Fix OOM in CI by reducing batch size in VLM SFT tests

[albertvillanova]

Fix OOM in CI by reducing batch size in VLM SFT tests.

Partial fix for:

• CI often fails with torch.OutOfMemoryError: CUDA out of memory #5207

Motivation

VLM training tests in test_sft_trainer.py were running with the default per_device_train_batch_size=8. For Gemma3, with vocab_size=262208 (production-scale, never reduced for tiny models) and mm_tokens_per_image=256, each training step computes logits of shape [8, 279, 262208].

PyTorch needs several float32 copies of this tensor for log-softmax and its gradient, pushing peak GPU memory to ~9 GiB per worker. With 4 parallel pytest-xdist workers this caused CUDA out-of-memory errors for other concurrent tests.

Solution

Set per_device_train_batch_size=1 in test_train_vlm, test_train_vlm_multi_image, and test_train_vlm_prompt_completion, following the pattern already used in test_train_vlm_gemma_3n. This drops peak GPU memory to ~1.1 GiB per worker for Gemma3, leaving ample headroom for parallel execution.

---

Note

Low Risk
Test-only change that lowers batch size to avoid CI OOMs; no production code paths are modified.

Overview
Reduces GPU memory pressure in vision-language SFT integration tests by explicitly setting per_device_train_batch_size=1 in test_train_vlm, test_train_vlm_multi_image, and test_train_vlm_prompt_completion.

This prevents CI CUDA OOMs during parallel test execution while keeping the VLM-specific max_length=None behavior unchanged.

^{Reviewed by Cursor Bugbot for commit 0480d77. Bugbot is set up for automated code reviews on this repo. 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.

[qgallouedec]

It seems to work!