DINOv2 — Meta's self-supervised ViT (with optional registers).
Public entry points:
- [
DinoV2Config] — model dimensions + variant factories (vit_small,vit_base,vit_large) - [
build_dinov2_graph_sized] — emits the IR graph and the host-side [DinoV2PreprocessWeights]. - [
assemble_hidden] / [rgb_u8_to_imagenet_nchw] — host-side image → encoder-input plumbing.
Weight keys match Meta / candle's safetensors so checkpoints from
the HF Hub (e.g. lmz/candle-dino-v2) load with no remapping.