vision-rs / Appendix

FAQ & Roadmap

Roadmap

vision-rs doesn’t have a published, authoritative roadmap yet. What’s implemented today:

  • Model support: YOLO26 (variants N/S/M/L/XL), inference and training (single-head and dual-assignment).
  • Deployment: cross-compilation and packaging for Jetson Orin Nano.
  • Kernels: Flash Attention 2, PSA attention wrappers, detect-decode, CIoU/classification loss.

DetectorConfig is deliberately structured as an enum over model families (see The Detection API) so additional model families can be added without breaking the top-level ObjectDetector API — check the repository for current activity if you need to know what’s actively being worked on.

FAQ

Why is cuda a default feature rather than optional?

Because vision-rs’s models are traced through teenygrad’s computational graph and lowered to GPU kernels — there’s currently no CPU execution path for the shipped models. --no-default-features still builds and documents the parts of the crate that aren’t behind cuda/training (mostly type definitions), which is what powers this crate’s docs.rs page, but it won’t get you a runnable detector without a GPU.

Why does this depend on a custom Rust compiler fork (teenyc)?

Only for compiling GPU kernels (see The teenyc Toolchain) — the crate itself builds with stable rustc. The kernel DSL compiles through an MLIR backend that isn’t part of upstream rustc, so a separate compiler handles just that piece, either just-in-time during development or ahead-of-time when packaging for deployment.

Can I install this from crates.io yet?

Not yet published — vision-rs depends on several teeny-* crates from teenygrad that aren’t live on crates.io yet either. Once they are, [patch.crates-io] in Cargo.toml comes out and vision-rs publishes normally.

I’m interested in contributing — how do I get started?

See Contributing to vision-rs.