【PIR Dist Op Reg No.22】 reg pull_sparse_v2 (PaddlePaddle#63014)

* feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2 * feat(pir): reg pull_sparse_v2
co63oc · May 10, 2024 · 9831401 · 9831401
1 parent 7898da1
commit 9831401
Show file tree

Hide file tree

Showing 8 changed files with 131 additions and 2 deletions.
diff --git a/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py b/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py
@@ -195,6 +195,7 @@
     'prune_gate_by_capacity',
     'push_sparse_v2',
     'push_sparse_v2_',
+    'pull_sparse_v2',
     'partial_concat',
     'partial_send',
     'partial_recv',

diff --git a/paddle/fluid/pir/dialect/operator/ir/ops.yaml b/paddle/fluid/pir/dialect/operator/ir/ops.yaml
@@ -1362,6 +1362,16 @@
   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 :
@@ -1373,7 +1383,7 @@
     data_type : DataType::FLOAT32
 
 - op : push_sparse_v2
-  args : (Tensor[] ids, Tensor[] w, Tensor[] out_grad_in, int embeddingdim = 11, int tableid = 0, str accessorclass = "", str ctrlabelname = "", int paddingid = 0, bool scalesparsegrad = true, str[] inputnames = {}, bool is_distributed = true)
+  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

diff --git a/paddle/phi/api/yaml/backward.yaml b/paddle/phi/api/yaml/backward.yaml
@@ -1870,6 +1870,12 @@
     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)

diff --git a/paddle/phi/api/yaml/op_compat.yaml b/paddle/phi/api/yaml/op_compat.yaml
@@ -2686,6 +2686,14 @@
   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_dense
   inputs :
     ids : Ids
@@ -2698,7 +2706,7 @@
   outputs :
     out : Out
   extra :
-    attrs : [int embeddingdim = 11, int tableid = 0, str accessorclass = "", str ctrlabelname = "", int paddingid = 0, bool scalesparsegrad = true, 'str[] inputnames = {}', bool is_distributed = true]
+    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

diff --git a/paddle/phi/infermeta/binary.cc b/paddle/phi/infermeta/binary.cc
@@ -2787,6 +2787,44 @@ void PullGpupsSparseInferMeta(const MetaTensor& w,
   }
 }
 
+void PullSparseV2InferMeta(const std::vector<const MetaTensor*>& ids,
+                           const std::vector<const MetaTensor*>& w,
+                           int embedding_dim,
+                           int table_id,
+                           const std::string& accessor_class,
+                           const std::string& ctrlabel_name,
+                           int padding_id,
+                           bool scale_sparse_grad,
+                           const std::vector<std::string>& input_names,
+                           bool is_distributed,
+                           std::vector<MetaTensor*> out) {
+  PADDLE_ENFORCE_GE(ids.size(),
+                    1UL,
+                    phi::errors::InvalidArgument(
+                        "Input(Ids) of PullSparseV2Op can not be null"));
+  PADDLE_ENFORCE_GE(out.size(),
+                    1UL,
+                    phi::errors::InvalidArgument(
+                        "Output(Out) of PullSparseV2Op can not be null"));
+
+  auto hidden_size = embedding_dim;
+  const size_t n_ids = ids.size();
+  std::vector<phi::DDim> outs_dims;
+  outs_dims.resize(n_ids);
+  for (size_t i = 0; i < n_ids; ++i) {
+    const auto ids_dims = ids[i]->dims();
+    auto out_dim = common::vectorize(ids_dims);
+    out_dim.push_back(hidden_size);
+    outs_dims[i] = common::make_ddim(out_dim);
+  }
+
+  for (size_t i = 0; i < n_ids; ++i) {
+    out[i]->set_dims(outs_dims[i]);
+    out[i]->share_lod(*ids[i], i);
+    out[i]->set_dtype(w[i]->dtype());
+  }
+}
+
 void ApplyPerChannelScaleInferMeta(const MetaTensor& x,
                                    const MetaTensor& scales,
                                    MetaTensor* out) {

diff --git a/paddle/phi/infermeta/binary.h b/paddle/phi/infermeta/binary.h
@@ -480,6 +480,18 @@ void PullGpupsSparseInferMeta(const MetaTensor& w,
                               bool is_distributed,
                               std::vector<MetaTensor*> out);
 
+void PullSparseV2InferMeta(const std::vector<const MetaTensor*>& ids,
+                           const std::vector<const MetaTensor*>& w,
+                           int embedding_dim,
+                           int table_id,
+                           const std::string& accessor_class,
+                           const std::string& ctrlabel_name,
+                           int padding_id,
+                           bool scale_sparse_grad,
+                           const std::vector<std::string>& input_names,
+                           bool is_distributed,
+                           std::vector<MetaTensor*> out);
+
 void RepeatInterleaveWithTensorIndexInferMeta(const MetaTensor& x,
                                               const MetaTensor& repeats,
                                               int dim,

diff --git a/test/ir/pir/translator/CMakeLists.txt b/test/ir/pir/translator/CMakeLists.txt
@@ -28,6 +28,7 @@ list(APPEND DISTRIBUTED_OP_TRANSLATOR_TEST test_partial_recv_translator)
 list(APPEND DISTRIBUTED_OP_TRANSLATOR_TEST
      test_prune_gate_by_capacity_translator)
 list(APPEND DISTRIBUTED_OP_TRANSLATOR_TEST test_pull_gpups_sparse_translator)
+list(APPEND DISTRIBUTED_OP_TRANSLATOR_TEST test_pull_sparse_v2_translator)
 list(APPEND DISTRIBUTED_OP_TRANSLATOR_TEST test_random_routing_translator)
 list(APPEND DISTRIBUTED_OP_TRANSLATOR_TEST test_limit_by_capacity_translator)
 list(APPEND DISTRIBUTED_OP_TRANSLATOR_TEST test_global_scatter_translator)

diff --git a/test/ir/pir/translator/test_pull_sparse_v2_translator.py b/test/ir/pir/translator/test_pull_sparse_v2_translator.py
@@ -0,0 +1,53 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import test_op_translator
+
+import paddle
+from paddle.base import core
+from paddle.base.layer_helper import LayerHelper
+
+
+class TestPullSparseV2OpTranslator(
+    test_op_translator.TestOpWithBackwardTranslator
+):
+    def setUp(self):
+        self.place = core.Place()
+        self.place.set_place(paddle.CPUPlace())
+        self.new_scope = paddle.static.Scope()
+        self.main_program = paddle.static.Program()
+        self.forward_op_type = "pull_sparse_v2"
+        self.backward_op_type = "push_sparse_v2"
+
+    def append_op(self):
+        self.op_type = "pull_sparse_v2"
+        ids = paddle.ones(shape=(1,), dtype='int64')
+        w = paddle.ones(shape=(1,), dtype='int64')
+        out = paddle.ones(shape=(1,), dtype='int64')
+        helper = LayerHelper(self.op_type)
+        helper.append_op(
+            type=self.op_type,
+            inputs={"Ids": [ids], "W": [w]},
+            outputs={"Out": [out]},
+        )
+        return out
+
+    def test_translator(self):
+        self.check()
+
+
+if __name__ == "__main__":
+    unittest.main()