Update README.md

README.md

@@ -20,4 +20,38 @@ Meta AI (FAIR)
 <ul>
  	<li><a href="https://github.com/facebookresearch/dinov2">Learn more</a></li>
  	<li><a href="https://github.com/facebookresearch/dinov2/blob/main/LICENSE">License</a></li>
-</ul>
+</ul>
+
+
+# Vendored DINOv2 source
+
+This repo bundles a slimmed copy of the DINOv2 architecture source alongside the model weights, so AddaxAI can load the architecture from this Hugging Face repo instead of fetching it from `github.com/facebookresearch/dinov2` at runtime.
+
+This is needed because many AddaxAI users sit behind networks that allow `huggingface.co` but block or throttle GitHub, and the first run of the embedding step would otherwise fail with no internet access to github.com.
+
+## What is included
+
+Only the files needed to load the vits14, vitb14, and vitl14 backbones for inference:
+
+- `hubconf.py` (slimmed to the three backbone entry points)
+- `LICENSE` (Apache 2.0)
+- `dinov2/__init__.py`
+- `dinov2/hub/__init__.py`, `dinov2/hub/backbones.py`, `dinov2/hub/utils.py`
+- `dinov2/models/__init__.py`, `dinov2/models/vision_transformer.py`
+- `dinov2/layers/__init__.py` and the 7 layer files it imports
+
+Training, evaluation, segmentation, depth, classifiers, cell-dino, xray-dino, and dinotxt code paths are not included.
+
+## How it is loaded
+
+AddaxAI calls:
+
+```python
+torch.hub.load(model_dir, model_arch, source="local", pretrained=False)
+```
+
+where `model_dir` is the local directory the weights and these source files have been downloaded into via the Hugging Face Hub.
+
+## License and provenance
+
+DINOv2 is licensed under Apache 2.0. The original code lives at https://github.com/facebookresearch/dinov2. See the `LICENSE` file for the full text. No source-level modifications have been made to the vendored files except for `hubconf.py`, which was slimmed to only re-export the three entry points AddaxAI uses.