Fix test_script.py on TPU v2/v3

[vanbasten23]

What does this PR do?

Fixes #2479

• This PR fix the dataloader test: The same test accelerate test succeeds on TPU v4 but fails on v2/v3. The reason why it fails on TPU v2/v3 is multithreading. TPU v2/v3 involves multithreading: there are 4 processes and 8 TPU devices on the TPU v3-8, and each process spawns 2 threads for each device respectively. However TPU v4 and further TPU generations doesn't use multithreading. On TPU v2/v3, due to multithreading, we need to explicitly set the seed on each thread. Otherwise, each thread will permute the dataset indices in a different way and therefore generate the duplicate indices as shown in the gh issue. This PR explicitly set the generator for each thread and fix the test failure.
• This PR also fix a xm.set_replication issue for TPU v2/v3.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[vanbasten23]

cc @will-cromar @muellerzr @SunMarc

[muellerzr]

Thanks for the fantastic fix!

Looks like there's some style nits here, can you do make style; make quality after doing pip install -e .[quality]? Thanks!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[will-cromar]

Thanks @vanbasten23!

[muellerzr]

Same reasoning, that doesn't do anything. We need to adjust sampler, we should never be modifying the dataloader directly

[vanbasten23]

Actually modifying the dataloader has some impact on the new dataloader: #2542 (comment)

If I don't do dataloader.generator = generator, the test would fail a later test with an error AssertionError: Did not obtain the same model on CPU or distributed training.: https://gist.github.com/vanbasten23/ce6d416859b89c23cfd72e272d356504.

So I think it's needed.

[muellerzr]

Thanks for doing this! Overall it looks fine to me, I'd like if possible for you to verify that things will still work fine if you run the code with seedable_sampler set to True though before merging to know if we need to include it there as well.

(I'd rather wish we kept things without needing to modify dataloader directly, but if that's not possible then it's okay)

[vanbasten23]

Thanks for doing this! Overall it looks fine to me, I'd like if possible for you to verify that things will still work fine if you run the code with seedable_sampler set to True though before merging to know if we need to include it there as well.

(I'd rather wish we kept things without needing to modify dataloader directly, but if that's not possible then it's okay)

Sure. I set use_seedable_sampler=True at

accelerate/src/accelerate/test_utils/scripts/test_script.py

Line 215 in ee163b6

dl = prepare_data_loader(dl, state.device, state.num_processes, state.process_index, put_on_device=True)

and also at line 230 and the test accelerate test passed. I also change the default value of use_seeable_sampler to True in def prepare_data_loader and the test accelerate test passed too.