robbyant

/

lingbot-vision-vit-giant

like 13

Follow

Robbyant 503

This PR adds native Transformers support alongside the existing reference checkpoint:

• model.safetensors + config.json — DINOv3ViT-compatible weights converted from model.pt (LingBot-Vision backbones are architecturally identical to the DINOv3 ViT implementation in transformers >= 4.56), loadable via AutoModel.from_pretrained(...) with no custom code
• preprocessor_config.json — 512x512 bilinear resize + ImageNet normalization, matching the reference preprocessing
• README.md — library_name: transformers, a Transformers usage snippet, dual-format wording (also fixes the stale lbot_vision_infer import in the existing snippet)

model.pt is unchanged and the reference lingbot_vision loader is unaffected.

Verification: converted weights are numerically equivalent to model.pt under the reference implementation (fp32, CPU): max |diff| = 2e-5 on patch/CLS tokens, including non-square inputs (dynamic RoPE). Load path tested on transformers 4.56.0 and 5.5.3. Conversion script: scripts/convert_to_transformers.py in the GitHub repo.