System Info

• transformers version: 4.57.3
• Platform: Linux-6.8.0-52-generic-x86_64-with-glibc2.35
• Python version: 3.12.9
• Huggingface_hub version: 0.34.3
• Safetensors version: 0.5.3
• Accelerate version: 1.9.0
• Accelerate config: - compute_environment: LOCAL_MACHINE
  - distributed_type: MULTI_GPU
  - mixed_precision: bf16
  - use_cpu: False
  - debug: False
  - num_processes: 8
  - machine_rank: 0
  - num_machines: 1
  - gpu_ids: all
  - rdzv_backend: static
  - same_network: True
  - main_training_function: main
  - enable_cpu_affinity: True
  - downcast_bf16: no
  - tpu_use_cluster: False
  - tpu_use_sudo: False
  - tpu_env: []
• DeepSpeed version: 0.17.4
• PyTorch version (accelerator?): 2.7.1+cu128 (CUDA)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?: no
• Using GPU in script?: no
• GPU type: NVIDIA H100 80GB HBM3

Who can help?

@zucchini-nlp

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

These are pure python logic bugs, they exist in multiple processors.

1. The video branch of _get_num_multimodal_tokens calls get_number_of_video_patches() but its not implemented for any of these processors.
2. The video branch of _get_num_multimodal_tokens uses merge_size without defining it in the video branch.

Affected processors/files:

• src/transformers/models/ernie4_5_vl_moe/processing_ernie4_5_vl_moe.py
• src/transformers/models/glm46v/processing_glm46v.py
• src/transformers/models/glm4v/processing_glm4v.py
• src/transformers/models/qwen2_5_vl/modular_qwen2_5_vl.py
• src/transformers/models/qwen2_5_vl/processing_qwen2_5_vl.py
• src/transformers/models/qwen2_vl/processing_qwen2_vl.py
• src/transformers/models/qwen3_vl/processing_qwen3_vl.py
• src/transformers/models/video_llama_3/processing_video_llama_3.py

Minimal repro (any affected processor):

# assumes you have a processor instance for any of the above models
processor._get_num_multimodal_tokens(video_sizes=[(8, 224, 224)])

This will results to 2 issues:

1. First issue: AttributeError: 'Qwen3VLVideoProcessor' object has no attribute 'get_number_of_video_patches'. All of the mentioned (affected processors) use this function, but its not defined -in comparison with the image counterpart get_number_of_image_patches which exists and all tests pass -video branch doesn't have tests.

2. Second issue: After you solve this by implementing 'get_number_of_video_patches' , then the undefined merge_size will arise: UnboundLocalError: cannot access local variable 'merge_size' where it is not associated with a value

Expected behavior

Calling _get_num_multimodal_tokens(video_sizes=...) should not crash.

The video branch should define merge_size (same as image branch) and get_number_of_video_patches (similarly as image branch).

Also in tests I think it would be a good idea to add a unit test that calls _get_num_multimodal_tokens(video_sizes=[...]) so CI actually exercises the video path.

I’m happy to send a PR with the fixes + tests.