mark test_save_load_fast_init_from_base as is_flaky

[ydshieh]

What does this PR do?

Mark test_save_load_fast_init_from_base as is_flaky.

• This test is known to be flaky, see Update max_diff in test_save_load_fast_init_to_base #19849
• The level of flakiness seems to get higher after Attempting to test automatically the _keys_to_ignore. #20042
• ran 5 times, and all passed.
• TODO: check why Attempting to test automatically the _keys_to_ignore. #20042 makes this test more flaky.

[HuggingFaceDocBuilder]

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

[Narsil]

:(.

Any ways you found a solution to make the error reproducible ?

I'm really thinking the issue is in the test with the random weight drop, but since I could never get a reproducible way to make it fail it was hard to debug.

[ydshieh]

I guess the best first step is to run it 1000 times (in a loop) and see how many times it fails (before and after your PR 20042).

Sometimes it is difficult to get reproducible failure if it is flaky.

[sgugger]

As I commented before, this probably comes from models having weights initialized outside of the _init_weights function. A way to debug would be to drop somewhere which weight was randomly dropped when the test has failed (if it's printed for instance, we can look for it in the logs)