It requires implementing training yourself to have something like Ultralytics one.
I used PP-YOLOE for my latest model and PaddleDetection. It's Apache-2.0 licensed, so we can copy from it, possibly, without lock in into AGPL.
Though this might be a separate project from tinygrad itself, was thinking about. But there is also no way as far as I understand exporting it to ONNX to run on smaller accelerators without python, like RKNN.