iron-sam3-npu2

SAM3 segmentation on the AMD NPU

An IRON export of facebook/sam3 for AMD Ryzen AI NPUs: the compiled NPU kernels (.xclbin + instruction streams) and the packed weights the taconite-sam3 Rust runtime replays, over XRT or directly over the amdxdna driver.

The kernels are compiled for NPU2 (AIE2P: Strix Point, Strix Halo, Krackan) and will not load on NPU1 (Phoenix, Hawk Point).

Download

hf download brishen/iron-sam3-npu2 --local-dir sam3
# or, from an IRON checkout:
python scripts/hf_models.py download sam3 --repo brishen/iron-sam3-npu2 --out sam3

This is bundle format 2 (the ViT's residual stream in fp32, and the more accurate backbone numerics of IRON's iron/applications/sam3/README.md): it needs taconite-sam3 >= 0.1.1. For 0.1.0, download the format-1 bundle instead: hf download brishen/iron-sam3-npu2 --revision bundle-v1 --local-dir sam3. A higher-precision export (export_sam3.py --precision high: the ViT's GEMM weights as bfp16 hi + lo; backbone cosine 0.9923 -> 0.9943 against float32, ~20% slower, 2.1 GB) is on the high branch, for the same runtime: hf download brishen/iron-sam3-npu2 --revision high --local-dir sam3.

Usage

cargo install taconite-sam3             # over XRT
# or, with no XRT at all (straight to the amdxdna driver):
cargo install taconite-sam3 --no-default-features --features cli,direct
sam3 segment sam3 photo.jpg "person" -o overlay.png
sam3 check sam3                        # every stage against the float32 reference

See taconite-sam3 (docs, source) for the full API. The bundle is exported by IRON's iron/applications/sam3.

Provenance

• Upstream model: facebook/sam3 (license: other; its terms apply to these weights)
• IRON commit: 6dc8c1f-dirty
• Uploaded: 2026-10-01
• Files: 46, 1.6 GB