- 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_api : matmul_double_grad
  forward : matmul_grad (Tensor x, Tensor y, Tensor out_grad, bool transpose_x, bool transpose_y) -> Tensor(dx), Tensor(dy)
  args : (Tensor x, Tensor y, Tensor out_grad, Tensor dx_grad, Tensor dy_grad, bool transpose_x, bool transpose_y)
  output : Tensor(d2x), Tensor(d2y), Tensor(dout_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, y, out_grad]
  kernel :
    func : matmul_double_grad
  optional : dx_grad, dy_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, float bias=0.0, bool bias_after_scale=true)
  output : Tensor(x_grad)
  invoke : scale(out_grad, scale, bias, bias_after_scale)


- backward_api : add_grad
  forward : add (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 : add_grad

- 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 : 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

- 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 : norm_grad
#   forward : norm (Tensor x, int axis, float epsilon, bool is_test) -> Tensor(out), Tensor(norm)
#   args : (Tensor out_grad, Tensor x, Tensor norm, int axis, float epsilon, bool is_test)
#   output : Tensor(x_grad)
#   infer_meta :
#     func : UnchangedInferMeta
#     param : [x]
#   kernel :
#     func : norm_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

# - backward_api : split_grad
#   forward : split (Tensor x, ScalarArray 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 : matmul_triple_grad
#   forward : matmul_double_grad (Tensor x, Tensor y, Tensor out_grad, Tensor dx_grad, Tensor dy_grad, bool transpose_x, bool transpose_y) -> Tensor(d2x), Tensor(d2y), Tensor(dout_grad)
#   args : (Tensor x, Tensor y, Tensor out_grad, Tensor dx_grad, Tensor dy_grad, Tensor d2x_grad, Tensor d2y_grad, Tensor dout_grad_grad, bool transpose_x, bool transpose_y)
#   output : Tensor(d3x), Tensor(d3y), Tensor(d2out_grad), Tensor(ddx_grad), Tensor(ddy_grad)
#   infer_meta :
#     func : MatmulTripleGradInferMeta
#   kernel :
#     func : matmul_triple_grad

# - 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 : 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 : lerp_grad
#   forward : transpose (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 : 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

- backward_api : scatter_nd_add_grad
  forward : scatter (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_grad

- backward_api : addmm_grad
  forward : scatter (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 : 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

# - 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 : [x, y]
#   kernel :
#     func : where_grad

# - backward_api : triangular_solve_grad
#   forward : triangular_solve (Tensor x, Tensor y, bool upper, bool tranpose, bool unitriangular) -> Tensor(out)
#   args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, bool upper, bool tranpose, bool unitriangular)
#   output : Tensor(x_grad), Tensor(y_grad)
#   infer_meta :
#     func : GeneralBinaryGradInferMeta
#     param : [x, y]
#   kernel :
#     func : triangular_solve_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

- 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 : atan2_grad
  forward : cross (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 : 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 : 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 : 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 : 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