SimCubeLtd/dinov2-vits14-candle

DINOv2 ViT-S/14 — candle format

dinov2-vits14.safetensors — DINOv2 ViT-S/14 weights packed into the layout expected by the candle framework's dinov2::vit_small model. Drop-in for on-device image feature extraction in Rust.

These are the numerically-identical parameters of facebook/dinov2-small plus Meta's official ImageNet linear-probe head, re-packed for candle (the separate query/key/value attention projections are fused into a single qkv tensor, and tensor keys are renamed to candle's convention). No weights were retrained or altered in value.

Why this exists

candle's vit_small expects a fused qkv and candle-style key names, which the upstream HF checkpoint doesn't use. Existing candle-format re-hosts are licensed CC-BY-NC or carry no license, which is unusable commercially. This repo provides the same weights under Apache-2.0, derived entirely from Apache-2.0 sources.

Contents

Tensors	Source (Apache-2.0)
Backbone (patch embed, 12 transformer blocks, final norm; fused qkv)	facebook/dinov2-small
head.weight (1000×768), head.bias (1000) — ImageNet linear probe	dinov2_vits14_linear_head.pth
mask_token (zeros, training-only)	facebook/dinov2-small

candle::dinov2::vit_small's forward pass returns the 1000-d linear-probe output; used here as a dense image descriptor for visual-similarity ranking.

Usage (candle)

use candle_transformers::models::dinov2;
let vb = unsafe {
    candle_nn::VarBuilder::from_mmaped_safetensors(
        &["dinov2-vits14.safetensors"], candle_core::DType::F32, &device)?
};
let model = dinov2::vit_small(vb)?;
let descriptor = model.forward(&image_tensor)?; // (1, 1000)

License & attribution

Apache-2.0 (see LICENSE). Derived from DINOv2 © Meta Platforms, Inc.; see NOTICE. Original work: facebookresearch/dinov2 (Oquab et al., 2023).