When I tanslate the following python code to cpp code with libtorch api

class _SphericalHarmonics(torch.autograd.Function):
    """Spherical Harmonics"""

    @staticmethod
    def forward(
        ctx, sh_degree: int, dirs: Tensor, coeffs: Tensor, masks: Tensor
    ) -> Tensor:
        colors = _make_lazy_cuda_func("compute_sh_fwd")(sh_degree, dirs, coeffs, masks)
        ctx.save_for_backward(dirs, coeffs, masks)
        ctx.sh_degree = sh_degree
        ctx.num_bases = coeffs.shape[-2]
        return colors

    @staticmethod
    def backward(ctx, v_colors: Tensor):
        dirs, coeffs, masks = ctx.saved_tensors
        sh_degree = ctx.sh_degree
        num_bases = ctx.num_bases
        compute_v_dirs = ctx.needs_input_grad[1]
        v_coeffs, v_dirs = _make_lazy_cuda_func("compute_sh_bwd")(
            num_bases,
            sh_degree,
            dirs,
            coeffs,
            masks,
            v_colors.contiguous(),
            compute_v_dirs,
        )
        if not compute_v_dirs:
            v_dirs = None
        return None, v_dirs, v_coeffs, None

I can’t find the AutogradContext *ctx has the method or attribute ctx->needs_input_grad corressponding to the ctx.needs_input_grad function in pytorch. Have some one know it?

I have try v_dirs.required_grad() to sub it in cpp with libtorch, I don’t know this is right.