update alloc_continuous_space_for_grad_pass (#18288)

test=release/1.5

Author: chengduo

paddle/fluid/framework/ir/alloc_continuous_space_for_grad_pass.cc

@@ -16,6 +16,7 @@
 #include <algorithm>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <utility>
 #include <vector>
 #include "paddle/fluid/framework/details/build_strategy.h"
@@ -84,24 +85,27 @@ class AllocContinuousSpaceForGradPass : public ir::Pass {
     }
 
     if (params_grads.size() == 0) {
-      LOG(WARNING) << "Doesn't find gradients";
+      LOG(INFO) << "Doesn't find gradients";
       return;
     }
 
-    std::unordered_map<std::string, ir::Node *> vars;
+    std::unordered_map<std::string, ir::Node *> var_name2node;
+    std::unordered_map<std::string, std::unordered_set<ir::Node *>>
+        var_name2node_set;
     for (ir::Node *node : result.Nodes()) {
       if (node->IsVar() && node->Var()) {
         // Note: The graph may have the same name node. For example, parameter
         // is the input of operator and it also is the output of optimizer;
-        vars.emplace(node->Var()->Name(), node);
+        var_name2node.emplace(node->Var()->Name(), node);
+        var_name2node_set[node->Var()->Name()].emplace(node);
       }
     }
 
     auto &group_grads_params =
         result.Get<details::GroupGradsAndParams>(details::kGroupGradsAndParams);
 
     // Note: the order of params_grads may be changed by SetGroupGradsAndParams.
-    SetGroupGradsAndParams(vars, params_grads, &group_grads_params);
+    SetGroupGradsAndParams(var_name2node, params_grads, &group_grads_params);
 
     params_grads.clear();
     for (auto &group_p_g : group_grads_params) {
@@ -116,9 +120,16 @@ class AllocContinuousSpaceForGradPass : public ir::Pass {
     auto dtype = kDefaultDtype;
     for (auto &p_g : params_grads) {
       // Get gradient var
-      auto iter = vars.find(p_g.second);
-      PADDLE_ENFORCE(iter != vars.end(), "%s is not found.", p_g.second);
-      iter->second->Var()->SetPersistable(true);
+      auto iter = var_name2node.find(p_g.second);
+      PADDLE_ENFORCE(iter != var_name2node.end(), "%s is not found.",
+                     p_g.second);
+      // Set persistable
+      auto same_nodes = var_name2node_set.find(p_g.second);
+      PADDLE_ENFORCE(same_nodes != var_name2node_set.end(), "%s is not found.",
+                     p_g.second);
+      for (auto it : same_nodes->second) {
+        it->Var()->SetPersistable(true);
+      }
 
       PADDLE_ENFORCE(IsSupportedVarType(iter->second->Var()->GetType()));
 
@@ -151,7 +162,7 @@ class AllocContinuousSpaceForGradPass : public ir::Pass {
                       "%s is duplicate in FusedVars.", fused_var_name);
     fused_var_set.insert(fused_var_name);
 
-    InitFusedVarsAndAllocSpaceForVars(places, local_scopes, vars,
+    InitFusedVarsAndAllocSpaceForVars(places, local_scopes, var_name2node,
                                       fused_var_name, params_grads);
   }
 

paddle/fluid/framework/ir/graph_helper.cc

@@ -103,6 +103,33 @@ bool HasCircle(const Graph &graph) {
   return HasCircleInternal(BuildOperationAdjList(graph), nullptr);
 }
 
+bool VarDescIsConsistency(const Graph &graph) {
+  std::unordered_map<std::string, std::unordered_set<ir::Node *>>
+      var_name2node_set;
+  for (ir::Node *node : graph.Nodes()) {
+    if (node->IsVar() && node->Var()) {
+      // Note: The graph may have the same name node. For example, parameter
+      // is the input of operator and it also is the output of optimizer;
+      var_name2node_set[node->Var()->Name()].emplace(node);
+    }
+  }
+  for (auto &iter : var_name2node_set) {
+    auto &first_node = *iter.second.begin();
+    bool is_persistable = std::any_of(iter.second.begin(), iter.second.end(),
+                                      [&first_node](const ir::Node *node) {
+                                        return node->Var()->Persistable();
+                                      });
+    if (is_persistable) {
+      bool is_consistency =
+          std::all_of(iter.second.begin(), iter.second.end(),
+                      [&first_node](const ir::Node *node) {
+                        return *node->Var() == *first_node->Var();
+                      });
+      if (!is_consistency) return false;
+    }
+  }
+  return true;
+}
 bool FindCircleSubGraph(const Graph &graph,
                         std::vector<std::vector<ir::Node *>> *circles) {
   return HasCircleInternal(BuildOperationAdjList(graph), circles);

paddle/fluid/framework/ir/graph_helper.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <map>
 #include <memory>
 #include <set>
+#include <string>
 #include <vector>
 
 #include "paddle/fluid/framework/ir/graph.h"
@@ -36,6 +37,9 @@ struct NodeComp {
 // Test if the graph contains circle.
 bool HasCircle(const Graph &graph);
 
+// Check if the var desc of node is consistency.
+bool VarDescIsConsistency(const Graph &graph);
+
 // Find All Circles for debugging,
 // store all subgraph in circles.
 bool FindCircleSubGraph(const Graph &graph,

paddle/fluid/framework/ir/pass.cc

@@ -38,6 +38,8 @@ Graph* Pass::Apply(Graph* graph) const {
   // TODO(panyx0718): Add more verifications.
   PADDLE_ENFORCE(!HasCircle(*graph),
                  "Illegal Pass. Generated graph shouldn't has cycle.");
+  PADDLE_ENFORCE(VarDescIsConsistency(*graph),
+                 "The VarDescs of persistable variable are not consistency.");
   PADDLE_ENFORCE(graph == native_graph,
                  "Pass::Apply() cannot delete the passed graph and shouldn't "
                  "return a new graph.(For the need of pybind11)");

paddle/fluid/framework/parallel_executor.cc

@@ -320,12 +320,14 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
   }
 #endif
   if (!member_->use_all_reduce_) {
-    PADDLE_ENFORCE(places.size() > 1,
-                   "If you set build_strategy.reduce with 'Reduce',"
-                   "the number of places must be greater than 1.");
+    if (places.size() == 1) {
+      LOG(INFO) << "If you set build_strategy.reduce with 'Reduce',"
+                   "the number of places should be greater than 1.";
+      member_->use_all_reduce_ = true;
+    }
   }
 
-  LOG(WARNING) << string::Sprintf(
+  LOG(INFO) << string::Sprintf(
       "The number of %s, which is used in ParallelExecutor, is %lu. And "
       "the Program will be copied %lu copies",
       (member_->use_cuda_ ? "CUDAPlace" : "CPUPlace"), places.size(),
@@ -364,10 +366,11 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
   // choice the execution strategy.
   build_strategy.enable_parallel_graph_ =
       EnableParallelGraphExecution(*graph, exec_strategy, build_strategy);
-  if (build_strategy.enable_parallel_graph_)
-    VLOG(0) << "The Executor would execute the graph by ParallelGraph "
-               "Execution which can get better performance,"
-            << "you can force it off by env FLAGS_enable_parallel_graph=0";
+  if (build_strategy.enable_parallel_graph_) {
+    LOG(INFO) << "The Executor would execute the graph by ParallelGraph "
+                 "Execution which can get better performance,"
+              << "you can force it off by env FLAGS_enable_parallel_graph=0";
+  }
 
   if (member_->use_cuda_ && member_->nranks_ > 1) {
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)

paddle/fluid/framework/var_desc.cc

@@ -264,5 +264,10 @@ std::vector<proto::VarType::TensorDesc *> VarDesc::mutable_tensor_descs() {
   }
 }
 
+bool operator==(const VarDesc &left, const VarDesc &right) {
+  return left.Proto()->SerializeAsString() ==
+         right.Proto()->SerializeAsString();
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/var_desc.h

@@ -67,6 +67,8 @@ class VarDesc {
 
   proto::VarDesc *Proto() { return &desc_; }
 
+  const proto::VarDesc *Proto() const { return &desc_; }
+
   std::string Name() const { return desc_.name(); }
 
   void SetName(std::string name) { desc_.set_name(name); }
@@ -116,5 +118,7 @@ class VarDesc {
 
   proto::VarDesc desc_;
 };
+
+bool operator==(const VarDesc &left, const VarDesc &right);
 }  // namespace framework
 }  // namespace paddle