v4.49.0

Migration Steps

1. Review the release notes for newly added models: Helium, Qwen2.5-VL, SuperGlue, Granite Vision, Zamba2, GOT-OCR 2.0, DAB-DETR, Depth PRO, and RT-DETRv2.
2. If you were using older versions or related models that have been superseded (e.g., Qwen2-VL in favor of Qwen2.5-VL), update your model loading configurations to use the new model identifiers.
3. If integrating Helium, note its supported languages (English, French, German, Italian, Portuguese, Spanish) and ensure your use case aligns.
4. If migrating to Qwen2.5-VL, be aware of architectural improvements like SwiGLU, RMSNorm in the ViT, and enhanced temporal dynamic resolution handling via upgraded MRoPE.
5. If using SuperGlue, integrate it alongside SuperPoint for feature matching and pose estimation tasks.
6. If migrating to Granite Vision, understand it uses a Granite LM and SigLIP visual encoder, leveraging multiple concatenated vision hidden states.
7. If adopting Zamba2, note its hybrid Mamba/Transformer architecture and that it uses the Mistral v0.1 tokenizer.
8. If using GOT-OCR 2.0, be aware that this implementation outputs plain text, requiring external processing for complex formats like tables or math formulas.
9. If migrating to DAB-DETR, note its use of dynamically updated anchor boxes for improved cross-attention computation in object detection.
10. If adopting Depth PRO, recognize it is a foundation model for zero-shot metric monocular depth estimation using a Dinov2 encoder and DPT-like fusion.
11. If migrating to RT-DETRv2, expect slight mAP improvements over RT-DETR due to selective multi-scale feature extraction and a discrete sampling operator.