[fsmt test] basic config test with online model + super tiny model

[stas00]

This PR does:

• Seeing the issue of [Dependencies|tokenizers] Make both SentencePiece and Tokenizers optional dependencies #7659 I realized fsmt didn't have a basic non-slow test that loads the tokenizer via online model files. Luckily a totally unrelated examples test caught this issue in that PR, so adding a very simple quick test in the main test suite, so that it runs by the normal CI.
• while I was building this test, I needed a new tiny model, so I refined the script https://github.com/huggingface/transformers/blob/master/scripts/fsmt/fsmt-make-tiny-model.py and made a new one that creates a 50 times smaller model, so we are now at 60KB, instead of 3MB.

@LysandreJik, @sshleifer

[thomwolf]

Hi Stas, thanks, #7659 will fix this (we will now require at least one example checkpoint for each tokenizer and we test it automatically).

[stas00]

@thomwolf, I'm not certain why you closed this. This is a tokenizer test that is needed - the issue caught in examples was just a flag that there was a missing test in the normal tests.

Actually, not only it's needed, I will have to expand this test to verify that it doesn't get the hardcoded default values, but fetches the correct values from tokenizer_config.json. And to change the default values to be different from TINY_FSMT, since otherwise it won't be testing the right thing.

Feel free to add it as part of #7659 but please make sure you used different from TINY_FSMT hardcoded defaults.

I hope this makes sense.

[stas00]

Hmm, but you copied TINY_FSMT in 1885ca7, how will then the tests check it can fetch that data if this is now hardcoded? I am not following this. Do I need to create TINY_FSMT2 with different values?

I haven't read the new code in depth, but my gut feeling is that the defaults may mask a problem.

[thomwolf]

Hi stas, I'll let you read the new code and then we can have a look together.

The basic idea is that we now require a full and working checkpoint for the tokenizers to be fully tested in various conditions and the slow vs. fast compared.

The question of testing that tokenizers load and use tokenizer_config.json is another question and we should indeed address it in a subsequent PR if it's not addressed already indeed.

[stas00]

That works.

But please re-open this PR, since we need it anyway. I will add more changes to it after your big PR merge to ensure that the loading of the tokenizer is properly tested.

[stas00]

This PR is complete now.

[LysandreJik]

LGTM