!6411 [Auto parallel] Add a flag to control whether to overwrite the right-node in triangle-elimination of DP algorithm

Merge pull request !6411 from Xiaoda/23-fix-the-triangle-elimination-problem

mindspore/ccsrc/frontend/parallel/auto_parallel/costmodel.h

@@ -246,14 +246,15 @@ struct SourceEliminationDecision : public Decision {
  */
 struct TriangleEliminationDecision : public Decision {
   TriangleEliminationDecision(StrategyPtr elimi_stra, CostPtr elimi_op_cost, CostPtr l_edge_cost, CostPtr r_edge_cost,
-                              StrategyPtr left_stra, CostPtr l_node_cost, StrategyPtr right_stra)
+                              StrategyPtr left_stra, CostPtr l_node_cost, StrategyPtr right_stra, CostPtr r_node_cost)
       : eliminated_op_strategy_(std::move(elimi_stra)),
         eliminated_op_cost_(std::move(elimi_op_cost)),
         left_edge_cost_(std::move(l_edge_cost)),
         right_edge_cost_(std::move(r_edge_cost)),
         left_node_strategy_(std::move(left_stra)),
         left_node_cost_(std::move(l_node_cost)),
-        right_node_strategy_(std::move(right_stra)) {
+        right_node_strategy_(std::move(right_stra)),
+        right_node_cost_(std::move(r_node_cost)) {
     type_ = DecisionType::TRIANGLE_ELIMINATION;
   }
 
@@ -264,6 +265,7 @@ struct TriangleEliminationDecision : public Decision {
   StrategyPtr left_node_strategy_;
   CostPtr left_node_cost_;
   StrategyPtr right_node_strategy_;
+  CostPtr right_node_cost_;
   MS_DECLARE_PARENT(TriangleEliminationDecision, Decision);
 };
 

mindspore/ccsrc/frontend/parallel/auto_parallel/dp_algo_costmodel.cc

@@ -199,9 +199,16 @@ Status RecoverStrategy(std::vector<EliminationPtr> eliminations) {
       eliminated_node->SetSelectedStrategyAndCost(decision->eliminated_op_strategy_, decision->eliminated_op_cost_);
       left_edge->set_selected_cost(decision->left_edge_cost_);
       right_edge->set_selected_cost(decision->right_edge_cost_);
-      // Since Triangle is eliminated into 'left_node', only 'left_node' is needed to recover the strategy.
+      // 'left_node' recovers the strategy.
       left_node->SetSelectedStrategyAndCost(decision->left_node_strategy_, decision->left_node_cost_);
+      if (TRIANGLE_STRATEGY_OVERWRITE) {
+        // 'right_node' recovers the strategy.
+        MS_LOG(INFO) << "Overwrite the right-node: " << right_node->name() << " in recovering triangle elimination.";
+        right_node->SetSelectedStrategyAndCost(decision->right_node_strategy_, decision->right_node_cost_);
+      } else {
+        // In this case, 'right_node' is not overwriten strategy, and it checks strategy consistency.
         right_node->CheckSelectedStrategy(decision->right_node_strategy_);
+      }
       MS_LOG(INFO) << "Recover triangleElimination succeeded.";
     } else if ((*rit)->isa<StarElimination>()) {
       auto elimination = (*rit)->cast<StarEliminationPtr>();

mindspore/ccsrc/frontend/parallel/auto_parallel/graph_costmodel.cc

@@ -40,6 +40,7 @@ bool FULLY_USE_DEVICES = DEFAULT_FULLY_USE_DEVICES;
 bool ELEMENTWISE_OP_STRA_FOLLOW = DEFAULT_ELEMENTWISE_OP_STRA_FOLLOW;
 bool MULTI_SUBGRAPHS = DEFAULT_IS_MULTI_SUBGRAPHS;
 int32_t RUN_PHASE = DEFAULT_RUN_PHASE;
+bool TRIANGLE_STRATEGY_OVERWRITE = DEFAULT_TRIANGLE_STRATEGY_OVERWRITE;
 
 void CostGraph::SetDeviceMemoryAndCostParameter() {
   MS_EXCEPTION_IF_NULL(CostModelContext::GetInstance());
@@ -154,6 +155,14 @@ void CostGraph::SetDeviceMemoryAndCostParameter() {
     MS_LOG(INFO) << "multi_subgraphs: false.";
   }
 
+  auto overwrite = CostModelContext::GetInstance()->triangle_strategy_overwrite();
+  TRIANGLE_STRATEGY_OVERWRITE = overwrite;
+  if (TRIANGLE_STRATEGY_OVERWRITE) {
+    MS_LOG(INFO) << "triangle_strategy_overwrite: true.";
+  } else {
+    MS_LOG(INFO) << "triangle_strategy_overwrite: false.";
+  }
+
   // RUN_PHASE
   auto phase = CostModelContext::GetInstance()->run_phase();
   if (phase != 0 && phase != 1) {
@@ -1294,8 +1303,17 @@ void CostGraph::CreateTriangleEliminationSubCostList(StrategyPtr elimi_op_stra,
           elimi_op_cost->communication_without_parameter_ + left_edge_cost->communication_without_parameter_ +
           left_node_cost->communication_without_parameter_ + right_edge_cost->communication_without_parameter_;
 
-        auto decision = std::make_shared<TriangleEliminationDecision>(
+        if (TRIANGLE_STRATEGY_OVERWRITE) {
-          elimi_op_stra, elimi_op_cost, left_edge_cost, right_edge_cost, left_op_stra, left_node_cost, right_op_stra);
+          new_computation += right_op_cost->computation_cost_;
+          new_memory += right_op_cost->memory_with_reuse_;
+          new_commu_cost += right_op_cost->communication_cost_;
+          new_commu_forward += right_op_cost->communication_forward_;
+          new_commu_without += right_op_cost->communication_without_parameter_;
+        }
+
+        auto decision =
+          std::make_shared<TriangleEliminationDecision>(elimi_op_stra, elimi_op_cost, left_edge_cost, right_edge_cost,
+                                                        left_op_stra, left_node_cost, right_op_stra, right_op_cost);
         auto new_cost = std::make_shared<Cost>(new_computation, new_commu_cost, decision);
         new_cost->communication_without_parameter_ = new_commu_without;
         new_cost->communication_with_partial_para_ =

mindspore/ccsrc/frontend/parallel/auto_parallel/graph_costmodel.h

@@ -46,6 +46,7 @@ extern bool FULLY_USE_DEVICES;
 extern bool ELEMENTWISE_OP_STRA_FOLLOW;
 extern bool MULTI_SUBGRAPHS;
 extern int32_t RUN_PHASE;
+extern bool TRIANGLE_STRATEGY_OVERWRITE;
 
 class CostGraph {
   // 'CostGraph' consists of Operators and edges between them. An edge is created between two Operators if they have

mindspore/ccsrc/frontend/parallel/costmodel_context.cc

@@ -64,6 +64,7 @@ void CostModelContext::ResetAlgoParameters() {
   tensor_slice_alignment_size_ = DEFAULT_TENSOR_SLICE_ALIGNMENT_SIZE;
   fully_use_device_ = DEFAULT_FULLY_USE_DEVICES;
   elementwise_stra_follow_ = DEFAULT_ELEMENTWISE_OP_STRA_FOLLOW;
+  triangle_strategy_overwrite_ = DEFAULT_TRIANGLE_STRATEGY_OVERWRITE;
 }
 
 void CostModelContext::set_costmodel_context_for_device(const std::string &device_target) {
@@ -133,6 +134,8 @@ void CostModelContext::set_elementwise_stra_follow(bool elementwise_follow) {
   elementwise_stra_follow_ = elementwise_follow;
 }
 
+void CostModelContext::set_triangle_strategy_overwrite(bool overwrite) { triangle_strategy_overwrite_ = overwrite; }
+
 void CostModelContext::set_run_phase(int32_t phase) { run_phase_ = phase; }
 
 struct CostRegister {

mindspore/ccsrc/frontend/parallel/costmodel_context.h

@@ -44,6 +44,8 @@ namespace parallel {
 #define DEFAULT_RUN_PHASE 0
 #define TRAINING_PHASE 0
 #define INFERENCE_PHASE 1
+#define DEFAULT_TRIANGLE_STRATEGY_OVERWRITE true;
+
 class CostModelContext {
  public:
   ~CostModelContext() = default;
@@ -133,6 +135,9 @@ class CostModelContext {
   void set_elementwise_stra_follow(bool);
   bool elementwise_stra_follow() const { return elementwise_stra_follow_; }
 
+  void set_triangle_strategy_overwrite(bool);
+  bool triangle_strategy_overwrite() const { return triangle_strategy_overwrite_; }
+
   void set_run_phase(int32_t);
   int32_t run_phase() const { return run_phase_; }
 
@@ -167,6 +172,10 @@ class CostModelContext {
   // MULTI_SUBGRAPHS
   bool is_multi_subgraphs_;
 
+  // In the recovery phase of DP algorithm, when encountering triangle structure,
+  // whether overwrite the right-node strategy
+  bool triangle_strategy_overwrite_;
+
   int32_t run_phase_;  // 0: 'training', 1: 'inference'
 
   int32_t costmodel_allreduce_fusion_algorithm_;

tests/ut/python/parallel/test_auto_parallel_triangle_overwrite.py

@@ -0,0 +1,73 @@
+# Copyright 2019 Huawei Technologies Co., Ltd
+#
+# 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 numpy as np
+import mindspore as ms
+import mindspore.nn as nn
+from mindspore.common.api import _executor
+from mindspore.ops import composite as C
+from mindspore.ops import operations as P
+from mindspore.parallel._utils import _reset_op_id as reset_op_id
+from mindspore import context, Tensor, Parameter
+from tests.ut.python.ops.test_math_ops import VirtualLoss
+
+grad_all = C.GradOperation(get_all=True)
+
+class NetWithLoss(nn.Cell):
+    def __init__(self, network):
+        super(NetWithLoss, self).__init__()
+        self.loss = VirtualLoss()
+        self.network = network
+
+    def construct(self, x):
+        predict = self.network(x)
+        return self.loss(predict)
+
+class GradWarp(nn.Cell):
+    def __init__(self, network):
+        super(GradWarp, self).__init__()
+        self.network = network
+
+    def construct(self, x):
+        return grad_all(self.network)(x)
+
+def test_triangle_strategy_consistency():
+    class Net(nn.Cell):
+        def __init__(self):
+            super(Net, self).__init__()
+            self.mul1 = P.Mul().shard(((2, 4), (2, 4)))
+            self.mul2 = P.Mul()
+            self.ba1 = P.BiasAdd()
+            self.weight = Parameter(Tensor(np.ones([128, 1000]), dtype=ms.float32), name="weight")
+            self.bias = Parameter(Tensor(np.ones([1000]), dtype=ms.float32), name="bias")
+            self.add = P.TensorAdd().shard(((1, 8), (1, 8)))
+            self.relu = P.ReLU()
+
+        def construct(self, x):
+            out = self.mul1(x, self.weight)
+            mul_out = self.mul2(out, self.weight)
+            ba_out = self.ba1(out, self.bias)
+            ta_out = self.add(mul_out, ba_out)
+            out = self.relu(ta_out)
+            return out
+
+    size = 8
+    context.set_auto_parallel_context(device_num=size, global_rank=0)
+    x = Tensor(np.ones([128, 1000]), dtype=ms.float32)
+    net = NetWithLoss(Net())
+    context.set_auto_parallel_context(parallel_mode="auto_parallel")
+    net.set_auto_parallel()
+    reset_op_id()
+
+    _executor.compile(net, x, phase='train')