Error during video handling with Gemma4 (E2B-Instruct)

[anand-kumar-consulting]

System Info

• transformers version: 5.15.0.dev0
• Platform: Linux-6.18.33.1-microsoft-standard-WSL2-x86_64-with-glibc2.43
• Python version: 3.13.14
• Huggingface_hub version: 1.24.0
• Safetensors version: 0.8.0
• Accelerate version: 1.14.0
• Accelerate config: not found
• DeepSpeed version: not installed
• PyTorch version (accelerator?): 2.13.0+cu132 (CUDA)
• Using distributed or parallel set-up in script?: No
• Using GPU in script?: Not clear.
• GPU type: NVIDIA GeForce RTX 4080 Laptop GPU

Who can help?

@molbap @guarin

The issue is with the vision model (Gemma4-E2B-IT)

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

File ".../transformers/models/gemma4/modeling_gemma4.py", line 2344, in forward
video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)
                 ^^^^^^^^^^^^^^^^^
AttributeError: 'tuple' object has no attribute 'to'

The error is resolved on changing the code to:

video_features = video_features[0].to(inputs_embeds.device, inputs_embeds.dtype)

Expected behavior

The code works well by making the aforementioned change (which should be default behaviour).

[jayzuccarelli]

Reproduced this — video_features is a tuple after torch.split, so the .to() call crashes. The image branch already does torch.cat(...).to(...); the video branch needs the same. Happy to open a PR with the fix + a regression test if that works for you.

[guarin]

Looks like this regression was accidentally introduced here: #46405 Will create a PR with the fix, thanks for raising the issue!