Introduction

vision-rs is a high-performance computer vision SDK for Rust, built on teenygrad, a memory-safe Rust ML training and inference runtime. It currently ships YOLO26, a real-time object detection model family, with custom GPU kernels (attention, detection decode, loss functions) written directly against teenygrad’s Triton-style kernel DSL.

What this book covers

• Getting a working detector running against a JPEG/PNG image.
• The ObjectDetector/DetectorConfig API surface.
• How the YOLO26 model is assembled from backbone/neck/head blocks, and how the N/S/M/L/XL variants differ.
• Training: the TaskAlignedAssigner, CIoU/classification losses, and the dual-head (one2many/one2one) assignment scheme.
• The custom kernels vision-rs ships (Flash Attention 2, position-sensitive attention, detect-decode) and the teenyc toolchain that compiles them.
• Cross-compiling and packaging a self-contained deployable bundle for a Jetson Orin Nano.

What this book does not cover

This book is a guide to using vision-rs. For the full API reference (every public type, field, and function), see the crate documentation. For the ML runtime vision-rs is built on — tensors, the computational graph, the kernel compiler — see The Teenygrad Book.

Status

vision-rs is under active development. The public API (vision_rs::detect) is the intended stable surface; everything under vision_rs::models is expected to grow additional model families over time. See the FAQ & Roadmap for what’s implemented today.