# - backward_api : gumbel_softmax_grad
#   forward : gumbel_softmax (Tensor x, float temperature, bool hard, int axis) -> Tensor(out)
#   args : (Tensor out, Tensor out_grad, int axis)
#   output : Tensor(x_grad)
#   infer_meta :
#     func : GumbelSoftmaxGradInferMeta
#     param : [out, out_grad, axis]
#   kernel :
#     func : gumbel_softmax_grad

- backward_api : abs_grad
  forward : abs (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : abs_grad
  data_transform:
    skip_transform : out_grad

- backward_api : acos_grad
  forward : acos (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : acos_grad

- backward_api : acosh_grad
  forward : acosh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : acosh_grad

- backward_api : add_grad
  forward : add (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : add_grad
  no_need_buffer : x, y

- backward_api : add_n_grad
  forward : add_n (Tensor[] x) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad)
  output : Tensor[](x_grad)
  invoke : add_n_grad_impl(x, out_grad)
  no_need_buffer : x

- backward_api : addmm_grad
  forward : addmm (Tensor input, Tensor x, Tensor y, float alpha, float beta) -> Tensor(out)
  args : (Tensor input, Tensor x, Tensor y, Tensor out_grad, float alpha, float beta)
  output : Tensor(input_grad), Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [input, x, y]
  kernel :
    func : addmm_grad

- backward_api : argsort_grad
  forward : argsort (Tensor x, int axis, bool descending) -> Tensor(out), Tensor(indices)
  args : (Tensor indices, Tensor x, Tensor out_grad, int axis, bool descending)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : argsort_grad
  no_need_buffer : x

- backward_api : asin_grad
  forward : asin (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : asin_grad

- backward_api : asinh_grad
  forward : asinh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : asinh_grad

- backward_api : atan2_grad
  forward : atan2 (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : atan2_grad

- backward_api : atan_grad
  forward : atan (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : atan_grad

- backward_api : atanh_grad
  forward : atanh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : atanh_grad

- backward_api : batch_norm_grad
  forward : batch_norm (Tensor x, Tensor scale, Tensor bias, Tensor mean, Tensor variance, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu) -> Tensor(out), Tensor(mean_out), Tensor(variance_out), Tensor(saved_mean), Tensor(saved_variance), Tensor(reserve_space)
  args : (Tensor x, Tensor scale, Tensor bias, Tensor mean_out, Tensor variance_out, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor out_grad, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu)
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, scale, bias]
  kernel :
    func : batch_norm_grad
    data_type : out_grad
  optional : mean_out, variance_out, reserve_space

- backward_api : bce_loss_grad
  forward : bce_loss (Tensor input, Tensor label) -> Tensor(out)
  args : (Tensor input, Tensor label, Tensor out_grad)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : bce_loss_grad

- backward_api : brelu_grad
  forward : brelu (Tensor x, float t_min, float t_max) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float t_min, float t_max)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : brelu_grad

- backward_api : cast_grad
  forward : cast (Tensor x, DataType out_dtype) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : cast_grad
    data_type : out_grad

- backward_api : ceil_grad
  forward : ceil(Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out_grad]
  kernel :
    func : ceil_grad

- backward_api : cholesky_grad
  forward : cholesky (Tensor x, bool upper) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, bool upper)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : cholesky_grad

- backward_api : cholesky_solve_grad
  forward : cholesky_solve (Tensor x, Tensor y, bool upper) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, bool upper)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : cholesky_solve_grad

- backward_api : clip_grad
  forward : clip (Tensor x, Scalar min, Scalar max) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, Scalar min = 0., Scalar max = 0.)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : clip_grad

- backward_api : concat_grad
  forward : concat (Tensor[] x, Scalar axis) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad, Scalar axis = 0)
  output : Tensor[](x_grad)
  invoke : concat_grad_impl(x, out_grad, axis)

- backward_api : conv2d_grad
  forward : conv2d (Tensor input, Tensor filter, int[] strides, int[] paddings, str paddding_algorithm, int groups, int[] dilations, str data_format, bool use_addto, int workspace_size_MB, bool exhaustive_search) -> Tensor(out)
  args : (Tensor input, Tensor filter, Tensor out_grad,  int[] strides, int[] paddings, str paddding_algorithm, int groups, int[] dilations, str data_format, bool use_addto, int workspace_size_MB, bool exhaustive_search)
  output : Tensor(input_grad), Tensor(filter_grad)
  invoke : conv2d_grad_impl(input, filter, out_grad,  strides, paddings, paddding_algorithm, groups, dilations, data_format, use_addto, workspace_size_MB, exhaustive_search)

- backward_api : conv2d_transpose_grad
  forward : conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
    func : ConvTransposeGradInferMeta
  kernel :
    func : conv2d_transpose_grad

- backward_api : conv3d_transpose_grad
  forward : conv3d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
    func : ConvTransposeGradInferMeta
  kernel :
    func : conv3d_transpose_grad

- backward_api : cos_grad
  forward : cos (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : cos_grad

- backward_api : cosh_grad
  forward : cosh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : cosh_grad

- backward_api : cross_entropy_with_softmax_grad
  forward : cross_entropy_with_softmax (Tensor input, Tensor label, bool soft_label, bool use_softmax, bool numeric_stable_mode, int ignore_index, int axis) -> Tensor(softmax), Tensor(loss)
  args : (Tensor label, Tensor softmax, Tensor loss_grad, bool soft_label, bool use_softmax, bool numeric_stable_mode, int ignore_index, int axis)
  output : Tensor(input_grad)
  infer_meta :
    func : CrossEntropyWithSoftmaxGradInferMeta
  kernel :
    func : cross_entropy_with_softmax_grad
    data_type : softmax

- backward_api : cross_grad
  forward : cross (Tensor x, Tensor y, int axis = 9) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : cross_grad

- backward_api : cumprod_grad
  forward : cumprod (Tensor x, int dim) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int dim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : cumprod_grad

- backward_api : cumsum_grad
  forward : cumsum(Tensor x, int axis, bool flatten, bool exclusive, bool reverse) -> Tensor(out)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  args : (Tensor out_grad, int axis, bool flatten, bool exclusive, bool reverse)
  output : Tensor(x_grad)
  invoke : cumsum(out_grad, axis, flatten, exclusive, !reverse)

- backward_api : depthwise_conv2d_transpose_grad
  forward : depthwise_conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
    func : ConvTransposeGradInferMeta
  kernel :
    func : depthwise_conv2d_transpose_grad

- backward_api : det_grad
  forward : det (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : determinant_grad

- backward_api : diagonal_grad
  forward : diagonal (Tensor x, int offset, int axis1, int axis2) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int offset = 0, int axis1 = 0, int axis2 = 1)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : diagonal_grad
  no_need_buffer : x

- backward_api : digamma_grad
  forward : digamma (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : digamma_grad

- backward_api : dist_grad
  forward : dist (Tensor x, Tensor y, float p) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, float p)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : dist_grad

- backward_api : divide_grad
  forward : divide (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : divide_grad

- backward_api : dropout_grad
  forward : dropout (Tensor x, Tensor seed_tensor, float p, bool is_test, str mode, int seed, bool fix_seed) -> Tensor(out), Tensor(mask)
  args : (Tensor mask, Tensor out_grad, float p, bool is_test, str mode)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : dropout_grad
  optional : seed_tensor

- backward_api : eigh_grad
  forward : eigh (Tensor x, str uplo) -> Tensor(out_w), Tensor(out_v)
  args : (Tensor out_w, Tensor out_v, Tensor out_w_grad, Tensor out_v_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_v]
  kernel :
    func : eigh_grad

- backward_api : elementwise_pow_grad
  forward : elementwise_pow(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : elementwise_pow_grad

- backward_api : elu_grad
  forward : elu (Tensor x, float alpha) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, float alpha)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : elu_grad

- backward_api : erf_grad
  forward : erf (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : erf_grad
    data_type : out_grad

- backward_api : erfinv_grad
  forward : erfinv (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : erfinv_grad

- backward_api : exp_grad
  forward : exp (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : exp_grad

- backward_api : expand_as_grad
  forward : expand_as (Tensor x, Tensor y, int[] target_shape) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] target_shape)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : expand_as_grad

- backward_api : expm1_grad
  forward : expm1 (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : expm1_grad

- backward_api : flatten_grad
  forward : flatten(Tensor x, int start_axis, int stop_axis) -> Tensor(out), Tensor(xshape)
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func :  KernelWithXShapeInferMeta
    param : [xshape]
  kernel :
    func : flatten_grad
    data_type: out_grad
    backend: out_grad
    layout: out_grad

- backward_api : floor_grad
  forward : floor(Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out_grad]
  kernel :
    func : floor_grad

- backward_api : fmax_grad
  forward : fmax(Tensor x, Tensor y, int axis) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmax_grad

- backward_api : fmin_grad
  forward : fmin(Tensor x, Tensor y, int axis) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmin_grad

- backward_api : frobenius_norm_grad
  forward : frobenius_norm(Tensor x, int64_t[] axis,  bool keep_dim,  bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis,  bool keep_dim,  bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : frobenius_norm_grad

- backward_api : gather_grad
  forward : gather(Tensor x, Tensor index, Scalar axis=0) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad, Scalar axis=0, bool overwrite=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    data_type: x
    func : gather_grad

- backward_api : gather_nd_grad
  forward : gather_nd (Tensor x, Tensor index) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : gather_nd_grad

- backward_api : gelu_grad
  forward : gelu(Tensor x,  bool approximate) -> Tensor(out)
  args : (Tensor x, Tensor out_grad,  bool approximate)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : gelu_grad

- backward_api : graph_send_recv_grad
  forward : graph_send_recv (Tensor x, Tensor src_index, Tensor dst_index, str pool_type = "SUM", int64_t out_size = 0) -> Tensor(out), Tensor(dst_count)
  args : (Tensor x, Tensor src_index, Tensor dst_index, Tensor out, Tensor dst_count, Tensor out_grad, str pool_type = "SUM")
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : graph_send_recv_grad
  optional: out, dst_count

- backward_api : hard_shrink_grad
  forward : hard_shrink (Tensor x, float threshold) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : hard_shrink_grad

- backward_api : hard_sigmoid_grad
  forward : hard_sigmoid (Tensor x, float slope, float offset) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, float slope, float offset)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : hard_sigmoid_grad

- backward_api : hard_swish_grad
  forward : hard_swish (Tensor x, float threshold = 6.0, float scale = 6.0, float offset = 3.0) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold, float scale, float offset)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : hard_swish_grad

- backward_api : huber_loss_grad
  forward : huber_loss (Tensor input, Tensor label, float delta) -> Tensor(out), Tensor(residual)
  args : (Tensor residual, Tensor out_grad, float delta)
  output : Tensor(input_grad), Tensor(label_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [residual, residual]
  kernel :
    func : huber_loss_grad

- backward_api : imag_grad
  forward : imag (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  invoke : imag_grad_impl(out_grad)

- backward_api : index_sample_grad
  forward : index_sample (Tensor x, Tensor index) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : index_sample_grad
    data_type : out_grad
  no_need_buffer : x

- backward_api : index_select_grad
  forward : index_select(Tensor x, Tensor index,  int dim) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad,  int dim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : index_select_grad
    data_type : x

- backward_api : kldiv_loss_grad
  forward : kldiv_loss(Tensor x, Tensor label, str reduction) -> Tensor(out)
  args : (Tensor x, Tensor label, Tensor out_grad, str reduction)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : kldiv_loss_grad

- backward_api : kron_grad
  forward : kron (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : kron_grad
    data_type : out_grad

- backward_api : kthvalue_grad
  forward : kthvalue(Tensor x, int k, int axis, bool keepdim) -> Tensor(out), Tensor(indices)
  args : (Tensor x, Tensor indices, Tensor out_grad, int k, int axis, bool keepdim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : kthvalue_grad

- backward_api : label_smooth_grad
  forward : label_smooth (Tensor label, Tensor prior_dist, float epsilon) -> Tensor(out)
  args : (Tensor out_grad, float epsilon)
  output : Tensor(label_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : label_smooth_grad
  optional : prior_dist

- backward_api : leaky_relu_grad
  forward : leaky_relu (Tensor x, float alpha) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float alpha)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : leaky_relu_grad

- backward_api : lerp_grad
  forward : lerp (Tensor x, Tensor y, Tensor weight) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor weight, Tensor out, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : lerp_grad

- backward_api : lgamma_grad
  forward : lgamma(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : lgamma_grad

- backward_api : log10_grad
  forward : log10 (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : log10_grad

- backward_api : log1p_grad
  forward : log1p (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : log1p_grad

- backward_api : log2_grad
  forward : log2 (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : log2_grad

- backward_api : log_grad
  forward : log (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : log_grad

- backward_api : log_loss_grad
  forward : log_loss (Tensor input, Tensor label, float epsilon) -> Tensor(out)
  args : (Tensor input, Tensor label, Tensor out_grad, float epsilon)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : log_loss_grad

- backward_api : log_softmax_grad
  forward : log_softmax(Tensor x,  int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad,  int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out]
  kernel :
    func : log_softmax_grad

- backward_api : logit_grad
  forward : logit (Tensor x, float eps = 1e-6f) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float eps)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : logit_grad

- backward_api : logsigmoid_grad
  forward : logsigmoid (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : logsigmoid_grad

- backward_api : logsumexp_grad
  forward : logsumexp(Tensor x, int64_t[] axis,  bool keepdim,  bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis,  bool keepdim,  bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : logsumexp_grad

- backward_api : masked_select_grad
  forward : masked_select (Tensor x, Tensor mask) -> Tensor(out)
  args : (Tensor x, Tensor mask, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : masked_select_grad
    data_type : x

- backward_api : matmul_double_grad
  forward : matmul_grad (Tensor x, Tensor y, Tensor grad_out, bool transpose_x=false, bool transpose_y=false) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor x, Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, y, grad_out]
  kernel :
    func : matmul_double_grad
  backward : matmul_triple_grad
  optional : grad_x_grad, grad_y_grad

- backward_api : matmul_grad
  forward : matmul (Tensor x, Tensor y, bool transpose_x=false, bool transpose_y=false) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : matmul_grad
  backward : matmul_double_grad

- backward_api : matmul_triple_grad
  forward : matmul_double_grad (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, bool transpose_x=false, bool transpose_y=false) -> Tensor(grad_x), Tensor(grad_y), Tensor(grad_grad_out)
  args : (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, Tensor grad_x_grad, Tensor grad_y_grad, Tensor grad_grad_out_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(fwd_grad_out_grad), Tensor(fwd_grad_grad_x_grad), Tensor(fwd_grad_grad_y_grad)
  infer_meta :
    func : GeneralQuinaryGradInferMeta
    param : [x, y, fwd_grad_out, fwd_grad_grad_x, fwd_grad_grad_y]
  kernel :
    func : matmul_triple_grad
  optional : grad_x_grad, grad_y_grad, grad_grad_out_grad

- backward_api : matrix_power_grad
  forward : matrix_power (Tensor x, int n) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int n)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : matrix_power_grad

- backward_api : max_grad
  forward: max (Tensor x,  int64_t[] dims={},  bool keep_dim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] dims={},  bool keep_dim=false, bool reduce_all=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : max_grad

- backward_api : max_pool2d_with_index_grad
  forward : max_pool2d_with_index(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive) -> Tensor(out), Tensor(mask)
  args : (Tensor x, Tensor mask, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive)
  output : Tensor(x_grad)
  infer_meta :
    func : MaxPoolWithIndexGradInferMeta
  kernel :
    func : max_pool2d_with_index_grad

- backward_api : max_pool3d_with_index_grad
  forward : max_pool3d_with_index(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive) -> Tensor(out), Tensor(mask)
  args : (Tensor x, Tensor mask, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive)
  output : Tensor(x_grad)
  infer_meta :
    func : MaxPoolWithIndexGradInferMeta
  kernel :
    func : max_pool3d_with_index_grad

- backward_api : maximum_grad
  forward : maximum(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : maximum_grad

- backward_api : mean_all_grad
  forward : mean_all(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mean_all_grad

- backward_api : mean_grad
  forward: mean (Tensor x,  int64_t[] dims={},  bool keep_dim=false) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int64_t[] dims={},  bool keep_dim=false, bool reduce_all=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mean_grad

- backward_api : meshgrid_grad
  forward : meshgrid (Tensor[] inputs) -> Tensor[](outputs)
  args : (Tensor[] inputs, Tensor[] outputs_grad)
  output : Tensor[](inputs_grad)
  invoke : meshgrid_grad_impl(inputs, outputs_grad)

- backward_api : min_grad
  forward: min (Tensor x,  int64_t[] dims={},  bool keep_dim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] dims={},  bool keep_dim=false, bool reduce_all=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : min_grad

- backward_api : minimum_grad
  forward : minimum(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : minimum_grad

- backward_api : mish_grad
  forward : mish (Tensor x, float threshold) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : mish_grad

- backward_api : mode_grad
  forward : mode(Tensor x,  int axis,  bool keepdim) -> Tensor(out), Tensor(indices)
  args : (Tensor x, Tensor indices, Tensor out_grad,  int axis,  bool keepdim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mode_grad

- backward_api : modulo_grad
  forward : modulo (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : modulo_grad
  no_need_buffer : x, y

- backward_api : multiply_grad
  forward : multiply (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : multiply_grad

- backward_api : mv_grad
  forward : mv (Tensor x, Tensor vec) -> Tensor(out)
  args : (Tensor x, Tensor vec, Tensor out_grad)
  output : Tensor(x_grad), Tensor(vec_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, vec]
  kernel :
    func : mv_grad

- backward_api : nll_loss_grad
  forward : nll_loss (Tensor input, Tensor label, Tensor weight, int64_t ignore_index, str reduction) -> Tensor(out), Tensor(total_weight)
  args : (Tensor input, Tensor label, Tensor weight, Tensor total_weight, Tensor out_grad, int64_t ignore_index, str reduction)
  output : Tensor(input_grad)
  infer_meta :
    func : NllLossGradInferMeta
  kernel :
    func : nll_loss_grad
    data_type : input
  optional : weight

- backward_api : p_norm_grad
  forward : p_norm(Tensor x,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : p_norm_grad

- backward_api : pad3d_grad
  forward : pad3d(Tensor x, IntArray paddings, str mode,  float pad_value, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray paddings, str mode,  float pad_value, str data_format)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pad3d_grad

- backward_api : pool2d_grad
  forward : pool2d(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
  output : Tensor(x_grad)
  infer_meta :
    func : PoolGradInferMeta
  kernel :
    func : pool2d_grad

- backward_api : pool3d_grad
  forward : pool3d(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
  output : Tensor(x_grad)
  infer_meta :
    func : PoolGradInferMeta
  kernel :
    func : pool3d_grad

- backward_api : pow_grad
  forward : pow(Tensor x, Scalar s) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, Scalar s=-1)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pow_grad

- backward_api : prelu_grad
  forward : prelu(Tensor x, Tensor alpha, str data_format, str mode) -> Tensor(out)
  args : (Tensor x, Tensor alpha, Tensor out_grad, str data_format, str mode)
  output : Tensor(x_grad), Tensor(alpha_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, alpha]
  kernel :
    func : prelu_grad

- backward_api : psroi_pool_grad
  forward : psroi_pool (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, int output_channels, float spatial_scale) -> Tensor(out)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor out_grad, int pooled_height, int pooled_width, int output_channels, float spatial_scale)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : psroi_pool_grad
  optional : boxes_num

# output is optional
- backward_api : put_along_axis_grad
  forward : put_along_axis (Tensor x, Tensor index, Tensor value, int axis, str reduce) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad, int axis, str reduce)
  output : Tensor(x_grad), Tensor(value_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, index]
  kernel :
    func : put_along_axis_grad

- backward_api : real_grad
  forward : real (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  invoke : real_grad_impl(out_grad)

- backward_api : reciprocal_grad
  forward : reciprocal (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : reciprocal_grad

- backward_api : reduce_prod_grad
  forward : reduce_prod (Tensor x, int64_t[] dims, bool keep_dim, bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] dims,  bool keep_dim, bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : prod_grad

- backward_api : relu_double_grad
  forward : relu_grad (Tensor out, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor out, Tensor grad_x_grad)
  output : Tensor(out_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [out, out]
  kernel :
    func : relu_double_grad

- backward_api : relu_grad
  forward : relu (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : relu_grad
  backward: relu_double_grad

- backward_api : reshape_grad
  forward : reshape_with_xshape (Tensor x, IntArray shape) -> Tensor(out), Tensor(xshape)
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param : [xshape]
  kernel :
    func : reshape_grad
    param : [out_grad]
    data_type: out_grad
    backend: out_grad
    layout: out_grad

- backward_api : roi_align_grad
  forward : roi_align (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, float spatial_scale, int sampling_ratio, bool aligned) -> Tensor(out)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor out_grad, int pooled_height, int pooled_width, float spatial_scale, int sampling_ratio, bool aligned)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roi_align_grad
  optional : boxes_num

- backward_api : roi_pool_grad
  forward : roi_pool (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, float spatial_scale) -> Tensor(out), Tensor(arg_max)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor arg_max, Tensor out_grad, int pooled_height, int pooled_width, float spatial_scale)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roi_pool_grad
  optional : boxes_num

- backward_api : roll_grad
  forward : roll(Tensor x, IntArray shifts, int64_t[] axis) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray shifts, int64_t[] axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roll_grad
    data_type : x

- backward_api : round_grad
  forward : round(Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out_grad]
  kernel :
    func : round_grad

- backward_api : rsqrt_grad
  forward : rsqrt (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : rsqrt_grad

- backward_api : scale_grad
  forward : scale (Tensor x, Scalar scale, float bias, bool bias_after_scale) -> Tensor(out)
  args : (Tensor out_grad, Scalar scale=1.0, float bias=0.0, bool bias_after_scale=true)
  output : Tensor(x_grad)
  invoke : scale(out_grad, scale, bias, bias_after_scale)

- backward_api : scatter_grad
  forward : scatter (Tensor x, Tensor index, Tensor updates, bool overwrite) -> Tensor(out)
  args : (Tensor index, Tensor updates, Tensor out_grad, bool overwrite)
  output : Tensor(x_grad), Tensor(updates_grad)
  infer_meta :
    func : ScatterGradInferMeta
    param : [index, updates, out_grad, overwrite]
  kernel :
    func : scatter_grad
  no_need_buffer : updates

- backward_api : scatter_nd_add_grad
  forward : scatter_nd_add (Tensor x, Tensor index, Tensor updates) -> Tensor(out)
  args : (Tensor index, Tensor updates, Tensor out_grad)
  output : Tensor(x_grad), Tensor(updates_grad)
  infer_meta :
    func : ScatterNdAddGradInferMeta
    param : [index, updates, out_grad]
  kernel :
    func : scatter_nd_add_grad
  no_need_buffer : updates

- backward_api : segment_pool_grad
  forward : segment_pool (Tensor x, Tensor segment_ids, str pooltype) -> Tensor(out), Tensor(summed_ids)
  args : (Tensor x, Tensor segment_ids, Tensor out, Tensor summed_ids, Tensor out_grad, str pooltype)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : segment_pool_grad

- backward_api : selu_grad
  forward : selu (Tensor x, float scale, float alpha) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, float scale, float alpha)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : selu_grad

- backward_api : sigmoid_cross_entropy_with_logits_grad
  forward : sigmoid_cross_entropy_with_logits (Tensor x, Tensor label, bool normalize, int ignore_index) -> Tensor(out)
  args : (Tensor x, Tensor label, Tensor out_grad, bool normalize, int ignore_index)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sigmoid_cross_entropy_with_logits_grad

- backward_api : sigmoid_double_grad
  forward : sigmoid_grad (Tensor out, Tensor fwd_grad_out) -> Tensor(grad_x)
  args : (Tensor out, Tensor fwd_grad_out, Tensor grad_x_grad)
  output : Tensor(out_grad), Tensor(fwd_grad_out_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [out, fwd_grad_out]
  kernel :
    func : sigmoid_double_grad
  backward : sigmoid_triple_grad

- backward_api : sigmoid_grad
  forward : sigmoid (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : sigmoid_grad
  backward : sigmoid_double_grad

- backward_api : sigmoid_triple_grad
  forward : sigmoid_double_grad (Tensor out, Tensor fwd_grad_out, Tensor grad_grad_x) -> Tensor(grad_out), Tensor(grad_grad_out)
  args : (Tensor out, Tensor fwd_grad_out, Tensor grad_grad_x, Tensor grad_out_grad, Tensor grad_grad_out_grad)
  output : Tensor(out_grad), Tensor(fwd_grad_out_grad), Tensor(grad_grad_x_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [out, fwd_grad_out, grad_grad_x]
  kernel :
    func : sigmoid_double_grad

- backward_api : silu_grad
  forward : silu (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : silu_grad

- backward_api : sin_grad
  forward : sin (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sin_grad

- backward_api : sinh_grad
  forward : sinh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sinh_grad

- backward_api : slice_grad
  forward : slice (Tensor input, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis) -> Tensor(out)
  args : (Tensor input, Tensor out_grad, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : slice_grad

- backward_api : soft_shrink_grad
  forward : soft_shrink (Tensor x, float lambda) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float lambda)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : soft_shrink_grad

- backward_api : softmax_grad
  forward : softmax (Tensor x, int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : softmax_grad

- backward_api : split_grad
  forward : split (Tensor x, IntArray num_or_sections, Scalar axis) -> Tensor[](out)
  args : (Tensor[] out_grad, Scalar axis)
  output : Tensor(x_grad)
  invoke : concat( out_grad, axis)
# TODO(zhangyunfei) The config of double grad and triple grad will be supported in the future.

- backward_api : sqrt_grad
  forward : sqrt (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : sqrt_grad

- backward_api : square_grad
  forward : square (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : square_grad

- backward_api : squeeze_grad
  forward : squeeze(Tensor x, int[] axes) -> Tensor(xshape), Tensor(out)
  args : (Tensor xshape, Tensor out_grad, int[] axes)
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param: [xshape]
  kernel :
    func : squeeze_grad

- backward_api : stack_grad
  forward : stack (Tensor[] x, int axis) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad, int axis)
  output : Tensor[](x_grad)
  invoke : stack_grad_impl(x, out_grad, axis)
  no_need_buffer : x

- backward_api : strided_slice_grad
  forward : strided_slice (Tensor x, int[] axes, IntArray starts, IntArray ends, IntArray strides) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] axes, IntArray starts, IntArray ends, IntArray strides)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : strided_slice_grad

- backward_api : subtract_grad
  forward : subtract (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : subtract_grad
  no_need_buffer : x, y

- backward_api : sum_grad
  forward : sum (Tensor x, int64_t[] dims={}, DataType out_dtype=paddle::experimental::DataType::UNDEFINED, bool keep_dim=false) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int64_t[] dims, bool keep_dim, bool reduce_all=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sum_grad

- backward_api : take_along_axis_grad
  forward : take_along_axis (Tensor x, Tensor index, int axis) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : take_along_axis_grad

- backward_api : tan_grad
  forward : tan (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : tan_grad

- backward_api : tanh_grad
  forward : tanh (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : tanh_grad

- backward_api : tanh_shrink_grad
  forward : tanh_shrink (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : tanh_shrink_grad

- backward_api : thresholded_relu_grad
  forward : thresholded_relu (Tensor x, float threshold) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : thresholded_relu_grad

- backward_api : tile_grad
  forward : tile (Tensor x, IntArray repeat_times) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray repeat_times)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : tile_grad
  no_need_buffer : x

- backward_api : top_k_grad
  forward : top_k (Tensor x, Scalar k, int axis = -1, bool largest = true, bool sorted = true) -> Tensor(out), Tensor(indices)
  args : (Tensor x, Tensor indices, Tensor out_grad, Scalar k = -1, int axis = -1, bool largest = true, bool sorted = true)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : top_k_grad

- backward_api : trace_grad
  forward : trace (Tensor x, int offset, int axis1, int axis2) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int offset, int axis1, int axis2)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : trace_grad
  no_need_buffer : x

- backward_api : transpose_grad
  forward : transpose (Tensor x, int[] axis) -> Tensor(out)
  args : (Tensor out_grad, int[] axis)
  output : Tensor(x_grad)
  infer_meta :
    func : TransposeGradInferMeta
    param : [out_grad, axis]
  kernel :
    func : transpose_grad

- backward_api : tril_triu_grad
  forward : tril_triu(Tensor x,  int diagonal,  bool lower) -> Tensor(out)
  args : (Tensor out_grad,  int diagonal,  bool lower)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : tril_triu_grad

- backward_api : trunc_grad
  forward : trunc (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : trunc_grad

- backward_api : unbind_grad
  forward : unbind (Tensor input, int axis) -> Tensor[](out)
  args : (Tensor[] out_grad, int axis)
  output : Tensor(input_grad)
  invoke : stack(out_grad, axis)

- backward_api : unfold_grad
  forward : unfold (Tensor x, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : unfold_grad
  no_need_buffer : x

- backward_api : unsqueeze_grad
  forward : unsqueeze(Tensor x, IntArray axes) -> Tensor(xshape), Tensor(out)
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param: [xshape]
  kernel :
    func : unsqueeze_grad

- backward_api : where_grad
  forward : where (Tensor condition, Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor condition, Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : where_grad