Changelog

All notable changes to the visynx Python SDK. Versions follow Semantic Versioning.

0.3.1 (2026-11-14)

Documentation and packaging only; no change to the API or its behaviour.

New documentation site at docs.visynx.com, with an installation guide, a quickstart, the API reference and this changelog.
README: an introduction, a features list, an installation section, the NVIDIA Jetson workflow, and links to the docs.
Package metadata: a clearer description and keywords, OS Independent, and links to the docs and changelog.

Run on NVIDIA Jetson: on a Jetson, or with VISYNX_TRT_PRECISION set, a portable model loads through onnxruntime's TensorRT provider. The engine is built on the device and cached per device and JetPack release, in INT8 when the export ships calibration and FP16 otherwise. If TensorRT can't build the graph, it falls back to CUDA with a warning.
model.download(format="jetson_fp16" | "jetson_int8") waits for INT8 calibration on the server and unpacks the bundle into the target folder.
visynx.open_camera("csi://0" | "usb://0" | url) opens Argus (GStreamer), V4L2 or any other OpenCV source.
python -m visynx pair checks a vf_ key and saves it for the device; Visynx() falls back to it. python -m visynx build-engine PATH builds the TensorRT engine ahead of the first frame.
LocalModel.providers shows the onnxruntime execution providers in use, and portable models handle more output shapes.

Run downloaded models offline: visynx.load_model() returns a LocalModel whose predict() gives the same Result as the hosted call.
New local extra (pip install "visynx[local]") and python -m visynx setup.

First release: Visynx(api_key=...), vf.model("project/version"), and predict() with segment=True (SAM 2 polygons) and depth=True.
Typed errors that quote X-Request-ID, and 429s retried on Retry-After and billed once.