yolo.cpp — ggml port of YOLOv8

C++ forward pass for YOLOv8 detection models built on ggml. Letterbox preprocessing, anchor-free output decode, and NMS stay in TypeScript (see plugins/plugin-vision/src/yolo-detector.ts); this library runs only the CNN.

Status

Phase 2 (current): FFI surface scaffolded; conversion script authored. Native lib + GGUF weights not yet built. The TS binding throws a clear error until both are present.

Build (when implemented)

bash

cd plugins/plugin-vision/native/yolo.cpp
cmake -B build -S . -DYOLO_WITH_METAL=ON   # macOS arm64
cmake --build build --config Release

Convert weights (when implemented)

bash

python scripts/convert.py --variant yolov8n --out vision/yolov8n.gguf

License

The runtime in this directory is a clean-room implementation. Ultralytics YOLOv8 weights are AGPL-3.0; this repo does not bundle them. End users fetch weights at runtime or via the model-publish workflow that already exists for the llama.cpp text/vision tiers.