Three things ship, on three clocks.
The engine, the models it can run, and the pack format are separate releases, because they change for separate reasons. Mixing them into one feed would hide which of the three moved — and that is usually the only thing you need to know.
Engine · NVIDIA CUDA
PREPARINGThe second lane. Running internally on image and video families; nothing published until it clears the same gate.
Engine · Apple Metal
OPENThe runtime itself: version, supported devices, kernel and load-path changes, compatibility, fixes, checksums, known limits.
Packs
PLANNEDThe calibrated artifact format that convert produces, and the compatibility rules between a pack and an engine.
What has shipped so far
Titles and dates only. A release is written once, on the page that owns it: the lane page for the engine, the family page for a model.
Loading…