Fix: Conditionally import torch.distributed.fsdp in trainer_seq2seq.py

[0xDELUXA]

What does this PR do?

torch.distributed.fsdp is not available in all PyTorch builds (for example, Windows ROCm). Importing it unconditionally at the top level causes an immediate crash with:

ModuleNotFoundError: No module named 'torch._C._distributed_c10d'; 'torch._C' is not a package

This makes import transformers fail entirely on those builds, even for users who have no intention of using FSDP or distributed training.

Changes

• Moved from torch.distributed.fsdp import FullyShardedDataParallel to after the unconditional imports, guarded with if torch.distributed.is_available()
• Added the same guard to the isinstance(self.model, FullyShardedDataParallel) check to avoid a NameError at runtime on non-distributed builds

Prior art

A previous fix for the same class of issue in this repo was made in:

• huggingface/transformers#40038

The same fix was applied to huggingface/diffusers in:

• huggingface/diffusers#12420
• huggingface/diffusers#12425

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?

@SunMarc @3outeille

[3outeille]

lgtm

[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.