vllm - 💡(How to fix) Fix [Bug]: Multiple model files still trigger ImageProcessorFast DeprecationWarning post transformers v5.4 refactor [1 pull requests]

What

Static imports of deprecated *ImageProcessorFast symbols still trigger transformers DeprecationWarning across four model files under vllm/model_executor/models/. #42700 fixed the same family for qwen3_vl.py (one file); the other four were not covered.

Reproduction

$ docker run --rm --gpus '"device=1"' --entrypoint python3 \
    vllm/vllm-openai:v0.20.0 -W default -c \
    "import vllm.model_executor.models.lfm2_vl"
[transformers] Accessing `Lfm2VlImageProcessorFast` from `.models.lfm2_vl.image_processing_lfm2_vl`.
Returning `Lfm2VlImageProcessor` instead. Behavior may be different and this alias
will be removed in future versions.

Same form of warning fires on each of:

Sanity check the inverse: vllm/transformers_utils/processors/{qwen_vl,glm4v}.py define internal classes that inherit BaseImageProcessorFast. The same probe does not raise any warning for these — BaseImageProcessorFast is still the canonical export. So the surface is import-site only, four files.

Background

• huggingface/transformers#43514 (merged 2026-03-19, "🚨🚨 Refactor Image Processors to support different backends") dropped the Fast suffix for image processors and reshaped the underlying class layout.
• vLLM's requirements/common.txt already pins around the breaking transitions (!= 5.4.*, != 5.5.0), but the import-site warnings on the four files above remain visible on every load.

Question for maintainers

Same as #42688's resolution — swap the imports for the un-suffixed canonical names across these four files, mirroring #42700's pattern.

I'll send the follow-up PR shortly. Happy to scope it differently if there's an internal sweep planned.