Crate rlx_dinov2 

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.

Re-exports

pub use builder::build_dinov2_graph_sized;

pub use config::DinoV2Config;

pub use config::IMAGENET_MEAN;

pub use config::IMAGENET_STD;

pub use flow::DinoV2Built;

pub use flow::DinoV2Flow;

pub use flow::build_dinov2_built;

pub use flow::build_dinov2_built_with_packed;

pub use packed_gguf::gguf_has_packed_linears;

pub use packed_gguf::load_dinov2_from_gguf;

pub use preprocess::DinoV2PreprocessWeights;

pub use preprocess::assemble_hidden;

pub use preprocess::rgb_u8_to_imagenet_nchw;

pub use runner::DinoV2Output;

pub use runner::DinoV2Runner;

pub use runner::DinoV2RunnerBuilder;

pub use runner::DinoV2Variant;

Modules

builder

DINOv2 graph builder — delegates to super::flow::DinoV2Flow.

cli

config

DINOv2 configuration. Mirrors Meta’s reference configs.

flow

Tier-0 DINOv2 flow — native ModelFlow ViT assembly.

packed_gguf

K-quant / legacy GGUF ViT weights for DINOv2 (DequantMatMul via rlx_flow::GgufPackedParams).

preprocess

Host-side DINOv2 preprocessing: patch projection + token assembly.

runner