transformers

27e7f6c8 - Add an opt-in per-frame pixel cap (cap_pixels_per_frame) to the Qwen3-VL video processor (#48071)

Add an opt-in per-frame pixel cap (cap_pixels_per_frame) to the Qwen3-VL video processor (#48071) * Add max_pixels_per_frame to Qwen3-VL video processor The video pixel budget (size.longest_edge) is spent across all sampled frames, so clips that sample fewer than max_frames frames keep near-native per-frame resolution and a short clip can cost almost as many tokens as a long video. The new optional kwarg caps the effective budget at num_frames * max_pixels_per_frame, making token cost scale with clip duration when set. Default None preserves current behavior. * Floor the frame count at 32 when applying max_pixels_per_frame Serving stacks probe processors with tiny synthetic clips (vLLM's boot memory profiling uses 2-frame dummies); capping those at num_frames * max_pixels_per_frame collapses them to near-minimum resolution. Apply the cap against max(num_frames, 32) so short and synthetic clips keep a usable floor while longer videos stay duration-proportional. * Regenerate CohereCompass video processor from modular source CohereCompassVideoProcessor inherits from Qwen3VLVideoProcessor, so the modular conversion check picks up the new max_pixels_per_frame kwarg. Commit the regenerated file so it matches what check_modular_conversion.py produces. * Replace max_pixels_per_frame with a cap_pixels_per_frame boolean Per review: instead of a numeric knob, a boolean that applies the qwen-vl-utils per-frame cost formula (768-patch frame ceiling or the budget's even share per frame, floored just above min_pixels). Unset warns that the capped behavior becomes the default in v5.22, when the argument will be removed; passing False keeps today's behavior silently. * Apply ruff formatting * Spell Qwen3VL in the deprecation warning so modular conversion renames it * Add cap_pixels_per_frame to the Qwen2VL video processor * Regenerate VideoLlama3 video processor from modular source * Address review: parameterize caps as class attrs, inherit Qwen2VL _preprocess, drop stray attrs - max_video_tokens (768) and video_total_seq_len (128000) become saved class attributes instead of magic numbers - Qwen3VLVideoProcessor now inherits Qwen2VLVideoProcessor and overrides only resize and sample_frames; _preprocess (with the deprecation warning) is copied by modular conversion - VideoLlama3 no longer inherits the cap attributes or kwargs; its resize already spreads the budget across frames - document min_frames/max_frames in Qwen3VL kwargs (silences the auto_docstring runtime lint) * Override get_num_of_video_patches for Qwen3VL's temporal smart_resize The copy inherited from Qwen2VL called this module's t*h*w smart_resize with per-frame arguments. Override with the temporal signature and apply the per-frame cap so the patch count matches what resize actually produces. * small unification in naming --------- Co-authored-by: raushan <raushan@huggingface.co>