Release 0.16 ·
NEURAX runs the designs it costs
Every design passes one gate before it trains: the generated code must match the analysis that approved it.
Until 0.16, NEURAX told you what a design would cost. Now it runs it. The design becomes real PyTorch files in your project folder (model, training script, request) that you can read, edit and keep.
Every model passes one gate before it is trained: the generated code is built, counted and pushed one batch through, and code whose parameter count disagrees with the analysis is refused rather than trained.
The hardware probe now reports what the machine can run, not just what it contains, and you choose where the data is and where the run happens. The Production workspace was removed; Decision is the workspace for the person who signs.