[albertvillanova]

Remove xfail for Qwen3VL CI tests.

After investigation, I found that the bug (huggingface/transformers#43334) never landed on an official release, just in v5.0.0.dev0. It was fixed in v5.0.0 by:

• Fix: adding pad_token_id in Qwen3VLTextConfig transformers#43398

In transformers v5.0.1, the bug was fixed for other models besides qwen3_vl: glm4v, glm_image, glm_ocr, phi, qwen2_5_omni, and qwen3_vl_moe. However, we don't test these models in our CI. The fixing PR was:

• Fix failing tests due to no attribute pad_token_id transformers#43453

I have tested locally with transformers v5.0.0 and the tests pass.

Follow-up to:

• CI fails with dev dependencies: AttributeError: 'Qwen3VLTextConfig' object has no attribute 'pad_token_id' #4842

Revert:

• Mark Qwen3VL tests as xfail for transformers 5.0.x #5029

---

Note

Low Risk
Low risk because changes are limited to test markers, but it may surface new CI failures if upstream transformers regressions reappear in certain versions.

Overview
Removes the pytest.mark.xfail conditions for trl-internal-testing/tiny-Qwen3VLForConditionalGeneration in the VLM test parametrizations for DPOTrainer, SFTTrainer, and forward_masked_logits utils tests.

As a result, Qwen3-VL now runs as a normal CI test (still skipif-gated on transformers>=4.57.0) instead of being treated as an expected failure for specific transformers version ranges.

^{Written by Cursor Bugbot for commit 304e22c. This will update automatically on new commits. Configure here.}

[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.

[albertvillanova]

The failing test is unrelated. It is caused by flaky CI and I opened a PR to fix it:

• Fix flaky CI test_rloo[fsdp2]: Replace non-deterministic xfail with skipif for transformers 5.4.0 #5403