Taggerine (DINOv3 ViT-H/16+) image tagger on the AMD NPU
An IRON export of lodestones/taggerine for AMD Ryzen AI NPUs: the compiled NPU kernels (.xclbin + instruction streams) and the packed weights a zero-Python Rust runtime replays. It produces the tagger's 74,625 raw logits; the tag vocabulary (tagger_vocab_with_categories_and_alias_updated.json) comes from the upstream repo.
The kernels are compiled for NPU2 (AIE2P: Strix Point, Strix Halo, Krackan) and will not load on NPU1 (Phoenix, Hawk Point).
The bundle is IRON's shared bundle format, version 9: manifest.txt (model params, one xclbin record per hardware context, kernel and matrix records), the tensor store tensors.txt + tensors.bin, and kernels/. iron/applications/taggerine/export_taggerine.py documents every record and tensor.
Download
hf download brishen/iron-taggerine-npu2 --local-dir taggerine
Usage
See iron/applications/taggerine/README.md in IRON for how the model runs on the NPU. The runtime that reads this bundle is image-organizer's tagger (iron backend).
Provenance
- Upstream model:
lodestones/taggerine(license:apache-2.0; its terms apply to these weights) - IRON commit:
65b32e2 - Uploaded: 2026-09-23
- Files: 15, 1.9 GB
- Reference: NPU vs float32 logits cosine 0.99965, top-50 overlap 47/50 (the
ref.*tensors)