Issue evidence · Run · huggingface/trl

Online DPO VLM training fails for processors with 5D pixel values, such as Idefics2

GitHub issue: https://github.com/huggingface/trl/issues/7702

Reproduced

OnlineDPOTrainer fails on the first step for VLMs whose processors return 5D pixel_values, such as Idefics2.

What Badgr ran

Ran the issue #7701 reproducer with tiny-Idefics2ForConditionalGeneration on TRL main (installed from git at run time), on a Mac CPU, with a one-step run.

What came back

  • Idefics2: RuntimeError: Number of dimensions of repeat dims can not be smaller than number of dimensions of tensor, the error in the issue.
  • The Llava control failed with a different error (RuntimeError: Unexpected floating ScalarType in at::autocast::prioritize), which looks specific to the CPU run.

The supported-configuration control did not pass here, so this does not show the Llava path working. No GPU was used.

Checked 2026-10-14 in Badgr’s local development environment.

← All issue evidence