[{"data":1,"prerenderedAt":38},["ShallowReactive",2],{"chapter:vision-rs\u002Fappendix\u002Ffaq-and-roadmap.json":3},{"project":4,"route":5,"title":6,"titleHtml":7,"navTitle":6,"part":8,"sourcePath":9,"editUrl":10,"html":11,"toc":12,"hasMermaid":33,"prev":34,"next":37},"vision-rs","\u002Fvision-rs\u002Fappendix\u002Ffaq-and-roadmap","FAQ & Roadmap","FAQ & Roadmap","Appendix","appendix\u002Ffaq-and-roadmap.md","https:\u002F\u002Fgithub.com\u002Fteenygrad\u002Fvision-rs\u002Fedit\u002Fmain\u002Fbook\u002Fsrc\u002Fappendix\u002Ffaq-and-roadmap.md","\u003Ch2 id=\"roadmap\">Roadmap\u003C\u002Fh2>\n\u003Cp>vision-rs doesn’t have a published, authoritative roadmap yet. What’s\nimplemented today:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Model support\u003C\u002Fstrong>: YOLO26 (variants N\u002FS\u002FM\u002FL\u002FXL), inference and training\n(single-head and dual-assignment).\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Deployment\u003C\u002Fstrong>: cross-compilation and packaging for Jetson Orin Nano.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Kernels\u003C\u002Fstrong>: Flash Attention 2, PSA attention wrappers, detect-decode,\nCIoU\u002Fclassification loss.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Ccode>DetectorConfig\u003C\u002Fcode> is deliberately structured as an enum over model families\n(see \u003Ca href=\"\u002Fvision-rs\u002Fcore-concepts\u002Fdetection-api\">The Detection API\u003C\u002Fa>) so additional\nmodel families can be added without breaking the top-level\n\u003Ccode>ObjectDetector\u003C\u002Fcode> API — check the\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fteenygrad\u002Fvision-rs\" target=\"_blank\" rel=\"noopener noreferrer\">repository\u003C\u002Fa> for current activity\nif you need to know what’s actively being worked on.\u003C\u002Fp>\n\u003Ch2 id=\"faq\">FAQ\u003C\u002Fh2>\n\u003Ch3 id=\"why-is-cuda-a-default-feature-rather-than-optional\">Why is \u003Ccode>cuda\u003C\u002Fcode> a default feature rather than optional?\u003C\u002Fh3>\n\u003Cp>Because vision-rs’s models are traced through teenygrad’s computational\ngraph and lowered to GPU kernels — there’s currently no CPU execution path\nfor the shipped models. \u003Ccode>--no-default-features\u003C\u002Fcode> still builds and documents\nthe parts of the crate that aren’t behind \u003Ccode>cuda\u003C\u002Fcode>\u002F\u003Ccode>training\u003C\u002Fcode> (mostly type\ndefinitions), which is what powers this crate’s docs.rs page, but it won’t\nget you a runnable detector without a GPU.\u003C\u002Fp>\n\u003Ch3 id=\"why-does-this-depend-on-a-custom-rust-compiler-fork-teenyc\">Why does this depend on a custom Rust compiler fork (\u003Ccode>teenyc\u003C\u002Fcode>)?\u003C\u002Fh3>\n\u003Cp>Only for compiling GPU kernels (see\n\u003Ca href=\"\u002Fvision-rs\u002Fkernels-and-performance\u002Fteenyc-toolchain\">The \u003Ccode>teenyc\u003C\u002Fcode> Toolchain\u003C\u002Fa>) —\nthe crate itself builds with stable \u003Ccode>rustc\u003C\u002Fcode>. The kernel DSL compiles\nthrough an MLIR backend that isn’t part of upstream \u003Ccode>rustc\u003C\u002Fcode>, so a separate\ncompiler handles just that piece, either just-in-time during development or\nahead-of-time when packaging for deployment.\u003C\u002Fp>\n\u003Ch3 id=\"can-i-install-this-from-cratesio-yet\">Can I install this from crates.io yet?\u003C\u002Fh3>\n\u003Cp>Not yet published — vision-rs depends on several \u003Ccode>teeny-*\u003C\u002Fcode> crates from\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fteenygrad\u002Fteenygrad\" target=\"_blank\" rel=\"noopener noreferrer\">teenygrad\u003C\u002Fa> that aren’t live on\ncrates.io yet either. Once they are, \u003Ccode>[patch.crates-io]\u003C\u002Fcode> in \u003Ccode>Cargo.toml\u003C\u002Fcode>\ncomes out and vision-rs publishes normally.\u003C\u002Fp>\n\u003Ch3 id=\"im-interested-in-contributing--how-do-i-get-started\">I’m interested in contributing — how do I get started?\u003C\u002Fh3>\n\u003Cp>See \u003Ca href=\"\u002Fvision-rs\u002Fcontributing\u002Fcontributing\">Contributing to vision-rs\u003C\u002Fa>.\u003C\u002Fp>\n",[13,17,20,24,27,30],{"id":14,"text":15,"level":16},"roadmap","Roadmap",2,{"id":18,"text":19,"level":16},"faq","FAQ",{"id":21,"text":22,"level":23},"why-is-cuda-a-default-feature-rather-than-optional","Why is cuda a default feature rather than optional?",3,{"id":25,"text":26,"level":23},"why-does-this-depend-on-a-custom-rust-compiler-fork-teenyc","Why does this depend on a custom Rust compiler fork (teenyc)?",{"id":28,"text":29,"level":23},"can-i-install-this-from-cratesio-yet","Can I install this from crates.io yet?",{"id":31,"text":32,"level":23},"im-interested-in-contributing--how-do-i-get-started","I’m interested in contributing — how do I get started?",false,{"title":35,"titleHtml":35,"route":36},"Contributing to vision-rs","\u002Fvision-rs\u002Fcontributing\u002Fcontributing",null,1786271830593]