def xp_pad(x, pad_width, mode='constant', *, xp, **kwargs):
    # xp.pad is available on numpy, cupy and jax.numpy; on torch, reuse
    # http://github.com/pytorch/pytorch/blob/main/torch/_numpy/_funcs_impl.py#L2045
    # for mode = 'constant'
    if mode != 'constant':
        raise NotImplementedError()

    value = kwargs.get("constant_values", 0)

    if is_array_api_strict(xp):
        np_x = np.asarray(x)
        padded = np.pad(np_x, pad_width, mode=mode, **kwargs)
        return xp.asarray(padded)
    elif is_torch(xp):
        pad_width = xp.asarray(pad_width)
        pad_width = xp.broadcast_to(pad_width, (x.ndim, 2))
        pad_width = xp.flip(pad_width, axis=(0,)).flatten()
        return xp.nn.functional.pad(x, tuple(pad_width), value=value)
    else:
        return xp.pad(x, pad_width, mode=mode, **kwargs)