sde (torch.nn.Module): Object with methods `f` and `g` representing the
            drift and diffusion. The output of `g` should be a single tensor of
            size (batch_size, d) for diagonal noise SDEs or (batch_size, d, m)
            for SDEs of other noise types; d is the dimensionality of state and
            m is the dimensionality of Brownian motion.
        y0 (Tensor): A tensor for the initial state.
        ts (Tensor or sequence of float): Query times in non-descending order.
            The state at the first time of `ts` should be `y0`.
        bm (Brownian, optional): A 'BrownianInterval', `BrownianPath` or
            `BrownianTree` object. Should return tensors of size (batch_size, m)
            for `__call__`. Defaults to `BrownianInterval`.
        method (str, optional): Numerical integration method to use. Must be
            compatible with the SDE type (Ito/Stratonovich) and the noise type
            (scalar/additive/diagonal/general). Defaults to a sensible choice
            depending on the SDE type and noise type of the supplied SDE.
        adjoint_method (str, optional): Name of numerical integration method for
            backward adjoint solve. Defaults to a sensible choice depending on
            the SDE type and noise type of the supplied SDE.
        dt (float, optional): The constant step size or initial step size for
            adaptive time-stepping.
        adaptive (bool, optional): If `True`, use adaptive time-stepping.
        adjoint_adaptive (bool, optional): If `True`, use adaptive time-stepping
            for the backward adjoint solve.
        rtol (float, optional): Relative tolerance.
        adjoint_rtol (float, optional): Relative tolerance for backward adjoint
            solve.
        atol (float, optional): Absolute tolerance.
        adjoint_atol (float, optional): Absolute tolerance for backward adjoint
            solve.
        dt_min (float, optional): Minimum step size during integration.
        options (dict, optional): Dict of options for the integration method.
        adjoint_options (dict, optional): Dict of options for the integration
            method of the backward adjoint solve.
        adjoint_params (Sequence of Tensors, optional): Tensors whose gradient
            should be obtained with the adjoint. If not specified, defaults to
            the parameters of `sde`.
        names (dict, optional): Dict of method names for drift and diffusion.
            Expected keys are "drift" and "diffusion". Serves so that users can
            use methods with names not in `("f", "g")`, e.g. to use the
            method "foo" for the drift, we supply `names={"drift": "foo"}`.
        logqp (bool, optional): If `True`, also return the log-ratio penalty.
        extra (bool, optional): If `True`, also return the extra hidden state
            used internally in the solver.
        extra_solver_state: (tuple of Tensors, optional): Additional state to
            initialise the solver with. Some solvers keep track of additional
            state besides y0, and this offers a way to optionally initialise
            that state.

    Returns:
        A single state tensor of size (T, batch_size, d).
        if logqp is True, then the log-ratio penalty is also returned.
        If extra is True, the any extra internal state of the solver is also
        returned.

    Raises:
        ValueError: An error occurred due to unrecognized noise type/method,
            or `sde` is missing required methods.

    Note:
        The backward pass is much more efficient with Stratonovich SDEs than
        with Ito SDEs.

    Note:
        Double-backward is supported for Stratonovich SDEs. Doing so will use
        the adjoint method to compute the gradient of the adjoint. (i.e. rather
        than backpropagating through the numerical solver used for the
        adjoint.) The same `adjoint_method`, `adjoint_adaptive`, `adjoint_rtol,
        `adjoint_atol`, `adjoint_options` will be used for the second-order
        adjoint as is used for the first-order adjoin