How to use the d2l.split_batch function in d2l

def train_batch_ch12(net, features, labels, loss, trainer, ctx_list):
    Xs, ys = d2l.split_batch(features, labels, ctx_list)
    with autograd.record():
        pys = [net(X) for X in Xs]
        ls = [loss(py, y) for py, y in zip(pys, ys)]
    for l in ls:
        l.backward()
    trainer.step(features.shape[0])
    train_loss_sum = sum([l.sum().asscalar() for l in ls])
    train_acc_sum = sum(d2l.accuracy(py, y) for py, y in zip(pys, ys))
    return train_loss_sum, train_acc_sum