Release 0.18 ·
One block, a fragment, or a whole model
A new compiler prices one block exactly as it prices a model, and a design with one broken block no longer erases the screen.
NEURAX 0.18 replaces the analysis engine with a compiler that lowers every block on the canvas to the primitive operations it is made of, and prices those. The consequence people notice first: one block, dropped on an empty canvas, is already a design with real figures: its operations, its weights, its memory and its step time on your machine.
The assistant was the reason. Asked to fix one block, it used to re-run a whole-design analysis that failed with the same sentence however many times it asked. Now it can ask about the block itself.
A design with one broken block keeps everything else. Every block that compiled keeps its numbers, and every block that did not is named with its own reason. What NEURAX cannot compute is said as unknown, with what is missing, never drawn as a zero.
It scales the other way too. A dense transformer past a thousand billion parameters is priced in under a tenth of a second, because the work is one lowering per drawn block, never one per weight.
The ground truth changed with it: 73 published architectures measured with PyTorch itself: parameters, FLOPs by operator and the bytes autograd keeps. An automated test compiles 56 of them on every change and fails if parameters drift more than 1% or FLOPs more than 5%.