate has no time-based stopping criteria, so it is natural that the test fails. I'd remove it.

[ydshieh]

Regarding the max_time/generate test -- when I first reviewed the GPT-J PR, I had little understanding of the current state of TF generate. Now I can tell that this test makes no sense :D Contrarily to PT's generate, TF generate has no time-based stopping criteria, so it is natural that the test fails. I'd remove it.

OK, thank you for the feedback.

But just curious (off-topic): TF generate has no time-based stopping criteria --> do we plan to support this in the future ??

(not very sure, but I remembered before TF can generate short sequences too. And if TF can't stop earlier, it looks like a quite big drawback ..? Anyway, we shouldn't discuss this generation thing in this PR.)

[gante]

But just curious (off-topic): TF generate has no time-based stopping criteria --> do we plan to support this in the future ??

(not very sure, but I remembered before TF can generate short sequences too. And if TF can't stop earlier, it looks like a quite big drawback ..? Anyway, we shouldn't discuss this generation thing in this PR.)

The plan we have for the refactoring does not mention extras like the stopping criteria, so I can only tell that it probably won't happen in the next 2-3 months :) We can generate short sentences with TF if we pass the max_length argument, where generate() generates up to max_length tokens.