[docs] update set-seed

[faaany]

What does this PR do?

The current set-seed in documentation only includes the cuda part, which is out-dated. This PR updates it with the current implementation.

@muellerzr

[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.

[faaany]

@muellerzr @BenjaminBossan

[BenjaminBossan]

Thanks for updating the docs.

IMO, it would also be fine to remove the code snippet completely (as it's just a matter of time until it gets out of date again) and just explain in words what's happening.

[faaany]

Thanks for updating the docs.

IMO, it would also be fine to remove the code snippet completely (as it's just a matter of time until it gets out of date again) and just explain in words what's happening.

I think code is easier to read and understand. Only using cuda is a bit unfair. Maybe I can update other devices in a comment?

[BenjaminBossan]

I think code is easier to read and understand. Only using cuda is a bit unfair. Maybe I can update other devices in a comment?

My main concern is mostly the example and the actual code getting out of sync. It's not a big deal, would just be nice to prevent it.

[muellerzr]

Thanks, there is always a fear of what would happen if code gets out of sync, esp as even this code doesn't quite fully align to what we're up to. (e.g. there's now a by_index).

While it's not per-se "fair", I'm not sure if it's more or less readable now than before for our docs. It is a concept guide, so we can show more code. Just wish we could show the source itself.

I think we could get by just saying:

    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed) # or torch.npu.manual_seed_all, etc
    # ^^ safe to call this function even if cuda/backend is not available
    if is_torch_xla_available():
        xm.set_rng_state(seed)

Since we care more about manual_seed_all than the backend specifically

[faaany]

Thanks for the feedback! Agree. Code looks much cleaner now. @muellerzr @BenjaminBossan

[muellerzr]

Thanks!