torch:step

(torch:step optimizer)

Applies the optimizer's rule to every parameter and returns the optimizer (PyTorch's optimizer.step()). The step counter is incremented first, so a bias correction reading torch:step-count sees 1 during the first step.

The update writes each parameter's data in place with no torch operation, so it records nothing on the tape: unlike a hand-written update built from torch:set-data, it needs no torch:no-grad around it. A parameter whose gradient is still NIL is skipped.