[fluid_ops] clean push_dense yaml config (#72056)

co63oc · web-flow · commit ba72d6374ea7 · 2025-04-07T14:44:30.000+08:00
diff --git a/paddle/fluid/ir_adaptor/translator/op_compat_gen.py b/paddle/fluid/ir_adaptor/translator/op_compat_gen.py
@@ -181,15 +181,6 @@ def insert_new_mutable_attributes(
         "out": "Out",
     }
     op_arg_name_mappings['fused_softmax_mask_grad'].update({"out": "Softmax"})
-    op_arg_name_mappings['push_sparse_v2'].update(
-        {"out_grad_in": "Out@GRAD", "out_grad_out": "Out@GRAD"}
-    )
-    op_arg_name_mappings['push_box_sparse'].update(
-        {"out_grad_in": "Out@GRAD", "out_grad_out": "Out@GRAD"}
-    )
-    op_arg_name_mappings['push_gpups_sparse'].update(
-        {"out_grad": "Out@GRAD", "out_grad_grad": "Out@GRAD"}
-    )
 
     sparse_op_yaml_files = sparse_op_yaml_file.split(",")
     for yaml_file in sparse_op_yaml_files:
diff --git a/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py b/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py
@@ -197,10 +197,6 @@
     'lod_reset_',
     'max_pool2d_v2',
     'partial_sum',
-    'pull_gpups_sparse',
-    'pull_gpups_sparse_',
-    'push_gpups_sparse',
-    'push_gpups_sparse_',
     'random_routing',
     'rnn_',
     'row_conv',
@@ -214,25 +210,17 @@
     'match_matrix_tensor',
     'c_scatter',
     "cross_entropy_grad2",
-    'push_sparse_v2',
-    'push_sparse_v2_',
-    'pull_sparse_v2',
     'partial_concat',
     'partial_send',
     'partial_recv',
     'partial_allgather',
     'partial_allgather_',
     'gemm_epilogue',
-    'push_dense',
     'legacy_matmul',
     'legacy_matmul_grad',
     'legacy_matmul_double_grad',
     'global_scatter',
     'global_gather',
-    'pull_box_sparse',
-    'pull_box_sparse_',
-    'push_box_sparse',
-    'push_box_sparse_',
     'send_and_recv',
     'send_and_recv_',
     'straight_through_estimator',
diff --git a/paddle/phi/api/generator/api_gen.py b/paddle/phi/api/generator/api_gen.py
@@ -18,7 +18,6 @@
 from api_base import PREFIX_TENSOR_NAME, BaseAPI
 
 backward_api_black_list = [
-    "pull_sparse_v2_grad",  # tensor = push_sparse_v2() is not implemented in api_custom_impl.cc
     "scale_grad",  # tensor = scale is not implemented in api_custom_impl.cc
 ]
 
diff --git a/paddle/phi/ops/yaml/backward.yaml b/paddle/phi/ops/yaml/backward.yaml
@@ -2538,22 +2538,6 @@
     data_type : x
   optional : boxes_num
 
-- backward_op : pull_sparse_v2_grad
-  forward : pull_sparse_v2 (Tensor[] ids, Tensor[] w, int embedding_dim = 11, int table_id = 0, str accessor_class = "", str ctrlabel_name = "", int padding_id = 0, bool scale_sparse_grad = true, str[] input_names = {}, bool is_distributed = true) -> Tensor[](out)
-  args : (Tensor[] ids, Tensor[] w, Tensor[] out_grad, int embedding_dim = 11, int table_id = 0, str accessor_class = "", str ctrlabel_name = "", int padding_id = 0, bool scale_sparse_grad = true, str[] input_names = {}, bool is_distributed = true)
-  output : Tensor[](out_grad_out)
-  invoke : push_sparse_v2(ids, w, out_grad, embedding_dim, table_id, accessor_class, ctrlabel_name, padding_id, scale_sparse_grad, input_names, is_distributed)
-
-- backward_op : push_gpups_sparse
-  forward : pull_gpups_sparse (Tensor w, Tensor[] ids, int[] size={}, bool is_sparse=false, bool is_distributed=false) -> Tensor[](out)
-  args : (Tensor[] ids, Tensor[] out_grad, int[] size, bool is_sparse, bool is_distributed)
-  output : Tensor[](out_grad_grad){out_grad.size()}
-  infer_meta :
-    func : PushGpupsSparseInferMeta
-  kernel :
-    func : push_gpups_sparse
-  inplace : (out_grad -> out_grad_grad)
-
 - backward_op : put_along_axis_grad
   forward : put_along_axis (Tensor arr, Tensor indices, Tensor values, int axis, str reduce = "assign", bool include_self = true) -> Tensor(out)
   args : (Tensor arr, Tensor indices, Tensor values, Tensor out, Tensor out_grad, int axis, str reduce, bool include_self)
diff --git a/paddle/phi/ops/yaml/inconsistent/static_backward.yaml b/paddle/phi/ops/yaml/inconsistent/static_backward.yaml
@@ -470,18 +470,6 @@
   output : Tensor(in_grad)
   invoke : print(out_grad, first_n, message, summarize, print_tensor_name, print_tensor_type, print_tensor_shape, print_tensor_layout, print_tensor_lod, print_phase, is_forward)
 
-- backward_op : push_box_sparse
-  forward : pull_box_sparse (Tensor w, Tensor[] ids, bool is_sparse = false, bool is_distributed = false, int size = 1) -> Tensor[](out){ids.size()}
-  args : (Tensor[] ids, Tensor[] out_grad_in, bool is_sparse = false, bool is_distributed = false, int size = 1)
-  output : Tensor[](out_grad_out){out_grad_in.size()}
-  infer_meta :
-    func : UnchangedMultiInferMeta
-    param : [out_grad_in]
-  kernel :
-    func : push_box_sparse
-    data_type : out_grad_in
-  inplace : (out_grad_in -> out_grad_out)
-
 - backward_op : remainder_grad
   forward : remainder (Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out_grad)
diff --git a/paddle/phi/ops/yaml/inconsistent/static_ops.yaml b/paddle/phi/ops/yaml/inconsistent/static_ops.yaml
@@ -631,57 +631,6 @@
   traits : pir::SideEffectTrait
   backward: print_grad
 
-- op : pull_box_sparse
-  args : (Tensor w, Tensor[] ids, bool is_sparse = false, bool is_distributed = false, int size = 1)
-  output : Tensor[](out){ids.size()}
-  infer_meta :
-    func : PullBoxSparseInferMeta
-  kernel :
-    func : pull_box_sparse
-    data_type : ids
-
-- op : pull_gpups_sparse
-  args : (Tensor w, Tensor[] ids, int[] size={}, bool is_sparse=false, bool is_distributed=false)
-  output : Tensor[](out){ids.size()}
-  infer_meta :
-    func : PullGpupsSparseInferMeta
-  kernel :
-    func : pull_gpups_sparse
-    data_type : PullGpupsSparseKernelKey
-  optional : w
-  backward: push_gpups_sparse
-
-- op : pull_sparse_v2
-  args : (Tensor[] ids, Tensor[] w, int embedding_dim = 11, int table_id = 0, str accessor_class = "", str ctr_label_name = "", int padding_id = 0, bool scale_sparse_grad = true, str[] input_names = {}, bool is_distributed = true)
-  output : Tensor[](out){w.size()}
-  infer_meta :
-    func : PullSparseV2InferMeta
-  kernel :
-    func : pull_sparse_v2
-    data_type : DataType::FLOAT32
-  backward : pull_sparse_v2_grad
-
-- op : push_dense
-  args : (Tensor[] ids, int table_id = -1, float scale_data_norm = -1.0f, str[] input_names = {})
-  output :
-  infer_meta :
-    func : PushDenseInferMeta
-    param : [ids, table_id, scale_data_norm, input_names]
-  kernel :
-    func : push_dense
-    data_type : DataType::FLOAT32
-
-- op : push_sparse_v2
-  args : (Tensor[] ids, Tensor[] w, Tensor[] out_grad_in, int embedding_dim = 11, int table_id = 0, str accessor_class = "", str ctr_label_name = "", int padding_id = 0, bool scale_sparse_grad = true, str[] input_names = {}, bool is_distributed = true)
-  output : Tensor[](out_grad_out){out_grad_in.size()}
-  infer_meta :
-    func : UnchangedMultiInferMeta
-    param : [out_grad_in]
-  kernel :
-    func : push_sparse_v2
-    data_type : out_grad_in
-  inplace: (out_grad_in -> out_grad_out)
-
 # Note: dequantize_linear and quantize_linear are supported using one op maker in fluid, the out_scale can't be used in dequantize_linear
 # so ,the out_scale is optional. Currently, we can't modify the op definition of dequantize_linear/quantize_linear and it can cause incompatibility problem
 # We need modify dequantize_linear/quantize_linear yaml and make it more reasonable when we abandon Fluid op.
diff --git a/paddle/phi/ops/yaml/op_compat.yaml b/paddle/phi/ops/yaml/op_compat.yaml
@@ -3025,53 +3025,6 @@
   outputs :
     out : Out
 
-- op : pull_box_sparse
-  inputs :
-    { w : W, ids: Ids}
-  outputs :
-    out : Out
-  attrs :
-    sparse : is_sparse
-  extra :
-    attrs : [bool is_sparse = false, bool is_distributed = false, int size = 1]
-
-- op : pull_gpups_sparse
-  backward : push_gpups_sparse
-  inputs :
-    {w : W, ids : Ids}
-  outputs :
-    out : Out
-
-- op : pull_sparse_v2
-  inputs :
-    { ids : Ids, w : W}
-  outputs :
-    out : Out
-  extra :
-    attrs : [int embedding_dim = 11, int table_id = 0, str accessor_class = "", str ctr_label_name = "", int padding_id = 0, bool scale_sparse_grad = true, 'str[] input_names = {}', bool is_distributed = true]
-
-- op : push_box_sparse
-  inputs :
-    ids: Ids
-  outputs :
-    out : Out
-  attrs :
-    sparse : is_sparse
-
-- op : push_dense
-  inputs :
-    ids : Ids
-  attrs :
-    {table_id : TableId, scale_data_norm : ScaleDataNorm, input_names: InputNames}
-
-- op : push_sparse_v2
-  inputs :
-    { x : Ids, W : w}
-  outputs :
-    out : Out
-  extra :
-    attrs : [int embedding_dim = 11, int table_id = 0, str accessor_class = "", str ctr_labe_lname = "", int padding_id = 0, bool scales_parse_grad = true, 'str[] input_names = {}', bool is_distributed = true]
-
 - op : put_along_axis
   backward : put_along_axis_grad
   inputs :

Original file line number	Diff line number	Diff line change
`@@ -18,7 +18,6 @@`
`18`	`18`	`from api_base import PREFIX_TENSOR_NAME, BaseAPI`
`19`	`19`
`20`	`20`	`backward_api_black_list = [`
`21`		`- "pull_sparse_v2_grad", # tensor = push_sparse_v2() is not implemented in api_custom_impl.cc`
`22`	`21`	`"scale_grad", # tensor = scale is not implemented in api_custom_impl.cc`
`23`	`22`	`]`
`24`	`23`