Hotfix CI: xfail the GRPO continuous batching test under DataParallel

[albertvillanova]

Hotfix:

• CI fails for slow tests on multiple GPUs: RuntimeError: Expected all tensors to be on the same device, but got index is on cuda:1, different from other tensors on cuda:0 #7421

This PR marks as xfail the GRPO continuous batching slow test when several accelerators are visible, where Trainer wraps the model in nn.DataParallel.

Motivation

Since #7389, GRPO patches the fused LM head by binding a new forward to the model instance. nn.DataParallel copies that bound method into every replica, so replica 1 still runs on the device-0 weights and the multi-GPU slow job fails:

RuntimeError: Expected all tensors to be on the same device, but got index is on cuda:1, different from other tensors on cuda:0

This is the same root cause as the SFT failures in #6836. See the investigation in #7421.

Solution

Apply the existing xfail_data_parallel marker, as done for SFT in #6841. It is conditional on more than one accelerator being visible, so the test keeps reporting normally in the single-GPU job.

Changes

• Apply xfail_data_parallel to test_train_with_transformers_continuous_batching in tests/test_grpo_trainer.py

---

Note

Low Risk
Test-only CI adjustment using an existing marker; no production or trainer logic changes.

Overview
Marks test_train_with_transformers_continuous_batching with the shared @xfail_data_parallel marker so CI stops failing when multiple accelerators are visible and Hugging Face Trainer wraps the model in nn.DataParallel.

That path hits a known TRL limitation: GRPO’s fused LM head rebinds forward on one module instance, which DataParallel replicas mishandle (cross-device tensor errors). Single-GPU runs are unchanged—the marker only applies when more than one device is present.

^{Reviewed by Cursor Bugbot for commit 80b1069. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

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.

[albertvillanova]

Merging this PR as a CI hotfix.