diff --git a/paddle/fluid/eager/auto_code_generator/generator/codegen_utils.py b/paddle/fluid/eager/auto_code_generator/generator/codegen_utils.py
index 580dc14f43b16..9022e800905d0 100644
--- a/paddle/fluid/eager/auto_code_generator/generator/codegen_utils.py
+++ b/paddle/fluid/eager/auto_code_generator/generator/codegen_utils.py
@@ -83,10 +83,10 @@ def ReadBwdFile(filepath):
     ret = {}
     if contents is not None:
         for content in contents:
-            assert 'backward_api' in content.keys(), AssertMessage(
-                'backward_api', content.keys())
-            if 'backward_api' in content.keys():
-                api_name = content['backward_api']
+            assert 'backward_op' in content.keys(), AssertMessage(
+                'backward_op', content.keys())
+            if 'backward_op' in content.keys():
+                api_name = content['backward_op']
 
             ret[api_name] = content
     f.close()
diff --git a/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py b/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
index fd1acbf932984..100dfd5740544 100644
--- a/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
+++ b/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
@@ -1485,7 +1485,7 @@ def GenerateHigherOrderNodeCreationCode(self):
         if next_grad_api_contents:
             # Fake forward_api_contents and backward_api_contents
             forward_api_contents = grad_api_contents
-            forward_api_contents['op'] = forward_api_contents['backward_api']
+            forward_api_contents['op'] = forward_api_contents['backward_op']
             backward_api_contents = next_grad_api_contents
 
             next_node_generator = DygraphFunctionGeneratorBase(
@@ -1959,8 +1959,7 @@ def GenerateCode(self):
                 forward_api_contents = backward_api_contents
 
                 # Fake forward_api_content
-                forward_api_contents['op'] = forward_api_contents[
-                    'backward_api']
+                forward_api_contents['op'] = forward_api_contents['backward_op']
                 backward_api_contents = next_grad_api_contents
 
         if len(namespace) > 0:
diff --git a/paddle/phi/api/yaml/backward.yaml b/paddle/phi/api/yaml/backward.yaml
index d2ed2533ae03e..9d81435169c6a 100644
--- a/paddle/phi/api/yaml/backward.yaml
+++ b/paddle/phi/api/yaml/backward.yaml
@@ -1,4 +1,4 @@
-- backward_api : atan2_grad
+- backward_op : 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)
@@ -8,7 +8,7 @@
   kernel :
     func : atan2_grad
 
-- backward_api : cholesky_grad
+- backward_op : cholesky_grad
   forward : cholesky (Tensor x, bool upper) -> Tensor(out)
   args : (Tensor out, Tensor out_grad, bool upper)
   output : Tensor(x_grad)
@@ -18,7 +18,7 @@
   kernel :
     func : cholesky_grad
 
-- backward_api : cholesky_solve_grad
+- backward_op : 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)
@@ -28,7 +28,7 @@
   kernel :
     func : cholesky_solve_grad
 
-- backward_api : cross_grad
+- backward_op : 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)
@@ -39,7 +39,7 @@
     func : cross_grad
     data_type : out_grad
 
-- backward_api : diag_grad
+- backward_op : diag_grad
   forward : diag (Tensor x, int offset, float padding_value) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, int offset)
   output : Tensor(x_grad)
@@ -51,7 +51,7 @@
     data_type : out_grad
   no_need_buffer : x
 
-- backward_api : diagonal_grad
+- backward_op : 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)
@@ -63,7 +63,7 @@
     data_type : out_grad
   no_need_buffer : x
 
-- backward_api : digamma_grad
+- backward_op : digamma_grad
   forward : digamma (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -73,7 +73,7 @@
   kernel :
     func : digamma_grad
 
-- backward_api : dist_grad
+- backward_op : 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)
@@ -83,7 +83,7 @@
   kernel :
     func : dist_grad
 
-- backward_api : dot_grad
+- backward_op : dot_grad
   forward : dot (Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad)
@@ -94,7 +94,7 @@
     func : dot_grad
     data_type : out_grad
 
-- backward_api : erf_grad
+- backward_op : erf_grad
   forward : erf (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -105,7 +105,7 @@
     func : erf_grad
     data_type : out_grad
 
-- backward_api : erfinv_grad
+- backward_op : erfinv_grad
   forward : erfinv (Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -115,7 +115,7 @@
   kernel :
     func : erfinv_grad
     
-- backward_api : fft_c2c_grad
+- backward_op : fft_c2c_grad
   forward: fft_c2c(Tensor x, int64_t[] axes, str normalization, bool forward) -> Tensor(out)
   args : (Tensor out_grad, int64_t[] axes, str normalization, bool forward)
   output: Tensor(x_grad)
@@ -125,7 +125,7 @@
   kernel :
     func : fft_c2c_grad
 
-- backward_api : fft_c2r_grad
+- backward_op : fft_c2r_grad
   forward: fft_c2r(Tensor x, int64_t[] axes, str normalization, bool forward, int64_t last_dim_size) -> Tensor(out)
   args : (Tensor out_grad, int64_t[] axes, str normalization, bool forward, int64_t last_dim_size)
   output: Tensor(x_grad)
@@ -135,7 +135,7 @@
     func : fft_c2r_grad
     data_type: out_grad
 
-- backward_api : fft_r2c_grad
+- backward_op : fft_r2c_grad
   forward: fft_r2c(Tensor x, int64_t[] axes, str normalization, bool forward, bool onesided) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, int64_t[] axes, str normalization, bool forward, bool onesided)
   output: Tensor(x_grad)
@@ -147,7 +147,7 @@
     data_type: out_grad
   no_need_buffer: x
 
-- backward_api : graph_send_uv_grad
+- backward_op : graph_send_uv_grad
   forward : graph_send_uv (Tensor x, Tensor y, Tensor src_index, Tensor dst_index, str message_op = "ADD") -> Tensor(out)
   args: (Tensor x, Tensor y, Tensor src_index, Tensor dst_index, Tensor out_grad, str message_op = "ADD")
   output : Tensor(x_grad), Tensor(y_grad)
@@ -158,7 +158,7 @@
     func : graph_send_uv_grad
     data_type : x
 
-- backward_api : lgamma_grad
+- backward_op : lgamma_grad
   forward : lgamma(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -168,7 +168,7 @@
   kernel :
     func : lgamma_grad
 
-- backward_api : mv_grad
+- backward_op : 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)
@@ -178,7 +178,7 @@
   kernel :
     func : mv_grad
 
-- backward_api : poisson_grad
+- backward_op : poisson_grad
   forward : poisson (Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
@@ -188,7 +188,7 @@
   kernel :
     func : poisson_grad
 
-- backward_api : solve_grad
+- backward_op : solve_grad
   forward : solve (Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad)
@@ -198,7 +198,7 @@
   kernel :
     func : solve_grad
 
-- backward_api : trace_grad
+- backward_op : 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)
@@ -210,7 +210,7 @@
     data_type : out_grad
   no_need_buffer : x
 
-- backward_api : trunc_grad
+- backward_op : trunc_grad
   forward : trunc (Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
diff --git a/paddle/phi/api/yaml/generator/backward_api_gen.py b/paddle/phi/api/yaml/generator/backward_api_gen.py
index f168430ee7c5c..1a8b9cc4d390b 100644
--- a/paddle/phi/api/yaml/generator/backward_api_gen.py
+++ b/paddle/phi/api/yaml/generator/backward_api_gen.py
@@ -28,7 +28,7 @@ def __init__(self, backward_item_yaml):
         self.no_need_buffer = self.parse_no_need_buffer(backward_item_yaml)
 
     def get_api_name(self, api_item_yaml):
-        return api_item_yaml['backward_api']
+        return api_item_yaml['backward_op']
 
     def parse_forward_config(self, forward_config):
         # api_name (const Tensor& input, ... , int attr, ...) -> Tensor(out)
diff --git a/paddle/phi/api/yaml/generator/parse_api.py b/paddle/phi/api/yaml/generator/parse_api.py
index bbecbcf26ada8..91fd44b709339 100644
--- a/paddle/phi/api/yaml/generator/parse_api.py
+++ b/paddle/phi/api/yaml/generator/parse_api.py
@@ -27,7 +27,7 @@ def main(api_yaml_path, output_path, backward):
             apis = []
         else:
             apis = [
-                parse_api_entry(api, "backward_api" if backward else "op")
+                parse_api_entry(api, "backward_op" if backward else "op")
                 for api in apis
             ]
 
diff --git a/paddle/phi/api/yaml/generator/parse_utils.py b/paddle/phi/api/yaml/generator/parse_utils.py
index 45aefdfd6dc05..f617f166dd13b 100644
--- a/paddle/phi/api/yaml/generator/parse_utils.py
+++ b/paddle/phi/api/yaml/generator/parse_utils.py
@@ -334,7 +334,7 @@ def parse_api_entry(api_entry: Dict[str, Any], name_field="op"):
     api["backward"] = backward
 
     # forward for backward_apis
-    is_backward_api = name_field == "backward_api"
+    is_backward_api = name_field == "backward_op"
     if is_backward_api:
         if "forward" in api_entry:
             forward = parse_forward(api_name, api_entry["forward"])
diff --git a/paddle/phi/api/yaml/legacy_backward.yaml b/paddle/phi/api/yaml/legacy_backward.yaml
index 7471d310a641e..76d60522d8633 100755
--- a/paddle/phi/api/yaml/legacy_backward.yaml
+++ b/paddle/phi/api/yaml/legacy_backward.yaml
@@ -1,4 +1,4 @@
-- backward_api : abs_double_grad
+- backward_op : abs_double_grad
   forward : abs_grad (Tensor x, Tensor grad_out) -> Tensor(grad_x)
   args : (Tensor x, Tensor grad_x_grad)
   output : Tensor(grad_out_grad)
@@ -10,7 +10,7 @@
   data_transform:
     skip_transform : grad_x_grad
 
-- backward_api : abs_grad
+- backward_op : abs_grad
   forward : abs (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -23,7 +23,7 @@
     skip_transform : out_grad
   backward : abs_double_grad
 
-- backward_api : acos_grad
+- backward_op : acos_grad
   forward : acos (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -34,7 +34,7 @@
     func : acos_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : acosh_grad
+- backward_op : acosh_grad
   forward : acosh (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -45,7 +45,7 @@
     func : acosh_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : add_double_grad
+- backward_op : add_double_grad
   forward : add_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
   args : (Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
   output : Tensor(grad_out_grad)
@@ -58,7 +58,7 @@
   backward : add_triple_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : add_grad
+- backward_op : 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)
@@ -71,7 +71,7 @@
   backward : add_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : add_triple_grad
+- backward_op : add_triple_grad
   forward : add_double_grad (Tensor y, Tensor grad_out, Tensor grad_grad_x, Tensor grad_grad_y, int axis = -1) -> Tensor(grad_grad_out)
   args : (Tensor grad_grad_x, Tensor grad_grad_y, Tensor grad_grad_out_grad, int axis = -1)
   output : Tensor(grad_grad_x_grad), Tensor(grad_grad_y_grad)
@@ -82,7 +82,7 @@
     func : add_triple_grad
   inplace : (grad_grad_out_grad -> grad_grad_x_grad)
 
-- backward_api : addmm_grad
+- backward_op : 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)
@@ -92,7 +92,7 @@
   kernel :
     func : addmm_grad
 
-- backward_api : affine_grid_grad
+- backward_op : affine_grid_grad
   forward : affine_grid (Tensor input, IntArray outputShape, bool use_cudnn=true, bool align_corners=true) -> Tensor(output)
   args : (Tensor output_grad, IntArray outputShape, bool use_cudnn=true, bool align_corners=true)
   output : Tensor(input_grad)
@@ -104,7 +104,7 @@
     param : [output_grad, outputShape, align_corners]
     use_gpudnn: use_cudnn
 
-- backward_api : amax_grad
+- backward_op : amax_grad
   forward: amax (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)
@@ -114,7 +114,7 @@
   kernel :
     func : amax_grad
 
-- backward_api : amin_grad
+- backward_op : amin_grad
   forward: amin (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)
@@ -124,7 +124,7 @@
   kernel :
     func : amin_grad
 
-- backward_api : angle_grad
+- backward_op : angle_grad
   forward : angle (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -136,7 +136,7 @@
   data_transform:
     skip_transform : out_grad
 
-- backward_api : argsort_grad
+- backward_op : 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)
@@ -148,19 +148,19 @@
     data_type : out_grad
   no_need_buffer : x
 
-- backward_api : as_complex_grad
+- backward_op : as_complex_grad
   forward : as_complex (Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
   invoke : as_real(out_grad)
 
-- backward_api : as_real_grad
+- backward_op : as_real_grad
   forward : as_real (Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
   invoke : as_complex(out_grad)
 
-- backward_api : asin_grad
+- backward_op : asin_grad
   forward : asin (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -171,7 +171,7 @@
     func : asin_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : asinh_grad
+- backward_op : asinh_grad
   forward : asinh (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -182,13 +182,13 @@
     func : asinh_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : assign_grad
+- backward_op : assign_grad
   forward : assign (Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
   invoke : assign(out_grad)
 
-- backward_api : assign_out__grad
+- backward_op : assign_out__grad
   forward : assign_out_ (Tensor x, Tensor output) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
@@ -198,7 +198,7 @@
     func : assign
   inplace : (out_grad -> x_grad)
 
-- backward_api : atan_grad
+- backward_op : atan_grad
   forward : atan (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -209,7 +209,7 @@
     func : atan_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : atanh_grad
+- backward_op : atanh_grad
   forward : atanh (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -220,7 +220,7 @@
     func : atanh_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : batch_norm_double_grad
+- backward_op : batch_norm_double_grad
   forward : batch_norm_grad (Tensor x, Tensor scale, Tensor bias, Tensor out_mean, Tensor out_variance, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor grad_out, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu) -> Tensor(grad_x), Tensor(grad_scale), Tensor(grad_bias)
   args : (Tensor x, Tensor scale, Tensor out_mean, Tensor out_variance, Tensor saved_mean, Tensor saved_variance, Tensor grad_out,  Tensor grad_x_grad, Tensor grad_scale_grad, Tensor grad_bias_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(grad_out_grad)
@@ -233,7 +233,7 @@
   optional : out_mean, out_variance
   inplace : (grad_out -> grad_out_grad)
 
-- backward_api : batch_norm_grad
+- backward_op : 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)
@@ -246,7 +246,7 @@
   optional : mean_out, variance_out, reserve_space
   backward : batch_norm_double_grad
 
-- backward_api : bce_loss_grad
+- backward_op : bce_loss_grad
   forward : bce_loss (Tensor input, Tensor label) -> Tensor(out)
   args : (Tensor input, Tensor label, Tensor out_grad)
   output : Tensor(input_grad)
@@ -257,7 +257,7 @@
     func : bce_loss_grad
   inplace : (out_grad -> input_grad)
 
-- backward_api : bicubic_interp_grad
+- backward_op : bicubic_interp_grad
   forward : bicubic_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
   args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
   output : Tensor(x_grad)
@@ -269,7 +269,7 @@
     func : bicubic_interp_grad
     data_type : output_grad
 
-- backward_api : bilinear_interp_grad
+- backward_op : bilinear_interp_grad
   forward : bilinear_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
   args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
   output : Tensor(x_grad)
@@ -281,7 +281,7 @@
     func : bilinear_interp_grad
     data_type : output_grad
 
-- backward_api : bilinear_tensor_product_grad
+- backward_op : bilinear_tensor_product_grad
   forward : bilinear_tensor_product (Tensor x, Tensor y, Tensor weight, Tensor bias) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor weight, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad), Tensor(weight_grad), Tensor(bias_grad)
@@ -290,7 +290,7 @@
   kernel :
     func : bilinear_tensor_product_grad
 
-- backward_api : bmm_grad
+- backward_op : bmm_grad
   forward : bmm (Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad)
@@ -299,7 +299,7 @@
   kernel :
     func : bmm_grad
 
-- backward_api : brelu_grad
+- backward_op : 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)
@@ -310,7 +310,7 @@
     func : brelu_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : broadcast_tensors_grad
+- backward_op : broadcast_tensors_grad
   forward : broadcast_tensors (Tensor[] x) -> Tensor[](out)
   args : (Tensor[] x, Tensor[] out_grad)
   output : Tensor[](x_grad)
@@ -322,7 +322,7 @@
     param : [out_grad]
   no_need_buffer : x
 
-- backward_api : cast_grad
+- backward_op : cast_grad
   forward : cast (Tensor x, DataType out_dtype) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -334,7 +334,7 @@
     data_type : out_grad
   no_need_buffer : x
 
-- backward_api : ceil_grad
+- backward_op : ceil_grad
   forward : ceil(Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
@@ -345,7 +345,7 @@
     func : ceil_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : celu_double_grad
+- backward_op : celu_double_grad
   forward : celu_grad(Tensor x, Tensor grad_out, float alpha) -> Tensor(grad_x)
   args : (Tensor x, Tensor grad_out, Tensor grad_x_grad, float alpha)
   output : Tensor(x_grad), Tensor(grad_out_grad)
@@ -356,7 +356,7 @@
     func : celu_double_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : celu_grad
+- backward_op : celu_grad
   forward : celu(Tensor x, float alpha) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float alpha)
   output : Tensor(x_grad)
@@ -368,7 +368,7 @@
   backward : celu_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : clip_double_grad
+- backward_op : clip_double_grad
   forward : clip_grad (Tensor x, Tensor grad_out, Scalar min = 0., Scalar max = 0.) -> Tensor(grad_x)
   args : (Tensor x, Tensor grad_x_grad, Scalar min = 0., Scalar max = 0.)
   output : Tensor(grad_out_grad)
@@ -378,7 +378,7 @@
   kernel :
     func : clip_grad
 
-- backward_api : clip_grad
+- backward_op : 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)
@@ -390,7 +390,7 @@
   backward : clip_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : complex_grad
+- backward_op : complex_grad
   forward : complex (Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad)
@@ -400,13 +400,13 @@
     func : complex_grad
     data_type : x
 
-- backward_api : concat_double_grad
+- backward_op : concat_double_grad
   forward : concat_grad (Tensor[] x, Tensor grad_out, Scalar axis) -> Tensor[](grad_x)
   args : (Tensor[] grad_x_grad, Scalar axis = 0)
   output : Tensor(grad_out_grad)
   invoke : concat(grad_x_grad, axis)
 
-- backward_api : concat_grad
+- backward_op : concat_grad
   forward : concat (Tensor[] x, Scalar axis) -> Tensor(out)
   args : (Tensor[] x, Tensor out_grad, Scalar axis = 0)
   output : Tensor[](x_grad){x.size()}
@@ -418,7 +418,7 @@
   no_need_buffer : x
   backward : concat_double_grad
 
-- backward_api : conj_grad
+- backward_op : conj_grad
   forward : conj (Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
@@ -428,7 +428,7 @@
   kernel :
     func : conj
 
-- backward_api : conv2d_grad
+- backward_op : 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)
@@ -440,7 +440,7 @@
     use_gpudnn : true
   backward : conv2d_grad_grad
 
-- backward_api : conv2d_grad_grad
+- backward_op : conv2d_grad_grad
   forward : conv2d_grad (Tensor input, Tensor filter, Tensor grad_out,  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(grad_input), Tensor(grad_filter)
   args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_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), Tensor(grad_out_grad)
@@ -452,7 +452,7 @@
     use_gpudnn : true
   optional : grad_input_grad, grad_filter_grad
 
-- backward_api : conv2d_transpose_double_grad
+- backward_op : conv2d_transpose_double_grad
   forward : conv2d_transpose_grad(Tensor x, Tensor filter, Tensor grad_out, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(grad_x), Tensor(grad_filter)
   args : (Tensor x, Tensor filter, Tensor grad_out, Tensor grad_x_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
   output : Tensor(x_grad), Tensor(filter_grad), Tensor(grad_out_grad)
@@ -462,7 +462,7 @@
     func : conv2d_transpose_grad_grad
     use_gpudnn : true
 
-- backward_api : conv2d_transpose_grad
+- backward_op : conv2d_transpose_grad
   forward : conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, IntArray 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, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
   output : Tensor(x_grad), Tensor(filter_grad)
@@ -473,7 +473,7 @@
     use_gpudnn : true
   backward : conv2d_transpose_double_grad
 
-- backward_api : conv3d_grad
+- backward_op : conv3d_grad
   forward : conv3d (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)
@@ -485,7 +485,7 @@
     use_gpudnn : true
   backward : conv3d_grad_grad
 
-- backward_api : conv3d_grad_grad
+- backward_op : conv3d_grad_grad
   forward : conv3d_grad (Tensor input, Tensor filter, Tensor grad_out,  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(grad_input), Tensor(grad_filter)
   args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_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), Tensor(grad_out_grad)
@@ -497,7 +497,7 @@
     use_gpudnn : true
   optional : grad_input_grad, grad_filter_grad
 
-- backward_api : conv3d_transpose_grad
+- backward_op : 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)
@@ -507,7 +507,7 @@
     func : conv3d_transpose_grad
     use_gpudnn : true
 
-- backward_api : cos_grad
+- backward_op : cos_grad
   forward : cos (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -518,7 +518,7 @@
     func : cos_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : cosh_grad
+- backward_op : cosh_grad
   forward : cosh (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -529,7 +529,7 @@
     func : cosh_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : crop_tensor_grad
+- backward_op : crop_tensor_grad
   forward : crop_tensor (Tensor x, IntArray shape, IntArray offsets) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, IntArray offsets)
   output : Tensor(x_grad)
@@ -539,7 +539,7 @@
     func : crop_tensor_grad
     data_type : x
 
-- backward_api : cross_entropy_with_softmax_grad
+- backward_op : 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)
@@ -550,7 +550,7 @@
     data_type : softmax
   inplace : (softmax -> input_grad)
 
-- backward_api : cumprod_grad
+- backward_op : cumprod_grad
   forward : cumprod (Tensor x, int dim) -> Tensor(out)
   args : (Tensor x, Tensor out, Tensor out_grad, int dim)
   output : Tensor(x_grad)
@@ -560,13 +560,13 @@
   kernel :
     func : cumprod_grad
 
-- backward_api : cumsum_grad
+- backward_op : cumsum_grad
   forward : cumsum(Tensor x, Scalar axis, bool flatten, bool exclusive, bool reverse) -> Tensor(out)
   args : (Tensor out_grad, Scalar axis, bool flatten, bool exclusive, bool reverse)
   output : Tensor(x_grad)
   invoke : cumsum(out_grad, axis, flatten, exclusive, !reverse)
 
-- backward_api : deformable_conv_grad
+- backward_op : deformable_conv_grad
   forward : deformable_conv(Tensor x, Tensor offset, Tensor filter, Tensor mask, int[] strides, int[] paddings, int[] dilations, int deformable_groups, int groups, int im2col_step) -> Tensor(out)
   args : (Tensor x, Tensor offset, Tensor filter, Tensor mask, Tensor out_grad, int[] strides, int[] paddings, int[] dilations, int deformable_groups, int groups, int im2col_step)
   output : Tensor(x_grad), Tensor(offset_grad), Tensor(filter_grad), Tensor(mask_grad)
@@ -577,7 +577,7 @@
     data_type : x
   optional : mask
 
-- backward_api : depthwise_conv2d_grad
+- backward_op : depthwise_conv2d_grad
   forward : depthwise_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, bool fuse_relu, bool use_gpudnn) -> 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, bool fuse_relu, bool use_gpudnn)
   output : Tensor(input_grad), Tensor(filter_grad)
@@ -590,7 +590,7 @@
     use_gpudnn : use_gpudnn
   backward : depthwise_conv2d_grad_grad
 
-- backward_api : depthwise_conv2d_grad_grad
+- backward_op : depthwise_conv2d_grad_grad
   forward : depthwise_conv2d_grad (Tensor input, Tensor filter, Tensor grad_out, int[] strides, int[] paddings, str paddding_algorithm, int groups, int[] dilations, str data_format, bool use_addto, int workspace_size_MB, bool exhaustive_search, bool fuse_relu, bool use_gpudnn) -> Tensor(grad_input), Tensor(grad_filter)
   args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_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, bool fuse_relu)
   output : Tensor(input_grad), Tensor(filter_grad), Tensor(grad_out_grad)
@@ -601,7 +601,7 @@
     func : depthwise_conv2d_grad_grad
   optional : grad_input_grad, grad_filter_grad
 
-- backward_api : depthwise_conv2d_transpose_grad
+- backward_op : depthwise_conv2d_transpose_grad
   forward : depthwise_conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, IntArray 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, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
   output : Tensor(x_grad), Tensor(filter_grad)
@@ -610,7 +610,7 @@
   kernel :
     func : depthwise_conv2d_transpose_grad
 
-- backward_api : det_grad
+- backward_op : det_grad
   forward : det (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -620,7 +620,7 @@
   kernel :
     func : determinant_grad
 
-- backward_api : divide_double_grad
+- backward_op : divide_double_grad
   forward : divide_grad (Tensor x, Tensor y, Tensor out, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
   args : (Tensor y, Tensor out, Tensor grad_x, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
   output : Tensor(y_grad), Tensor(out_grad), Tensor(grad_out_grad)
@@ -633,7 +633,7 @@
   optional : grad_x_grad, grad_y_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : divide_grad
+- backward_op : 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)
@@ -644,7 +644,7 @@
     func : divide_grad
   backward : divide_double_grad
 
-- backward_api : dropout_grad
+- backward_op : dropout_grad
   forward : dropout (Tensor x, Tensor seed_tensor, Scalar p, bool is_test, str mode, int seed, bool fix_seed) -> Tensor(out), Tensor(mask)
   args : (Tensor mask, Tensor out_grad, Scalar p, bool is_test, str mode)
   output : Tensor(x_grad)
@@ -654,7 +654,7 @@
   kernel :
     func : dropout_grad
 
-- backward_api : eig_grad
+- backward_op : eig_grad
   forward : eig (Tensor x) -> 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)
@@ -667,7 +667,7 @@
   data_transform:
     skip_transform : out_w, out_w_grad
 
-- backward_api : eigh_grad
+- backward_op : 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)
@@ -680,7 +680,7 @@
   data_transform:
     skip_transform : out_w, out_w_grad
 
-- backward_api : eigvalsh_grad
+- backward_op : eigvalsh_grad
   forward : eigvalsh (Tensor x, str uplo, bool is_test) -> Tensor(eigenvalues), Tensor(eigenvectors)
   args : (Tensor eigenvectors, Tensor eigenvalues_grad, str uplo, bool is_test)
   output : Tensor(x_grad)
@@ -692,7 +692,7 @@
   data_transform :
     skip_transform : eigenvalues_grad
 
-- backward_api : einsum_grad
+- backward_op : einsum_grad
   forward : einsum (Tensor[] x, str equation) -> Tensor(out), Tensor[](inner_cache), Tensor[](x_shape)
   args : (Tensor[] x_shape, Tensor[] inner_cache, Tensor out_grad, str equation)
   output : Tensor[](x_grad){x.size()}
@@ -702,7 +702,7 @@
   kernel :
     func : einsum_grad
 
-- backward_api : elementwise_pow_grad
+- backward_op : 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)
@@ -712,7 +712,7 @@
   kernel :
     func : elementwise_pow_grad
 
-- backward_api : elu_double_grad
+- backward_op : elu_double_grad
   forward : elu_grad (Tensor x, Tensor out, Tensor grad_out, float alpha)-> Tensor(grad_x)
   args : (Tensor x, Tensor grad_out, Tensor grad_x_grad, float alpha)
   output : Tensor(x_grad), Tensor(grad_out_grad)
@@ -723,7 +723,7 @@
     func : elu_double_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : elu_grad
+- backward_op : elu_grad
   forward : elu (Tensor x, float alpha) -> Tensor(out)
   args : (Tensor x, Tensor out, Tensor out_grad, float alpha)
   output : Tensor(x_grad)
@@ -735,13 +735,13 @@
   backward : elu_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : embedding_grad
+- backward_op : embedding_grad
   forward : embedding (Tensor x, Tensor weight, int64_t padding_idx=-1, bool sparse=false) -> Tensor(out)
   args : (Tensor x, Tensor weight, Tensor out_grad, int64_t padding_idx=-1, bool sparse=false)
   output : Tensor(weight_grad)
   invoke : embedding_grad_impl(x, weight, out_grad, padding_idx, sparse, weight_grad)
 
-- backward_api : exp_grad
+- backward_op : exp_grad
   forward : exp (Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -752,7 +752,7 @@
     func : exp_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : expand_as_grad
+- backward_op : 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)
@@ -763,13 +763,13 @@
     func : expand_as_grad
   no_need_buffer : x
 
-- backward_api : expand_double_grad
+- backward_op : expand_double_grad
   forward : expand_grad (Tensor x, Tensor grad_out, IntArray shape) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, IntArray shape)
   output : Tensor(grad_out_grad)
   invoke : expand(grad_x_grad, shape)
 
-- backward_api : expand_grad
+- backward_op : expand_grad
   forward : expand (Tensor x, IntArray shape) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, IntArray shape)
   output : Tensor(x_grad)
@@ -781,7 +781,7 @@
   no_need_buffer : x
   backward : expand_double_grad
 
-- backward_api : expm1_grad
+- backward_op : expm1_grad
   forward : expm1 (Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -792,7 +792,7 @@
     func : expm1_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : exponential__grad
+- backward_op : exponential__grad
   forward : exponential_ (Tensor x, float lambda) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
@@ -800,7 +800,7 @@
     func : UnchangedInferMeta
   invoke : zeros_like(out_grad)
 
-- backward_api : fill_diagonal_grad
+- backward_op : fill_diagonal_grad
   forward : fill_diagonal (Tensor x, float value, int offset, bool wrap) -> Tensor(out)
   args : (Tensor out_grad, float value, int offset, bool wrap)
   output : Tensor(x_grad)
@@ -809,7 +809,7 @@
   kernel :
     func : fill_diagonal_grad
 
-- backward_api : fill_diagonal_tensor_grad
+- backward_op : fill_diagonal_tensor_grad
   forward : fill_diagonal_tensor (Tensor x, Tensor y, int64_t offset, int dim1, int dim2) -> Tensor(out)
   args : (Tensor out_grad, int64_t offset, int dim1, int dim2)
   output : Tensor(x_grad)
@@ -819,7 +819,7 @@
     func : fill_diagonal_tensor_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : fill_grad
+- backward_op : fill_grad
   forward : fill (Tensor x, Scalar value) -> Tensor(out)
   args : (Tensor out_grad, Scalar value)
   output : Tensor(x_grad)
@@ -830,7 +830,7 @@
     func : fill_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : flatten_grad
+- backward_op : 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)
@@ -844,13 +844,13 @@
     layout: out_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : flip_grad
+- backward_op : flip_grad
   forward : flip (Tensor x, int[] axis) -> Tensor(out)
   args : (Tensor out_grad, int[] axis)
   output : Tensor(x_grad)
   invoke : flip(out_grad, axis)
 
-- backward_api : floor_grad
+- backward_op : floor_grad
   forward : floor(Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
@@ -861,7 +861,7 @@
     func : floor_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : fmax_grad
+- backward_op : 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)
@@ -871,7 +871,7 @@
   kernel :
     func : fmax_grad
 
-- backward_api : fmin_grad
+- backward_op : 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)
@@ -881,7 +881,7 @@
   kernel :
     func : fmin_grad
 
-- backward_api : frame_grad
+- backward_op : frame_grad
   forward : frame(Tensor x, int frame_length, int hop_length, int axis) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, int frame_length, int hop_length, int axis)
   output : Tensor(x_grad)
@@ -891,7 +891,7 @@
   kernel :
     func : frame_grad
 
-- backward_api : frobenius_norm_grad
+- backward_op : 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)
@@ -901,7 +901,7 @@
   kernel :
     func : frobenius_norm_grad
 
-- backward_api : gather_grad
+- backward_op : 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)
@@ -913,7 +913,7 @@
     func : gather_grad
   no_need_buffer : x
 
-- backward_api : gather_nd_grad
+- backward_op : gather_nd_grad
   forward : gather_nd (Tensor x, Tensor index) -> Tensor(out)
   args : (Tensor x, Tensor index, Tensor out_grad)
   output : Tensor(x_grad)
@@ -924,7 +924,7 @@
     func : gather_nd_grad
   no_need_buffer : x
 
-- backward_api : gelu_grad
+- backward_op : gelu_grad
   forward : gelu(Tensor x,  bool approximate) -> Tensor(out)
   args : (Tensor x, Tensor out_grad,  bool approximate)
   output : Tensor(x_grad)
@@ -934,7 +934,7 @@
   kernel :
     func : gelu_grad
 
-- backward_api : graph_send_recv_grad
+- backward_op : graph_send_recv_grad
   forward : graph_send_recv (Tensor x, Tensor src_index, Tensor dst_index, str reduce_op = "SUM", IntArray 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 reduce_op = "SUM")
   output : Tensor(x_grad)
@@ -946,7 +946,7 @@
     data_type : out_grad
   optional: out, dst_count
 
-- backward_api : graph_send_ue_recv_grad
+- backward_op : graph_send_ue_recv_grad
   forward : graph_send_ue_recv (Tensor x, Tensor y, Tensor src_index, Tensor dst_index, str message_op, str reduce_op, IntArray out_size) -> Tensor(out), Tensor(dst_count)
   args : (Tensor x, Tensor y, Tensor src_index, Tensor dst_index, Tensor out, Tensor dst_count, Tensor out_grad, str message_op, str reduce_op)
   output : Tensor(x_grad), Tensor(y_grad)
@@ -958,7 +958,7 @@
     data_type : out_grad
   optional: out, dst_count
 
-- backward_api : grid_sample_grad
+- backward_op : grid_sample_grad
   forward : grid_sample (Tensor x, Tensor grid, str mode, str padding_mode, bool align_corners) -> Tensor(out)
   args : (Tensor x, Tensor grid, Tensor out_grad, str mode, str padding_mode, bool align_corners)
   output : Tensor(x_grad), Tensor(grid_grad)
@@ -969,7 +969,7 @@
     func : grid_sample_grad
     data_type : x
 
-- backward_api : group_norm_grad
+- backward_op : group_norm_grad
   forward : group_norm (Tensor x, Tensor scale, Tensor bias, float epsilon, int groups, str data_layout) -> Tensor(y), Tensor(mean), Tensor(variance)
   args : (Tensor x, Tensor scale, Tensor bias, Tensor y, Tensor mean, Tensor variance, Tensor y_grad, float epsilon, int groups, str data_layout)
   output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
@@ -982,7 +982,7 @@
   optional: scale, bias
   inplace : (y_grad -> x_grad)
 
-- backward_api : gumbel_softmax_grad
+- backward_op : 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)
@@ -992,7 +992,7 @@
   kernel :
     func : gumbel_softmax_grad
 
-- backward_api : hard_shrink_grad
+- backward_op : hard_shrink_grad
   forward : hard_shrink (Tensor x, float threshold) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float threshold)
   output : Tensor(x_grad)
@@ -1003,7 +1003,7 @@
     func : hard_shrink_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : hard_sigmoid_grad
+- backward_op : 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)
@@ -1014,7 +1014,7 @@
     func : hard_sigmoid_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : hard_swish_grad
+- backward_op : 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)
@@ -1025,7 +1025,7 @@
     func : hard_swish_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : hierarchical_sigmoid_grad
+- backward_op : hierarchical_sigmoid_grad
   forward : hierarchical_sigmoid (Tensor x, Tensor w, Tensor label, Tensor path, Tensor code, Tensor bias, int num_classes, bool remote_prefetch, int trainer_id, int64_t[] height_sections, str[] epmap, str[] table_names, bool is_sparse) -> Tensor(out), Tensor(pre_out), Tensor(w_out)
   args : (Tensor x, Tensor w, Tensor label, Tensor path, Tensor code, Tensor bias, Tensor pre_out, Tensor out_grad, int num_classes, bool remote_prefetch, int trainer_id, int64_t[] height_sections, str[] epmap, str[] table_names, bool is_sparse)
   output : Tensor(x_grad), Tensor(w_grad), Tensor(bias_grad)
@@ -1036,7 +1036,7 @@
   kernel :
     func : hierarchical_sigmoid_grad
 
-- backward_api : huber_loss_grad
+- backward_op : 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)
@@ -1046,13 +1046,13 @@
   kernel :
     func : huber_loss_grad
 
-- backward_api : imag_grad
+- backward_op : imag_grad
   forward : imag (Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
   invoke : imag_grad_impl(out_grad, x_grad)
 
-- backward_api : index_add_grad
+- backward_op : index_add_grad
   forward : index_add(Tensor x, Tensor index,  Tensor add_value, int axis) -> Tensor(out)
   args : (Tensor index, Tensor add_value, Tensor out_grad, int axis)
   output : Tensor(x_grad), Tensor(add_value_grad)
@@ -1063,7 +1063,7 @@
     data_type : out_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : index_sample_grad
+- backward_op : index_sample_grad
   forward : index_sample (Tensor x, Tensor index) -> Tensor(out)
   args : (Tensor x, Tensor index, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1075,7 +1075,7 @@
     data_type : out_grad
   no_need_buffer : x
 
-- backward_api : index_select_grad
+- backward_op : 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)
@@ -1087,7 +1087,7 @@
     data_type : x
   no_need_buffer : x
 
-- backward_api : instance_norm_double_grad
+- backward_op : instance_norm_double_grad
   forward : instance_norm_grad(Tensor x, Tensor fwd_scale, Tensor saved_mean, Tensor saved_variance, Tensor grad_y, float epsilon) -> Tensor(grad_x), Tensor(grad_scale), Tensor(grad_bias)
   args : (Tensor x, Tensor fwd_scale, Tensor saved_mean, Tensor saved_variance, Tensor grad_y, Tensor grad_x_grad, Tensor grad_scale_grad, Tensor grad_bias_grad, float epsilon)
   output : Tensor(x_grad), Tensor(fwd_scale_grad), Tensor(grad_y_grad)
@@ -1098,7 +1098,7 @@
     data_type : x
   optional : fwd_scale, grad_x_grad, grad_scale_grad, grad_bias_grad
 
-- backward_api : instance_norm_grad
+- backward_op : instance_norm_grad
   forward : instance_norm(Tensor x, Tensor scale, Tensor bias, float epsilon) -> Tensor(y), Tensor(saved_mean), Tensor(saved_variance)
   args : (Tensor x, Tensor scale, Tensor saved_mean, Tensor saved_variance, Tensor y_grad, float epsilon)
   output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
@@ -1110,7 +1110,7 @@
   optional : scale
   backward : instance_norm_double_grad
 
-- backward_api : inverse_grad
+- backward_op : inverse_grad
   forward : inverse(Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1119,7 +1119,7 @@
   kernel :
     func : inverse_grad
 
-- backward_api : kldiv_loss_grad
+- backward_op : 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)
@@ -1130,7 +1130,7 @@
     func : kldiv_loss_grad
   no_need_buffer : x
 
-- backward_api : kron_grad
+- backward_op : 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)
@@ -1141,7 +1141,7 @@
     func : kron_grad
     data_type : out_grad
 
-- backward_api : kthvalue_grad
+- backward_op : 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)
@@ -1151,7 +1151,7 @@
   kernel :
     func : kthvalue_grad
 
-- backward_api : label_smooth_grad
+- backward_op : 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)
@@ -1161,7 +1161,7 @@
   kernel :
     func : label_smooth_grad
 
-- backward_api : layer_norm_grad
+- backward_op : layer_norm_grad
   forward : layer_norm (Tensor x, Tensor scale, Tensor bias, float epsilon, int begin_norm_axis, bool is_test) -> Tensor(out), Tensor(mean), Tensor(variance)
   args : (Tensor x,  Tensor scale, Tensor bias, Tensor mean, Tensor variance, Tensor out_grad, float epsilon, int begin_norm_axis, bool is_test)
   output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
@@ -1174,7 +1174,7 @@
   no_need_buffer : bias
   optional : scale, bias
 
-- backward_api : leaky_relu_double_grad
+- backward_op : leaky_relu_double_grad
   forward : leaky_relu_grad (Tensor x, Tensor grad_out, float alpha) -> Tensor(grad_x)
   args : (Tensor x, Tensor grad_x_grad, float alpha)
   output : Tensor(grad_out_grad)
@@ -1185,7 +1185,7 @@
     func : leaky_relu_double_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : leaky_relu_grad
+- backward_op : leaky_relu_grad
   forward : leaky_relu (Tensor x, float alpha) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float alpha)
   output : Tensor(x_grad)
@@ -1197,7 +1197,7 @@
   backward : leaky_relu_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : lerp_grad
+- backward_op : 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)
@@ -1207,7 +1207,7 @@
   kernel :
     func : lerp_grad
 
-- backward_api : linear_interp_grad
+- backward_op : linear_interp_grad
   forward : linear_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
   args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
   output : Tensor(x_grad)
@@ -1219,7 +1219,7 @@
     func : linear_interp_grad
     data_type : output_grad
 
-- backward_api : log10_grad
+- backward_op : log10_grad
   forward : log10 (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1230,7 +1230,7 @@
     func : log10_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : log1p_grad
+- backward_op : log1p_grad
   forward : log1p (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1241,7 +1241,7 @@
     func : log1p_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : log2_grad
+- backward_op : log2_grad
   forward : log2 (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1252,7 +1252,7 @@
     func : log2_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : log_double_grad
+- backward_op : log_double_grad
   forward : log_grad (Tensor x, Tensor grad_out) -> Tensor(grad_x)
   args : (Tensor x, Tensor grad_out, Tensor grad_x_grad)
   output : Tensor(x_grad), Tensor(grad_out_grad)
@@ -1263,7 +1263,7 @@
     func : log_double_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : log_grad
+- backward_op : log_grad
   forward : log (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1275,7 +1275,7 @@
   backward : log_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : log_loss_grad
+- backward_op : 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)
@@ -1285,7 +1285,7 @@
   kernel :
     func : log_loss_grad
 
-- backward_api : log_softmax_grad
+- backward_op : log_softmax_grad
   forward : log_softmax(Tensor x,  int axis) -> Tensor(out)
   args : (Tensor out, Tensor out_grad,  int axis)
   output : Tensor(x_grad)
@@ -1295,7 +1295,7 @@
   kernel :
     func : log_softmax_grad
 
-- backward_api : logcumsumexp_grad
+- backward_op : logcumsumexp_grad
   forward : logcumsumexp(Tensor x, int axis, bool flatten, bool exclusive, bool reverse) -> Tensor(out)
   infer_meta :
     func : UnchangedInferMeta
@@ -1305,7 +1305,7 @@
   kernel :
     func : logcumsumexp_grad
 
-- backward_api : logit_grad
+- backward_op : logit_grad
   forward : logit (Tensor x, float eps = 1e-6f) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float eps)
   output : Tensor(x_grad)
@@ -1315,7 +1315,7 @@
   kernel :
     func : logit_grad
 
-- backward_api : logsigmoid_grad
+- backward_op : logsigmoid_grad
   forward : logsigmoid (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1326,7 +1326,7 @@
     func : logsigmoid_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : logsumexp_grad
+- backward_op : 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)
@@ -1336,7 +1336,7 @@
   kernel :
     func : logsumexp_grad
 
-- backward_api : lu_grad
+- backward_op : lu_grad
   forward : lu (Tensor x, bool pivot) -> Tensor(out), Tensor(pivots), Tensor(infos)
   args : (Tensor x, Tensor out, Tensor pivots, Tensor out_grad, bool pivot)
   output : Tensor(x_grad)
@@ -1345,7 +1345,7 @@
   kernel :
     func : lu_grad
 
-- backward_api : lu_unpack_grad
+- backward_op : lu_unpack_grad
   forward : lu_unpack (Tensor x, Tensor pivots, bool unpack_ludata, bool unpack_pivots) -> Tensor(pmat), Tensor(l), Tensor(u)
   args : (Tensor x, Tensor pivots, Tensor l, Tensor u, Tensor pmat, Tensor l_grad, Tensor u_grad, bool unpack_ludata, bool unpack_pivots)
   output : Tensor(x_grad)
@@ -1354,7 +1354,7 @@
   kernel :
     func : lu_unpack_grad
 
-- backward_api : margin_cross_entropy_grad
+- backward_op : margin_cross_entropy_grad
   forward : margin_cross_entropy (Tensor logits, Tensor label, bool return_softmax, int ring_id, int rank, int nranks, float margin1, float margin2, float margin3, float scale) -> Tensor(softmax), Tensor(loss)
   args : (Tensor logits, Tensor label, Tensor softmax, Tensor loss_grad, bool return_softmax, int ring_id, int rank, int nranks, float margin1, float margin2, float margin3, float scale)
   output : Tensor(logits_grad)
@@ -1365,7 +1365,7 @@
     data_type : softmax
   inplace : (softmax -> logits_grad)
 
-- backward_api : masked_select_grad
+- backward_op : masked_select_grad
   forward : masked_select (Tensor x, Tensor mask) -> Tensor(out)
   args : (Tensor x, Tensor mask, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1377,7 +1377,7 @@
     data_type : x
   no_need_buffer : x
 
-- backward_api : matmul_double_grad
+- backward_op : 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)
@@ -1389,7 +1389,7 @@
   backward : matmul_triple_grad
   optional : grad_x_grad, grad_y_grad
 
-- backward_api : matmul_grad
+- backward_op : 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)
@@ -1400,7 +1400,7 @@
     func : matmul_grad
   backward : matmul_double_grad
 
-- backward_api : matmul_triple_grad
+- backward_op : 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)
@@ -1411,7 +1411,7 @@
     func : matmul_triple_grad
   optional : grad_x_grad, grad_y_grad, grad_grad_out_grad
 
-- backward_api : matrix_power_grad
+- backward_op : 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)
@@ -1421,7 +1421,7 @@
   kernel :
     func : matrix_power_grad
 
-- backward_api : max_grad
+- backward_op : max_grad
   forward: max (Tensor x,  IntArray dims={},  bool keep_dim=false) -> Tensor(out)
   args : (Tensor x, Tensor out, Tensor out_grad, IntArray dims={}, bool keep_dim=false, bool reduce_all=false)
   output : Tensor(x_grad)
@@ -1431,7 +1431,7 @@
   kernel :
     func : max_grad
 
-- backward_api : max_pool2d_with_index_grad
+- backward_op : 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)
@@ -1440,7 +1440,7 @@
   kernel :
     func : max_pool2d_with_index_grad
 
-- backward_api : max_pool3d_with_index_grad
+- backward_op : 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)
@@ -1449,7 +1449,7 @@
   kernel :
     func : max_pool3d_with_index_grad
 
-- backward_api : maximum_grad
+- backward_op : 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)
@@ -1459,7 +1459,7 @@
   kernel :
     func : maximum_grad
 
-- backward_api : maxout_grad
+- backward_op : maxout_grad
   forward : maxout(Tensor x, int groups, int axis) -> Tensor(out)
   args : (Tensor x, Tensor out, Tensor out_grad, int groups, int axis)
   output : Tensor(x_grad)
@@ -1469,7 +1469,7 @@
   kernel :
     func : maxout_grad
 
-- backward_api : mean_all_grad
+- backward_op : mean_all_grad
   forward : mean_all(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1479,13 +1479,13 @@
   kernel :
     func : mean_all_grad
 
-- backward_api : mean_double_grad
+- backward_op : mean_double_grad
   forward: mean_grad (Tensor x, Tensor grad_out, IntArray dims={},  bool keep_dim=false, bool reduce_all = false) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, IntArray dims={},  bool keep_dim=false)
   output : Tensor(grad_out_grad)
   invoke : mean(grad_x_grad, dims, keep_dim)
 
-- backward_api : mean_grad
+- backward_op : mean_grad
   forward: mean (Tensor x,  IntArray dims={},  bool keep_dim=false) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, IntArray dims={},  bool keep_dim=false, bool reduce_all=false)
   output : Tensor(x_grad)
@@ -1497,7 +1497,7 @@
   backward : mean_double_grad
   no_need_buffer : x
 
-- backward_api : meshgrid_grad
+- backward_op : meshgrid_grad
   forward : meshgrid (Tensor[] inputs) -> Tensor[](outputs)
   args : (Tensor[] inputs, Tensor[] outputs_grad)
   output : Tensor[](inputs_grad){inputs.size()}
@@ -1506,7 +1506,7 @@
   kernel :
     func : meshgrid_grad
 
-- backward_api : min_grad
+- backward_op : min_grad
   forward: min (Tensor x,  IntArray dims={},  bool keep_dim=false) -> Tensor(out)
   args : (Tensor x, Tensor out, Tensor out_grad, IntArray dims={}, bool keep_dim=false, bool reduce_all=false)
   output : Tensor(x_grad)
@@ -1516,7 +1516,7 @@
   kernel :
     func : min_grad
 
-- backward_api : minimum_grad
+- backward_op : 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)
@@ -1526,7 +1526,7 @@
   kernel :
     func : minimum_grad
 
-- backward_api : mish_grad
+- backward_op : mish_grad
   forward : mish (Tensor x, float threshold) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float threshold)
   output : Tensor(x_grad)
@@ -1537,7 +1537,7 @@
     func : mish_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : mode_grad
+- backward_op : 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)
@@ -1547,7 +1547,7 @@
   kernel :
     func : mode_grad
 
-- backward_api : multi_dot_grad
+- backward_op : multi_dot_grad
   forward : multi_dot (Tensor[] x) -> Tensor(out)
   args : (Tensor[] x, Tensor out_grad)
   output : Tensor[](x_grad) {x.size()}
@@ -1556,7 +1556,7 @@
   kernel :
     func : multi_dot_grad
 
-- backward_api : multiplex_grad
+- backward_op : multiplex_grad
   forward : multiplex (Tensor[] ins, Tensor ids) -> Tensor(out)
   args : (Tensor[] ins, Tensor ids, Tensor out_grad)
   output : Tensor[](ins_grad){ins.size()}
@@ -1567,7 +1567,7 @@
     func : multiplex_grad
     param : [ids, out_grad]
 
-- backward_api : multiply_double_grad
+- backward_op : multiply_double_grad
   forward : multiply_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
   args : (Tensor x, Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
   output : Tensor(x_grad), Tensor(y_grad), Tensor(grad_out_grad)
@@ -1580,7 +1580,7 @@
   backward : multiply_triple_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : multiply_grad
+- backward_op : 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)
@@ -1591,7 +1591,7 @@
     func : multiply_grad
   backward : multiply_double_grad
 
-- backward_api : multiply_triple_grad
+- backward_op : multiply_triple_grad
   forward : multiply_double_grad (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, int aixs = -1) -> 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, int axis = -1)
   output : Tensor(x_grad), Tensor(y_grad), Tensor(fwd_grad_out_grad), Tensor(fwd_grad_grad_x_grad), Tensor(fwd_grad_grad_y_grad)
@@ -1602,7 +1602,7 @@
     func : multiply_triple_grad
   optional : fwd_grad_grad_x, fwd_grad_grad_y, grad_grad_out_grad
 
-- backward_api : nearest_interp_grad
+- backward_op : nearest_interp_grad
   forward : nearest_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
   args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
   output : Tensor(x_grad)
@@ -1614,7 +1614,7 @@
     func : nearest_interp_grad
     data_type : output_grad
 
-- backward_api : nll_loss_grad
+- backward_op : 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)
@@ -1625,7 +1625,7 @@
     data_type : input
   optional : weight
 
-- backward_api : norm_grad
+- backward_op : norm_grad
   forward : norm (Tensor x, int axis, float epsilon, bool is_test) -> Tensor(out), Tensor(norm)
   args : (Tensor x, Tensor norm, Tensor out_grad, int axis, float epsilon, bool is_test)
   output : Tensor(x_grad)
@@ -1635,7 +1635,7 @@
   kernel :
     func : norm_grad
 
-- backward_api : overlap_add_grad
+- backward_op : overlap_add_grad
   forward : overlap_add(Tensor x, int hop_length, int axis) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, int hop_length, int axis)
   output : Tensor(x_grad)
@@ -1645,7 +1645,7 @@
     func : overlap_add_grad
     data_type : x
 
-- backward_api : p_norm_grad
+- backward_op : 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)
@@ -1655,7 +1655,7 @@
   kernel :
     func : p_norm_grad
 
-- backward_api : pad3d_double_grad
+- backward_op : pad3d_double_grad
   forward : pad3d_grad(Tensor x, Tensor grad_out, IntArray paddings, str mode, float pad_value, str data_format) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, IntArray paddings, str mode, float pad_value, str data_format)
   output : Tensor(grad_out_grad)
@@ -1664,7 +1664,7 @@
   kernel :
     func : pad3d
 
-- backward_api : pad3d_grad
+- backward_op : 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)
@@ -1676,7 +1676,7 @@
   no_need_buffer : x
   backward : pad3d_double_grad
 
-- backward_api : pad_double_grad
+- backward_op : pad_double_grad
   forward : pad_grad(Tensor x, Tensor grad_out, int[] paddings, Scalar pad_value) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, int[] paddings, Scalar pad_value)
   output : Tensor(grad_out_grad)
@@ -1685,7 +1685,7 @@
   kernel :
     func : pad
 
-- backward_api : pad_grad
+- backward_op : pad_grad
   forward : pad(Tensor x, int[] paddings, Scalar pad_value) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, int[] paddings, Scalar pad_value)
   output : Tensor(x_grad)
@@ -1698,7 +1698,7 @@
   no_need_buffer : x
   backward : pad_double_grad
 
-- backward_api : pixel_shuffle_grad
+- backward_op : pixel_shuffle_grad
   forward : pixel_shuffle (Tensor x, int upscale_factor, str data_format) -> Tensor(out)
   args : (Tensor out_grad, int upscale_factor, str data_format)
   output : Tensor(x_grad)
@@ -1707,7 +1707,7 @@
   kernel :
     func : pixel_shuffle_grad
 
-- backward_api : pool2d_double_grad
+- backward_op : pool2d_double_grad
   forward : pool2d_grad(Tensor x, Tensor out, Tensor grad_out, IntArray 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, bool use_gpudnn) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, IntArray 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, bool use_gpudnn)
   output : Tensor(grad_out_grad)
@@ -1719,7 +1719,7 @@
     param : [grad_x_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
     use_gpudnn : use_gpudnn
 
-- backward_api : pool2d_grad
+- backward_op : pool2d_grad
   forward : pool2d(Tensor x, IntArray 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, bool use_gpudnn) -> Tensor(out)
   args : (Tensor x, Tensor out, Tensor out_grad, IntArray 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, bool use_gpudnn)
   output : Tensor(x_grad)
@@ -1732,7 +1732,7 @@
     use_gpudnn : use_gpudnn
   backward : pool2d_double_grad
 
-- backward_api : pool3d_grad
+- backward_op : 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, bool use_gpudnn) -> 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, bool use_gpudnn)
   output : Tensor(x_grad)
@@ -1744,7 +1744,7 @@
     param : [x, out, out_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
     use_gpudnn : use_gpudnn
 
-- backward_api : pow_grad
+- backward_op : pow_grad
   forward : pow(Tensor x, Scalar s) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, Scalar s=-1)
   output : Tensor(x_grad)
@@ -1755,7 +1755,7 @@
     func : pow_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : prelu_grad
+- backward_op : 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)
@@ -1765,7 +1765,7 @@
   kernel :
     func : prelu_grad
 
-- backward_api : psroi_pool_grad
+- backward_op : 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)
@@ -1778,7 +1778,7 @@
   optional : boxes_num
 
 # output is optional
-- backward_api : put_along_axis_grad
+- backward_op : 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)
@@ -1788,7 +1788,7 @@
   kernel :
     func : put_along_axis_grad
 
-- backward_api : qr_grad
+- backward_op : qr_grad
   forward : qr (Tensor x, str mode) -> Tensor(q), Tensor(r)
   args : (Tensor x, Tensor q, Tensor r, Tensor q_grad, Tensor r_grad, str mode)
   output : Tensor(x_grad)
@@ -1798,13 +1798,13 @@
   kernel :
     func : qr_grad
 
-- backward_api : real_grad
+- backward_op : real_grad
   forward : real (Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
   invoke : real_grad_impl(out_grad, x_grad)
 
-- backward_api : reciprocal_grad
+- backward_op : reciprocal_grad
   forward : reciprocal (Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1815,7 +1815,7 @@
     func : reciprocal_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : reduce_prod_grad
+- backward_op : reduce_prod_grad
   forward : reduce_prod (Tensor x, IntArray dims, bool keep_dim, bool reduce_all) -> Tensor(out)
   args : (Tensor x, Tensor out, Tensor out_grad, IntArray dims,  bool keep_dim, bool reduce_all)
   output : Tensor(x_grad)
@@ -1825,7 +1825,7 @@
   kernel :
     func : prod_grad
 
-- backward_api : relu6_grad
+- backward_op : relu6_grad
   forward : relu6 (Tensor x, float threshold) -> Tensor(out)
   args : (Tensor out, Tensor out_grad, float threshold)
   output : Tensor(x_grad)
@@ -1836,7 +1836,7 @@
     func : relu6_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : relu_double_grad
+- backward_op : relu_double_grad
   forward : relu_grad (Tensor out, Tensor grad_out) -> Tensor(grad_x)
   args : (Tensor out, Tensor grad_x_grad)
   output : Tensor(grad_out_grad)
@@ -1847,7 +1847,7 @@
     func : relu_double_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : relu_grad
+- backward_op : relu_grad
   forward : relu (Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1859,7 +1859,7 @@
   backward: relu_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : renorm_grad
+- backward_op : renorm_grad
   forward : renorm (Tensor x, float p, int axis, float max_norm) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float p, int axis, float max_norm)
   output : Tensor(x_grad)
@@ -1869,7 +1869,7 @@
   kernel :
     func : renorm_grad
 
-- backward_api : repeat_interleave_grad
+- backward_op : repeat_interleave_grad
   forward : repeat_interleave(Tensor x, int repeats, int dim) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, int repeats, int dim)
   output : Tensor(x_grad)
@@ -1879,7 +1879,7 @@
   kernel :
     func : repeat_interleave_grad
 
-- backward_api : repeat_interleave_with_tensor_index_grad
+- backward_op : repeat_interleave_with_tensor_index_grad
   forward : repeat_interleave_with_tensor_index(Tensor x, Tensor repeats, int dim) -> Tensor(out)
   args : (Tensor x, Tensor repeats, Tensor out_grad, int dim)
   output : Tensor(x_grad)
@@ -1890,7 +1890,7 @@
     func : repeat_interleave_with_tensor_index_grad
     data_type : x
 
-- backward_api : reshape_double_grad
+- backward_op : reshape_double_grad
   forward : reshape_grad (Tensor xshape, Tensor grad_out) -> Tensor(grad_x)
   args : (Tensor grad_out, Tensor grad_x_grad)
   output : Tensor(grad_out_grad)
@@ -1902,7 +1902,7 @@
   no_need_buffer : grad_out
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : reshape_grad
+- backward_op : reshape_grad
   forward : reshape (Tensor x, IntArray shape) -> Tensor(out), Tensor(xshape)
   args : (Tensor xshape, Tensor out_grad)
   output : Tensor(x_grad)
@@ -1918,7 +1918,7 @@
   backward : reshape_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : reverse_array_grad
+- backward_op : reverse_array_grad
   forward : reverse_array (Tensor[] x, IntArray axis) -> Tensor[](out)
   args : (Tensor[] out_grad, IntArray axis)
   output : Tensor[](x_grad){out_grad.size()}
@@ -1927,13 +1927,13 @@
   kernel :
     func : reverse
 
-- backward_api : reverse_grad
+- backward_op : reverse_grad
   forward : reverse (Tensor x, IntArray axis) -> Tensor(out)
   args : (Tensor out_grad, IntArray axis)
   output : Tensor(x_grad)
   invoke : reverse(out_grad, axis)
 
-- backward_api : roi_align_grad
+- backward_op : 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)
@@ -1946,7 +1946,7 @@
   no_need_buffer : x
   optional : boxes_num
 
-- backward_api : roi_pool_grad
+- backward_op : 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)
@@ -1958,7 +1958,7 @@
     data_type : x
   optional : boxes_num
 
-- backward_api : roll_grad
+- backward_op : 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)
@@ -1970,7 +1970,7 @@
     data_type : x
   no_need_buffer : x
 
-- backward_api : round_grad
+- backward_op : round_grad
   forward : round(Tensor x) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
@@ -1981,7 +1981,7 @@
     func : round_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : rsqrt_double_grad
+- backward_op : rsqrt_double_grad
   forward : rsqrt_grad (Tensor out, Tensor grad_out) -> Tensor(grad_x)
   args : (Tensor out, Tensor grad_x, Tensor grad_x_grad)
   output : Tensor(out_grad), Tensor(grad_out_grad)
@@ -1992,7 +1992,7 @@
     func : rsqrt_double_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : rsqrt_grad
+- backward_op : rsqrt_grad
   forward : rsqrt (Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2004,13 +2004,13 @@
   backward : rsqrt_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : scale_grad
+- backward_op : scale_grad
   forward : scale (Tensor x, Scalar scale, float bias, bool bias_after_scale) -> Tensor(out)
   args : (Tensor out_grad, Scalar scale=1.0, bool bias_after_scale=true)
   output : Tensor(x_grad)
   invoke : scale(out_grad, scale, 0.0, bias_after_scale)
 
-- backward_api : scatter_grad
+- backward_op : 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)
@@ -2021,7 +2021,7 @@
     func : scatter_grad
   no_need_buffer : updates
 
-- backward_api : scatter_nd_add_grad
+- backward_op : 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)
@@ -2032,7 +2032,7 @@
     func : scatter_nd_add_grad
   no_need_buffer : updates
 
-- backward_api : segment_pool_grad
+- backward_op : 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)
@@ -2044,7 +2044,7 @@
     data_type : x
   optional : summed_ids
 
-- backward_api : selu_grad
+- backward_op : 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)
@@ -2054,7 +2054,7 @@
   kernel :
     func : selu_grad
 
-- backward_api : sigmoid_cross_entropy_with_logits_grad
+- backward_op : 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)
@@ -2065,7 +2065,7 @@
     func : sigmoid_cross_entropy_with_logits_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : sigmoid_double_grad
+- backward_op : 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)
@@ -2077,7 +2077,7 @@
   backward : sigmoid_triple_grad
   inplace : (grad_x_grad -> fwd_grad_out_grad)
 
-- backward_api : sigmoid_grad
+- backward_op : sigmoid_grad
   forward : sigmoid (Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2089,7 +2089,7 @@
   backward : sigmoid_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : sigmoid_triple_grad
+- backward_op : 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)
@@ -2101,7 +2101,7 @@
   optional : grad_grad_out_grad
   inplace : (grad_grad_x -> fwd_grad_out_grad)
 
-- backward_api : silu_grad
+- backward_op : silu_grad
   forward : silu (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2112,7 +2112,7 @@
     func : silu_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : sin_grad
+- backward_op : sin_grad
   forward : sin (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2123,7 +2123,7 @@
     func : sin_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : sinh_grad
+- backward_op : sinh_grad
   forward : sinh (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2134,13 +2134,13 @@
     func : sinh_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : slice_double_grad
+- backward_op : slice_double_grad
   forward : slice_grad (Tensor input, Tensor grad_out, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis) -> Tensor(grad_input)
   args : (Tensor grad_input_grad, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)
   output : Tensor(grad_out_grad)
   invoke : slice(grad_input_grad, axes, starts, ends, infer_flags, decrease_axis)
 
-- backward_api : slice_grad
+- backward_op : 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)
@@ -2152,7 +2152,7 @@
   backward : slice_double_grad
   no_need_buffer : input
 
-- backward_api : slogdet_grad
+- backward_op : slogdet_grad
   forward : slogdet (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2162,7 +2162,7 @@
   kernel :
     func : slogdeterminant_grad
 
-- backward_api : soft_shrink_grad
+- backward_op : soft_shrink_grad
   forward : soft_shrink (Tensor x, float lambda) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float lambda)
   output : Tensor(x_grad)
@@ -2173,7 +2173,7 @@
     func : soft_shrink_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : softmax_grad
+- backward_op : softmax_grad
   forward : softmax (Tensor x, int axis) -> Tensor(out)
   args : (Tensor out, Tensor out_grad, int axis)
   output : Tensor(x_grad)
@@ -2184,7 +2184,7 @@
     func : softmax_grad
     use_gpudnn : true
 
-- backward_api : softplus_grad
+- backward_op : softplus_grad
   forward : softplus (Tensor x, float beta, float threshold) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float beta, float threshold)
   output : Tensor(x_grad)
@@ -2195,7 +2195,7 @@
     func : softplus_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : softsign_grad
+- backward_op : softsign_grad
   forward : softsign (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2206,7 +2206,7 @@
     func : softsign_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : spectral_norm_grad
+- backward_op : spectral_norm_grad
   forward : spectral_norm (Tensor weight, Tensor u, Tensor v, int dim, int power_iters, float eps) -> Tensor(out)
   args : (Tensor weight, Tensor u, Tensor v, Tensor out_grad, int dim, int power_iters, float eps)
   output : Tensor(weight_grad)
@@ -2216,20 +2216,20 @@
     func : spectral_norm_grad
     data_type : out_grad
 
-- backward_api : split_grad
+- backward_op : split_grad
   forward : split (Tensor x, IntArray num_or_sections, Scalar axis) -> Tensor[](out)
   args : (Tensor[] out_grad, Scalar axis = -1)
   output : Tensor(x_grad)
   invoke : concat( out_grad, axis)
 
-- backward_api : split_with_num_grad
+- backward_op : split_with_num_grad
   forward : split_with_num (Tensor x, int num, Scalar axis) -> Tensor[](out)
   args : (Tensor[] out_grad, Scalar axis = -1)
   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_double_grad
+- backward_op : sqrt_double_grad
   forward : sqrt_grad (Tensor out, Tensor grad_out) -> Tensor(grad_x)
   args : (Tensor out, Tensor grad_x, Tensor grad_x_grad)
   output : Tensor(out_grad), Tensor(grad_out_grad)
@@ -2240,7 +2240,7 @@
     func : sqrt_double_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : sqrt_grad
+- backward_op : sqrt_grad
   forward : sqrt (Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2252,7 +2252,7 @@
   backward : sqrt_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : square_double_grad
+- backward_op : square_double_grad
   forward : square_grad (Tensor x, Tensor grad_out) -> Tensor(grad_x)
   args : (Tensor x, Tensor grad_out, Tensor grad_x_grad)
   output : Tensor(x_grad), Tensor(grad_out_grad)
@@ -2263,7 +2263,7 @@
     func : square_double_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : square_grad
+- backward_op : square_grad
   forward : square (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2275,7 +2275,7 @@
   backward : square_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : squared_l2_norm_grad
+- backward_op : squared_l2_norm_grad
   forward : squared_l2_norm(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2285,13 +2285,13 @@
   kernel :
     func : squared_l2_norm_grad
 
-- backward_api : squeeze_double_grad
+- backward_op : squeeze_double_grad
   forward : squeeze_grad(Tensor xshape, Tensor grad_out, IntArray axes) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, IntArray axes)
   output : Tensor(grad_out_grad)
   invoke: squeeze(grad_x_grad, axes)
 
-- backward_api : squeeze_grad
+- backward_op : squeeze_grad
   forward : squeeze(Tensor x, IntArray axes) -> Tensor(out), Tensor(xshape)
   args : (Tensor xshape, Tensor out_grad, IntArray axes)
   output : Tensor(x_grad)
@@ -2303,7 +2303,7 @@
   inplace : (out_grad -> x_grad)
   backward: squeeze_double_grad
 
-- backward_api : stack_grad
+- backward_op : stack_grad
   forward : stack (Tensor[] x, int axis) -> Tensor(out)
   args : (Tensor[] x, Tensor out_grad, int axis)
   output : Tensor[](x_grad){x.size()}
@@ -2315,7 +2315,7 @@
     param : [out_grad, axis]
   no_need_buffer : x
 
-- backward_api : strided_slice_grad
+- backward_op : 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)
@@ -2326,7 +2326,7 @@
     func : strided_slice_grad
   no_need_buffer : x
 
-- backward_api : subtract_double_grad
+- backward_op : subtract_double_grad
   forward : subtract_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
   args : (Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
   output : Tensor(grad_out_grad)
@@ -2339,7 +2339,7 @@
   no_need_buffer : y, grad_out
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : subtract_grad
+- backward_op : 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)
@@ -2352,13 +2352,13 @@
   backward : subtract_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : sum_double_grad
+- backward_op : sum_double_grad
   forward : sum_grad (Tensor x, Tensor grad_out, IntArray dims, bool keep_dim, bool reduce_all=false) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, IntArray dims={}, bool keep_dim=false)
   output : Tensor(grad_out_grad)
   invoke : sum(grad_x_grad, dims, grad_x_grad.dtype(), keep_dim)
 
-- backward_api : sum_grad
+- backward_op : sum_grad
   forward : sum (Tensor x, IntArray dims={}, DataType out_dtype=DataType::UNDEFINED, bool keep_dim=false) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, IntArray dims, bool keep_dim, bool reduce_all=false)
   output : Tensor(x_grad)
@@ -2370,7 +2370,7 @@
   no_need_buffer : x
   backward : sum_double_grad
 
-- backward_api : svd_grad
+- backward_op : svd_grad
   forward : svd (Tensor x, bool full) -> Tensor(u), Tensor(s), Tensor(vh)
   args : (Tensor x, Tensor u, Tensor vh, Tensor s, Tensor u_grad, Tensor vh_grad, Tensor s_grad, bool full)
   output : Tensor(x_grad)
@@ -2381,7 +2381,7 @@
     func : svd_grad
   optional: u_grad, vh_grad, s_grad
 
-- backward_api : swish_grad
+- backward_op : swish_grad
   forward : swish (Tensor x, float beta=1.0) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float bete=1.0)
   output : Tensor(x_grad)
@@ -2392,7 +2392,7 @@
     func : swish_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : sync_batch_norm_grad
+- backward_op : sync_batch_norm_grad
   forward : sync_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 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)
@@ -2404,7 +2404,7 @@
     data_type : out_grad
   optional : reserve_space
 
-- backward_api : take_along_axis_grad
+- backward_op : 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)
@@ -2414,7 +2414,7 @@
   kernel :
     func : take_along_axis_grad
 
-- backward_api : tan_grad
+- backward_op : tan_grad
   forward : tan (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2425,7 +2425,7 @@
     func : tan_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : tanh_double_grad
+- backward_op : tanh_double_grad
   forward : tanh_grad (Tensor out, Tensor grad_out) -> Tensor(grad_x)
   args : (Tensor out, Tensor grad_out, Tensor grad_x_grad)
   output : Tensor(out_grad), Tensor(grad_out_grad)
@@ -2437,7 +2437,7 @@
   backward : tanh_triple_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
-- backward_api : tanh_grad
+- backward_op : tanh_grad
   forward : tanh (Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2449,7 +2449,7 @@
   backward : tanh_double_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : tanh_shrink_grad
+- backward_op : tanh_shrink_grad
   forward : tanh_shrink (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -2460,7 +2460,7 @@
     func : tanh_shrink_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : tanh_triple_grad
+- backward_op : tanh_triple_grad
   forward : tanh_double_grad (Tensor out, Tensor grad_out_forward, Tensor grad_x_grad_forward) -> Tensor(grad_out_new), Tensor(grad_out_grad)
   args : (Tensor out, Tensor grad_out_forward, Tensor grad_x_grad_forward, Tensor grad_out_new_grad, Tensor grad_out_grad_grad)
   output : Tensor(out_grad), Tensor(grad_out_forward_grad), Tensor(grad_x_grad_forward_grad)
@@ -2471,7 +2471,7 @@
     func : tanh_triple_grad
   inplace : (grad_x_grad_forward -> grad_out_forward_grad)
 
-- backward_api : temporal_shift_grad
+- backward_op : temporal_shift_grad
   forward : temporal_shift(Tensor x, int seg_num, float shift_ratio, str data_format_str) -> Tensor(out)
   args : (Tensor out_grad, int seg_num, float shift_ratio, str data_format_str)
   output : Tensor(x_grad)
@@ -2481,7 +2481,7 @@
   kernel :
     func : temporal_shift_grad
 
-- backward_api : thresholded_relu_grad
+- backward_op : thresholded_relu_grad
   forward : thresholded_relu (Tensor x, float threshold) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float threshold)
   output : Tensor(x_grad)
@@ -2492,13 +2492,13 @@
     func : thresholded_relu_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : tile_double_grad
+- backward_op : tile_double_grad
   forward : tile_grad (Tensor x, Tensor grad_out, IntArray repeat_times) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, IntArray repeat_times)
   output : Tensor(grad_out_grad)
   invoke : tile(grad_x_grad, repeat_times)
 
-- backward_api : tile_grad
+- backward_op : tile_grad
   forward : tile (Tensor x, IntArray repeat_times) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, IntArray repeat_times)
   output : Tensor(x_grad)
@@ -2510,7 +2510,7 @@
   no_need_buffer : x
   backward : tile_double_grad
 
-- backward_api : top_k_grad
+- backward_op : 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)
@@ -2520,13 +2520,13 @@
   kernel :
     func : top_k_grad
 
-- backward_api : transpose_double_grad
+- backward_op : transpose_double_grad
   forward : transpose_grad (Tensor grad_out, int[] axis) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, int[] axis)
   output : Tensor(grad_out_grad)
   invoke : transpose(grad_x_grad, axis)
 
-- backward_api : transpose_grad
+- backward_op : transpose_grad
   forward : transpose (Tensor x, int[] axis) -> Tensor(out)
   args : (Tensor out_grad, int[] axis)
   output : Tensor(x_grad)
@@ -2537,7 +2537,7 @@
     func : transpose_grad
   backward : transpose_double_grad
 
-- backward_api : triangular_solve_grad
+- backward_op : 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)
@@ -2547,7 +2547,7 @@
   kernel :
     func : triangular_solve_grad
 
-- backward_api : tril_triu_grad
+- backward_op : 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)
@@ -2557,7 +2557,7 @@
   kernel :
     func : tril_triu_grad
 
-- backward_api : trilinear_interp_grad
+- backward_op : trilinear_interp_grad
   forward : trilinear_interp (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode) -> Tensor(output)
   args : (Tensor x, Tensor out_size, Tensor[] size_tensor, Tensor scale_tensor, Tensor output_grad, str data_layout, int out_d, int out_h, int out_w, float[] scale, str interp_method, bool align_corners, int align_mode)
   output : Tensor(x_grad)
@@ -2569,13 +2569,13 @@
     func : trilinear_interp_grad
     data_type : output_grad
 
-- backward_api : unbind_grad
+- backward_op : 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
+- backward_op : 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)
@@ -2586,7 +2586,7 @@
     func : unfold_grad
   no_need_buffer : x
 
-- backward_api : uniform_random_inplace_grad
+- backward_op : uniform_random_inplace_grad
   forward : uniform_random_inplace(Tensor x, float min, float max, int seed, int diag_num, int diag_step, float diag_val) -> Tensor(out)
   args : (Tensor out_grad, float min, float max, int seed, int diag_num, int diag_step, float diag_val)
   output : Tensor(x_grad)
@@ -2596,13 +2596,13 @@
     func : uniform_random_inplace_grad
   inplace : (out_grad -> x_grad)
 
-- backward_api : unsqueeze_double_grad
+- backward_op : unsqueeze_double_grad
   forward : unsqueeze_grad(Tensor xshape, Tensor grad_out, IntArray axes) -> Tensor(grad_x)
   args : (Tensor grad_x_grad, IntArray axes)
   output : Tensor(grad_out_grad)
   invoke : unsqueeze(grad_x_grad, axes)
 
-- backward_api : unsqueeze_grad
+- backward_op : unsqueeze_grad
   forward : unsqueeze(Tensor x, IntArray axes) -> Tensor(out), Tensor(xshape)
   args : (Tensor xshape, Tensor out_grad, IntArray axes)
   output : Tensor(x_grad)
@@ -2615,7 +2615,7 @@
   inplace : (out_grad -> x_grad)
   backward : unsqueeze_double_grad
 
-- backward_api : unstack_grad
+- backward_op : unstack_grad
   forward : unstack (Tensor x, int axis, int num) -> Tensor[](out)
   args : (Tensor[] out_grad, int axis)
   output : Tensor(x_grad)
@@ -2625,7 +2625,7 @@
   kernel :
     func : unstack_grad
 
-- backward_api : warpctc_grad
+- backward_op : warpctc_grad
   forward : warpctc (Tensor logits, Tensor label, Tensor logits_length, Tensor labels_length, int blank, bool norm_by_times) -> Tensor(loss), Tensor(warpctcgrad)
   args : (Tensor logits, Tensor logits_length, Tensor warpctcgrad, Tensor loss_grad, int blank, bool norm_by_times)
   output : Tensor(logits_grad)
@@ -2637,7 +2637,7 @@
   optional : logits_length
   no_need_buffer : logits
 
-- backward_api : where_grad
+- backward_op : 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)
@@ -2648,7 +2648,7 @@
     func : where_grad
   no_need_buffer : x, y
 
-- backward_api : yolov3_loss_grad
+- backward_op : yolov3_loss_grad
   forward : yolov3_loss(Tensor x, Tensor gt_box, Tensor gt_label, Tensor gt_score, int[] anchors, int[] anchor_mask, int class_num, float ignore_thresh, int downsample_ratio, bool use_label_smooth=true, float scale_x_y=1.0) -> Tensor(loss), Tensor(objectness_mask), Tensor(gt_match_mask)
   args : (Tensor x, Tensor gt_box, Tensor gt_label, Tensor gt_score, Tensor objectness_mask, Tensor gt_match_mask, Tensor loss_grad, int[] anchors, int[] anchor_mask, int class_num, float ignore_thresh, int downsample_ratio, bool use_label_smooth=true, float scale_x_y=1.0)
   output : Tensor(x_grad), Tensor(gt_box_grad), Tensor(gt_label_grad), Tensor(gt_score_grad)
@@ -2658,7 +2658,7 @@
     func : yolov3_loss_grad
   optional : gt_score
 
-- backward_api: fold_grad
+- backward_op: fold_grad
   forward: fold (Tensor x, int[] output_sizes, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations) -> Tensor(out)
   args: (Tensor x, Tensor out_grad, int[] output_sizes, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations)
   output: Tensor(x_grad)
@@ -2669,7 +2669,7 @@
     func: fold_grad
   no_need_buffer : x
 
-- backward_api: unpool3d_grad
+- backward_op: unpool3d_grad
   forward: unpool3d (Tensor x, Tensor indices, int[] ksize, int[] strides, int[] padding, int[] output_size, str data_format) -> Tensor(out)
   args: (Tensor x, Tensor indices, Tensor out, Tensor out_grad, int[] ksize, int[] strides, int[] padding, int[] output_size, str data_format)
   output: Tensor(x_grad)
@@ -2680,7 +2680,7 @@
     func: unpool3d_grad
     data_type: x
 
-- backward_api: unpool_grad
+- backward_op: unpool_grad
   forward: unpool (Tensor x, Tensor indices, int[] ksize, int[] strides, int[] padding,  IntArray output_size, str data_format) -> Tensor(out)
   args: (Tensor x, Tensor indices, Tensor out, Tensor out_grad, int[] ksize, int[] strides, int[] padding, IntArray output_size, str data_format)
   output: Tensor(x_grad)
diff --git a/paddle/phi/api/yaml/sparse_backward.yaml b/paddle/phi/api/yaml/sparse_backward.yaml
index e6242f178e540..4bc306388d128 100644
--- a/paddle/phi/api/yaml/sparse_backward.yaml
+++ b/paddle/phi/api/yaml/sparse_backward.yaml
@@ -1,4 +1,4 @@
-- backward_api : abs_grad
+- backward_op : abs_grad
   forward : tanh(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -6,7 +6,7 @@
     func : abs_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            abs_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : acos_grad
+- backward_op : acos_grad
   forward : acos(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -14,7 +14,7 @@
     func : acos_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            acos_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : acosh_grad
+- backward_op : acosh_grad
   forward : acosh(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -22,7 +22,7 @@
     func : acosh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            acosh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : add_grad
+- backward_op : 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)
@@ -30,7 +30,7 @@
     func : add_coo_coo_grad{sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo},
            add_csr_csr_grad{sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr}
 
-- backward_api : addmm_grad
+- backward_op : addmm_grad
   forward : addmm(Tensor input, Tensor x, Tensor y, float alpha=1.0, float beta=1.0) -> Tensor(out)
   args : (Tensor input, Tensor x, Tensor y, Tensor out_grad, float alpha=1.0, float beta=1.0)
   output : Tensor(input_grad), Tensor(x_grad), Tensor(y_grad)
@@ -40,7 +40,7 @@
            addmm_coo_dense_grad {dense, sparse_coo, dense, dense -> dense, sparse_coo, dense},
            addmm_coo_coo_grad {sparse_coo, sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo, sparse_coo}
 
-- backward_api : asin_grad
+- backward_op : asin_grad
   forward : asin(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -48,7 +48,7 @@
     func : asin_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            asin_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : asinh_grad
+- backward_op : asinh_grad
   forward : asinh(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -56,7 +56,7 @@
     func : asinh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            asinh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : atan_grad
+- backward_op : atan_grad
   forward : atan(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -64,7 +64,7 @@
     func : atan_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            atan_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : atanh_grad
+- backward_op : atanh_grad
   forward : atanh(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -72,7 +72,7 @@
     func : atanh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            atanh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : cast_grad
+- backward_op : cast_grad
   forward : cast(Tensor x, DataType index_dtype, DataType value_dtype) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, DataType value_dtype)
   output : Tensor(x_grad)
@@ -81,14 +81,14 @@
            cast_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
     data_type : out_grad
 
-- backward_api : conv3d_coo_grad
+- backward_op : conv3d_coo_grad
   forward : conv3d_coo (Tensor x, Tensor kernel, int[] paddings, int[] dilations, int[] strides, int groups, bool subm, str key) -> Tensor(out), Tensor(rulebook), Tensor(counter)
   args : (Tensor x, Tensor kernel, Tensor out, Tensor rulebook, Tensor counter, Tensor out_grad, int[] paddings, int[] dilations, int[] strides, int groups, bool subm, str key)
   output : Tensor(x_grad), Tensor(kernel_grad)
   kernel :
     func : conv3d_coo_grad{sparse_coo, dense, sparse_coo, dense, dense, sparse_coo -> sparse_coo, dense}
 
-- backward_api : divide_grad
+- backward_op : divide_grad
   forward : divide(Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad)
@@ -96,13 +96,13 @@
     func : divide_coo_coo_grad{sparse_coo, sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo},
            divide_csr_csr_grad{sparse_csr, sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr}
 
-- backward_api : divide_scalar_grad
+- backward_op : divide_scalar_grad
   forward : divide_scalar (Tensor x, float scalar) -> Tensor(out)
   args : (Tensor out_grad, float scalar)
   output : Tensor(x_grad)
   invoke : divide_scalar(out_grad, scalar)
 
-- backward_api : expm1_grad
+- backward_op : expm1_grad
   forward : expm1(Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -110,7 +110,7 @@
     func : expm1_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            expm1_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : leaky_relu_grad
+- backward_op : leaky_relu_grad
   forward : leaky_relu(Tensor x, float alpha) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float alpha)
   output : Tensor(x_grad)
@@ -118,7 +118,7 @@
     func : leaky_relu_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            leaky_relu_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : log1p_grad
+- backward_op : log1p_grad
   forward : log1p(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -126,14 +126,14 @@
     func : log1p_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            log1p_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : masked_matmul_grad
+- backward_op : masked_matmul_grad
   forward : masked_matmul(Tensor x, Tensor y, Tensor mask) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad)
   kernel :
     func : masked_matmul_csr_grad{dense, dense, sparse_csr -> dense, dense}
 
-- backward_api : matmul_grad
+- backward_op : matmul_grad
   forward : matmul(Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad)
@@ -143,14 +143,14 @@
            matmul_coo_dense_grad {sparse_coo, dense, dense -> sparse_coo, dense},
            matmul_coo_coo_grad {sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo}
 
-- backward_api : maxpool_grad
+- backward_op : maxpool_grad
   forward : maxpool(Tensor x, int[] kernel_sizes, int[] paddings, int[] dilations, int[] strides) -> Tensor(out), Tensor(rulebook), Tensor(counter)
   args : (Tensor x, Tensor rulebook, Tensor counter, Tensor out, Tensor out_grad, int[] kernel_sizes)
   output : Tensor(x_grad)
   kernel :
     func : maxpool_coo_grad {sparse_coo, dense, dense, sparse_coo, sparse_coo -> sparse_coo}
 
-- backward_api : multiply_grad
+- backward_op : multiply_grad
   forward : multiply(Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad)
@@ -158,7 +158,7 @@
     func : multiply_coo_coo_grad{sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo},
            multiply_csr_csr_grad{sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr}
 
-- backward_api : mv_grad
+- backward_op : 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)
@@ -166,7 +166,7 @@
     func : mv_coo_grad{sparse_coo, dense, dense -> sparse_coo, dense},
            mv_csr_grad{sparse_csr, dense, dense -> sparse_csr, dense}
 
-- backward_api : pow_grad
+- backward_op : pow_grad
   forward : pow(Tensor x, float factor) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, float factor)
   output : Tensor(x_grad)
@@ -174,7 +174,7 @@
     func : pow_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            pow_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : relu6_grad
+- backward_op : relu6_grad
   forward : relu6(Tensor x, float threshold) -> Tensor(out)
   args : (Tensor out, Tensor out_grad, float threshold)
   output : Tensor(x_grad)
@@ -182,7 +182,7 @@
     func : relu6_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            relu6_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : relu_grad
+- backward_op : relu_grad
   forward : relu(Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -190,13 +190,13 @@
     func : relu_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            relu_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : scale_grad
+- backward_op : scale_grad
   forward : scale(Tensor x, float scale, float bias, bool bias_after_scale) -> Tensor(out)
   args : (Tensor out_grad, float scale)
   output : Tensor(x_grad)
   invoke : scale(out_grad, scale, 0.0, true)
 
-- backward_api : sin_grad
+- backward_op : sin_grad
   forward : sin(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -204,7 +204,7 @@
     func : sin_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            sin_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : sinh_grad
+- backward_op : sinh_grad
   forward : sinh(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -212,21 +212,21 @@
     func : sinh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            sinh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : softmax_grad
+- backward_op : softmax_grad
   forward : softmax(Tensor x, int axis=-1) -> Tensor(out)
   args : (Tensor out, Tensor out_grad, int axis)
   output : Tensor(x_grad)
   kernel :
     func : softmax_csr_grad{sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : sparse_coo_tensor_grad
+- backward_op : sparse_coo_tensor_grad
   forward : sparse_coo_tensor(Tensor values, Tensor indices, IntArray dense_shape) -> Tensor(out)
   args : (Tensor indices, Tensor out_grad)
   output : Tensor(values_grad)
   kernel :
     func : sparse_coo_tensor_grad{dense, sparse_coo -> dense}
 
-- backward_api : sqrt_grad
+- backward_op : sqrt_grad
   forward : sqrt(Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -234,7 +234,7 @@
     func : sqrt_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            sqrt_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : square_grad
+- backward_op : square_grad
   forward : square(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -242,7 +242,7 @@
     func : square_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            square_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : subtract_grad
+- backward_op : subtract_grad
   forward : subtract(Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out_grad)
   output : Tensor(x_grad), Tensor(y_grad)
@@ -250,7 +250,7 @@
     func : subtract_coo_coo_grad{sparse_coo, sparse_coo, sparse_coo -> sparse_coo, sparse_coo},
            subtract_csr_csr_grad{sparse_csr, sparse_csr, sparse_csr -> sparse_csr, sparse_csr}
 
-- backward_api : tan_grad
+- backward_op : tan_grad
   forward : tan(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
@@ -258,7 +258,7 @@
     func : tan_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            tan_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : tanh_grad
+- backward_op : tanh_grad
   forward : tanh(Tensor x) -> Tensor(out)
   args : (Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
@@ -266,28 +266,28 @@
     func : tanh_coo_grad {sparse_coo, sparse_coo -> sparse_coo},
            tanh_csr_grad {sparse_csr, sparse_csr -> sparse_csr}
 
-- backward_api : to_dense_grad
+- backward_op : to_dense_grad
   forward : to_dense(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
   kernel :
     func : coo_to_dense_grad{sparse_coo, dense -> sparse_coo}
 
-- backward_api : to_sparse_coo_grad
+- backward_op : to_sparse_coo_grad
   forward : to_sparse_coo(Tensor x, int64_t sparse_dim) -> Tensor(out)
   args : (Tensor out_grad)
   output : Tensor(x_grad)
   kernel :
     func : coo_to_dense { sparse_coo -> dense }
 
-- backward_api : values_grad
+- backward_op : values_grad
   forward : values_coo(Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
   output : Tensor(x_grad)
   kernel :
     func : values_coo_grad{sparse_coo, dense-> sparse_coo}
 
-- backward_api: fused_attention_grad
+- backward_op: fused_attention_grad
   forward : fused_attention_csr(Tensor query, Tensor key, Tensor value, Tensor sparse_mask, Tensor key_padding_mask, Tensor attn_mask) -> Tensor(out), Tensor(softmax)
   args: (Tensor query, Tensor key, Tensor value, Tensor softmax, Tensor out_grad)
   output : Tensor(query_grad), Tensor(key_grad), Tensor(value_grad)