with torch.no_grad():
    for p in model.parameters(): p -= p.grad * lr
    model.zero_grad()