PyTorch models

(Stage 2)

import sf.torch as sft

model = sft.convert(model, dtype=sf.sf16)             # Linear and Conv2d now compute in sf16
fast = torch.compile(model, backend="superfloat")      # sends the SF layers to the device
out = fast(x)
sft.last_status(model)                                 # per-layer saturation report

Weights and activations are scaled so the chip’s -1..1 range is used well, and by default the scale leaves enough headroom that nothing can saturate inside a layer. Training still works (straight-through gradients).

PyTorch (Stage 2)

Kernels (Stage 3)