Fix Glm4vModelTest::test_eager_matches_fa2_generate

[ydshieh]

What does this PR do?

This test is failing on GPU due to

FAILED tests/models/glm4v/test_modeling_glm4v.py::Glm4vModelTest::test_eager_matches_fa2_generate - RuntimeError: cu_seqlens_q must be on CUDA

see here

We need to change from

"image_grid_thw": torch.tensor([[1, patches_per_side, patches_per_side]] * self.batch_size),

to

"image_grid_thw": torch.tensor([[1, patches_per_side, patches_per_side]] * self.batch_size, device=torch_device),it

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: glm4v, glm4v_moe

[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.

[vasqu]

Makes sense, thank you

Just to be sure, other vlms don't have the issue? Just checked with qwen3 vl and they do the same in the prep function as you do here (with the fix)

[ydshieh]

I did a search (when opening the PR) "image_grid_thw": and only find these two without using torch_device. Should be good.

[vasqu]

Perfect, thx for looking 🙏