[Tests common] Fix flaky test

[patrickvonplaten]

The test test_model_outputs_equivalence fails quite often at the moment because of a problem with nan - nan. This should solve the issue.

Also added more explicit error message in case test error occurs.

Flaky error happens here for example: https://app.circleci.com/pipelines/github/huggingface/transformers/10798/workflows/44e689b2-f4b3-49be-88b3-a5b214eac6c5/jobs/75173

[sgugger]

OK for the temp fix. It might warrant some deeper investigation though.

[codecov]

Codecov Report

Merging #6593 into master will increase coverage by 0.01%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #6593      +/-   ##
==========================================
+ Coverage   79.42%   79.43%   +0.01%     
==========================================
  Files         156      156              
  Lines       28127    28127              
==========================================
+ Hits        22339    22343       +4     
+ Misses       5788     5784       -4     

Impacted Files	Coverage Δ
src/transformers/modeling_tf_albert.py	21.47% <0.00%> (-69.44%)	⬇️
src/transformers/tokenization_roberta.py	76.71% <0.00%> (-21.92%)	⬇️
src/transformers/tokenization_utils_base.py	86.58% <0.00%> (-7.19%)	⬇️
src/transformers/tokenization_transfo_xl.py	38.73% <0.00%> (-3.76%)	⬇️
src/transformers/tokenization_utils_fast.py	92.14% <0.00%> (-2.15%)	⬇️
src/transformers/tokenization_openai.py	82.57% <0.00%> (-1.52%)	⬇️
src/transformers/tokenization_bert.py	91.07% <0.00%> (-0.45%)	⬇️
src/transformers/tokenization_utils.py	90.00% <0.00%> (-0.41%)	⬇️
src/transformers/file_utils.py	82.44% <0.00%> (+0.25%)	⬆️
src/transformers/modeling_tf_utils.py	87.29% <0.00%> (+0.97%)	⬆️
... and 2 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ab42d74...d243891. Read the comment docs.

[sshleifer]

Long term, I would rather fill nan with a sentinel value like -11.27 (my negative birthday, dont forget!) and check equality but I am also fine with the temp fix.

[patrickvonplaten]

Ok pinging @LysandreJik here that we should take a look next week. Will note it down as well.