ONNX Export Support for Qwen2,2.5,3 and Gemma3 VLM

[satabios]

No description provided.

[satabios]

Files Modified

1. optimum/exporters/onnx/input_generators.py
   Added imports: DEFAULT_DUMMY_SHAPES, DummyInputGenerator
   Added DummyQwen2VLVisionInputGenerator class that generates:
   pixel_values: pre-flattened patches (total_patches, 1176) — Qwen2-VL's non-standard format where 1176 = 3 × 2 × 14 × 14
   image_grid_thw: grid dimensions (num_images, 3) with [grid_t, grid_h, grid_w] per image
2. optimum/exporters/onnx/model_configs.py
   Added import of DummyQwen2VLVisionInputGenerator
   Added Qwen2VLOnnxConfig class registered as "qwen2_vl" with tasks: feature-extraction, feature-extraction-with-past, text-generation, text-generation-with-past
   Design Decisions
   Decision Choice Rationale
   Base class TextDecoderWithPositionIdsOnnxConfig Qwen2-VL is a decoder-based VLM with position_ids support
   Normalized config NormalizedTextConfigWithGQA.with_args() Maps text attributes through text_config.* to handle the composite Qwen2VLConfig
   PKV generator MistralDummyPastKeyValuesGenerator Handles GQA-style key-value heads (same as Llama/Qwen2)
   Position IDs 3D (3, batch_size, seq_len) Qwen2-VL uses M-RoPE with temporal/height/width dimensions
   Vision inputs Only in initial encoding Excluded during cached generation (use_past_in_inputs=True)
   Verification Results
   Test Result
   Import & registration qwen2_vl registered with TasksManager for ONNX
   Config with real Qwen2-VL-2B Normalized config correctly resolves all text/vision attributes
   Dummy input generation Correct shapes: input_ids [2,20], position_ids [3,2,20], pixel_values [32,1176], image_grid_thw [2,3]
   PyTorch forward pass Produces logits [2, 20, 151936]
   ONNX export 2.4 MB graph + 9.7 GB weights exported at opset 18
   ONNX Runtime inference Produces matching logits shape (2, 20, 151936)
   Numerical accuracy Max diff 0.036, mean diff 0.0015 (acceptable for 2B model)
   Known Limitation
   The vision encoder's internal operations (iteration over grid_thw, cu_seqlens computation) use data-dependent shapes that become constants during ONNX tracing. This means the exported model expects the same number of images per inference as used during tracing. A model patcher could address this in a follow-up by making the vision encoder's attention mechanism ONNX-friendly for truly dynamic batch sizes.

[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.