【PIR Dist Op Reg No.19】 reg pull_box_sparse (PaddlePaddle#62982)

* fix * fix * fix * fix * fix * fix * add test * add * fix * fix * add out * fix * codestyle * fix * fix backward * merge
yinfan98 · May 7, 2024 · 0d57c4f · 0d57c4f
1 parent 6f35286
commit 0d57c4f
Show file tree

Hide file tree

Showing 9 changed files with 133 additions and 0 deletions.
diff --git a/paddle/fluid/ir_adaptor/translator/op_compat_gen.py b/paddle/fluid/ir_adaptor/translator/op_compat_gen.py
@@ -170,6 +170,9 @@ def insert_new_mutable_attributes(
     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"}
     )

diff --git a/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py b/paddle/fluid/pir/dialect/op_generator/ops_api_gen.py
@@ -206,6 +206,10 @@
     'push_dense',
     'limit_by_capacity',
     'global_scatter',
+    'pull_box_sparse',
+    'pull_box_sparse_',
+    'push_box_sparse',
+    'push_box_sparse_',
 ]
 
 

diff --git a/paddle/fluid/pir/dialect/operator/ir/ops.yaml b/paddle/fluid/pir/dialect/operator/ir/ops.yaml
@@ -1351,6 +1351,15 @@
     func : prune_gate_by_capacity
     data_type : gate_idx
 
+- 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()}

diff --git a/paddle/fluid/pir/dialect/operator/ir/ops_backward.yaml b/paddle/fluid/pir/dialect/operator/ir/ops_backward.yaml
@@ -659,6 +659,18 @@
     func : prod_grad
   composite: prod_grad(x, out, out_grad, dims, keep_dim, reduce_all, x_grad)
 
+- 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 : rank_attention_grad
   forward : rank_attention (Tensor x, Tensor rank_offset, Tensor rank_param, int max_rank = 3, int max_size = 0) -> Tensor(input_help), Tensor(out), Tensor(ins_rank)
   args : (Tensor x, Tensor rank_offset, Tensor rank_param, Tensor input_help, Tensor ins_rank, Tensor out_grad, int max_rank = 3, int max_size = 0)

diff --git a/paddle/fluid/pir/dialect/operator/utils/utils.cc b/paddle/fluid/pir/dialect/operator/utils/utils.cc
@@ -105,6 +105,8 @@ const std::unordered_set<std::string> LegacyOpList = {
     CReduceMinOp::name(),
     CReduceProdOp::name(),
     CScatterOp::name(),
+    PullBoxSparseOp::name(),
+    PushBoxSparseOp::name(),
     PushSparseV2Op::name(),
     PartialSendOp::name(),
     PartialRecvOp::name()};

diff --git a/paddle/phi/api/yaml/op_compat.yaml b/paddle/phi/api/yaml/op_compat.yaml
@@ -2691,6 +2691,16 @@
   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 :
@@ -2706,6 +2716,14 @@
   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

diff --git a/paddle/phi/infermeta/binary.cc b/paddle/phi/infermeta/binary.cc
@@ -2998,6 +2998,33 @@ void PruneGateByCapacityInferMeta(const MetaTensor& gate_idx,
   new_gate_idx->set_dtype(gate_idx.dtype());
 }
 
+void PullBoxSparseInferMeta(const MetaTensor& w,
+                            const std::vector<const MetaTensor*>& ids,
+                            bool is_sparse,
+                            bool is_distributed,
+                            int size,
+                            std::vector<MetaTensor*> out) {
+  auto hidden_size = static_cast<int64_t>(size);
+  const size_t n_ids = ids.size();
+  for (size_t i = 0; i < n_ids; ++i) {
+    MetaTensor* output = out[i];
+    auto ids_dims = ids[i]->dims();
+    int ids_rank = ids_dims.size();
+    PADDLE_ENFORCE_EQ(ids_dims[ids_rank - 1],
+                      1UL,
+                      phi::errors::InvalidArgument(
+                          "Shape error in %lu id, the last dimension of the "
+                          "'Ids' tensor must be 1.",
+                          i));
+    auto out_dim =
+        common::vectorize(common::slice_ddim(ids_dims, 0, ids_rank - 1));
+    out_dim.push_back(hidden_size);
+    output->set_dims(common::make_ddim(out_dim));
+    output->share_lod(*ids[i]);
+    output->set_dtype(w.dtype());
+  }
+}
+
 void RepeatInterleaveWithTensorIndexInferMeta(const MetaTensor& x,
                                               const MetaTensor& repeats,
                                               int dim,

diff --git a/paddle/phi/infermeta/binary.h b/paddle/phi/infermeta/binary.h
@@ -479,6 +479,13 @@ void PReluInferMeta(const MetaTensor& x,
                     MetaTensor* out,
                     MetaConfig config = MetaConfig());
 
+void PullBoxSparseInferMeta(const MetaTensor& w,
+                            const std::vector<const MetaTensor*>& ids,
+                            bool is_sparse,
+                            bool is_distributed,
+                            int size,
+                            std::vector<MetaTensor*> out);
+
 void PullGpupsSparseInferMeta(const MetaTensor& w,
                               const std::vector<const MetaTensor*>& ids,
                               const std::vector<int>& size,

diff --git a/test/ir/pir/translator/test_pull_box_sparse_translator.py b/test/ir/pir/translator/test_pull_box_sparse_translator.py
@@ -0,0 +1,51 @@
+# 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.layer_helper import LayerHelper
+
+
+class TestPullBoxSparseOpTranslator(
+    test_op_translator.TestOpWithBackwardTranslator
+):
+    def append_op(self):
+        self.forward_op_type = "pull_box_sparse"
+        self.backward_op_type = "push_box_sparse"
+        ids = paddle.ones(shape=(1, 1), dtype='float32')
+        w = paddle.ones(shape=(1, 1), dtype='float32')
+        out = paddle.ones(shape=(1, 1), dtype='float32')
+        attrs = {
+            'is_sparse': False,
+            'is_distributed': False,
+            'size': 1,
+        }
+        forward_helper = LayerHelper(self.forward_op_type)
+        forward_helper.append_op(
+            type=self.forward_op_type,
+            inputs={"W": w, "Ids": [ids]},
+            outputs={"Out": [out]},
+            attrs=attrs,
+        )
+        return out
+
+    def test_translator(self):
+        self.check()
+
+
+if __name__ == "__main__":
+    unittest.main()