• https://arxiv.org/abs/2411.00666

with torch.no_grad():
    mf = 0.3
    lr = 0.7
    for param_old, param, m in zip(model_old.parameters(), model.parameters(), momentum):
        grad = param - param_old
        m.copy_(mf * m + grad)
        param.copy_(param_old + lr * (m + mf * grad))
        param_old.copy_(param)