Support disable GC for some vars in interpretercore (#43546)

* Support disable GC for some vars in standalone executor

* Setting skip_gc_vars in interprecore construction

paddle/fluid/framework/new_executor/interpretercore.cc

@@ -56,12 +56,15 @@ bool IsInterpretercoreFastGCEnabled() {
 
 InterpreterCore::InterpreterCore(const platform::Place& place,
                                  const BlockDesc& block,
+                                 const std::set<std::string>& skip_gc_vars,
                                  VariableScope* global_scope)
     : place_(place),
       block_(block),
+      skip_gc_vars_(skip_gc_vars),
       global_scope_(global_scope),
       stream_analyzer_(place) {
   VLOG(4) << "InterpreterCore(): " << this << " on " << place_;
+
   is_build_ = false;
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -173,7 +176,8 @@ paddle::framework::FetchList InterpreterCore::Run(
                                                          create_local_scope_);
     std::vector<paddle::framework::OpFuncNode> op_func_nodes;
     paddle::framework::interpreter::build_op_func_list(
-        place_, block_, &op_func_nodes, global_scope_, create_local_scope_);
+        place_, block_, skip_gc_vars_, &op_func_nodes, global_scope_,
+        create_local_scope_);
     is_build_ = true;
     SetFeedVarsInplaceSkip(feed_names);
     // convert vec func_list to graph
@@ -309,6 +313,15 @@ void InterpreterCore::Convert(
     }
   }
 
+  // clear the last_live_ops list for all vars in skip_gc_vars
+  for (const std::string& skip_gc_var : skip_gc_vars_) {
+    int var_id = global_scope_->GetIdByName(skip_gc_var);
+    if (var_id != -1) {
+      last_live_ops_[var_id].clear();
+      VLOG(8) << "Skip gc for var: " << skip_gc_var;
+    }
+  }
+
   // shrink, find the downstream op that has no other op in the
   // downstream list happens before it
   // For example,
@@ -934,7 +947,8 @@ void InterpreterCore::Prepare(
     FeedInput();
     std::vector<paddle::framework::OpFuncNode> op_func_nodes;
     paddle::framework::interpreter::build_op_func_list(
-        place_, block_, &op_func_nodes, global_scope_, create_local_scope_);
+        place_, block_, skip_gc_vars_, &op_func_nodes, global_scope_,
+        create_local_scope_);
     is_build_ = true;
     SetFeedVarsInplaceSkip(feed_names);
     // convert vec func_list to graph
@@ -986,5 +1000,22 @@ void InterpreterCore::SetFeedVarsInplaceSkip(
   }
 }
 
+std::shared_ptr<InterpreterCore> CreateInterpreterCore(
+    const platform::Place& place, const ProgramDesc& prog,
+    VariableScope* global_scope, const std::vector<std::string>& fetch_names,
+    const std::set<std::string>& skip_gc_vars) {
+  std::shared_ptr<InterpreterCore> core = nullptr;
+  // NOTE(Aurelius84): `add_fetch` will modify BlockDesc, so we should copy
+  // a new program.
+  auto new_prog = std::make_shared<framework::ProgramDesc>(prog);
+  auto* block = new_prog->MutableBlock(0);
+  interpreter::add_fetch(fetch_names, block);
+
+  core = std::make_shared<InterpreterCore>(place, *block, skip_gc_vars,
+                                           global_scope);
+  core->SetCopyProgram(new_prog);
+  return core;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/new_executor/interpretercore.h

@@ -39,6 +39,7 @@ using AtomicVectorSizeT = std::vector<std::unique_ptr<std::atomic<size_t>>>;
 class InterpreterCore {
  public:
   InterpreterCore(const platform::Place& place, const BlockDesc& block,
+                  const std::set<std::string>& skip_gc_vars,
                   VariableScope* global_scope);
 
   ~InterpreterCore();
@@ -99,6 +100,8 @@ class InterpreterCore {
 
   const platform::Place& place_;
   const BlockDesc& block_;  // not owned
+  const std::set<std::string> skip_gc_vars_;
+
   // NOTE(zhiqiu): when add fetch ops in GetInterpreterCore, we will
   // copy a new program and block, the copy_program_ here is used to
   // hold the program, otherwise block_ maybe not valid after the
@@ -130,5 +133,12 @@ class InterpreterCore {
   bool create_local_scope_{true};
   Scope* local_scope_{nullptr};  // not owned
 };
+
+std::shared_ptr<InterpreterCore> CreateInterpreterCore(
+    const platform::Place& place, const ProgramDesc& prog,
+    VariableScope* global_scope,
+    const std::vector<std::string>& fetch_names = {},
+    const std::set<std::string>& skip_gc_vars = {});
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/new_executor/interpretercore_util.cc

@@ -323,6 +323,7 @@ void deal_operator_base(const platform::Place& place,
 
 void build_op_func_list(const platform::Place& place,
                         const framework::BlockDesc& block,
+                        const std::set<std::string>& skip_gc_vars,
                         std::vector<OpFuncNode>* vec_func_list,
                         VariableScope* var_scope, bool use_local_scope) {
   Scope* local_scope = use_local_scope ? var_scope->GetMutableLocalScope()
@@ -562,7 +563,7 @@ void build_op_func_list(const platform::Place& place,
 
     for (auto& var_name : delete_vars) {
       auto* var = var_scope->FindVar(var_name);
-      if (var == nullptr) {
+      if (var == nullptr || skip_gc_vars.find(var_name) != skip_gc_vars.end()) {
         continue;
       }
 

paddle/fluid/framework/new_executor/interpretercore_util.h

@@ -114,6 +114,7 @@ void build_variable_scope(const framework::BlockDesc& block,
 
 void build_op_func_list(const platform::Place& place,
                         const framework::BlockDesc& block,
+                        const std::set<std::string>& skip_gc_vars,
                         std::vector<OpFuncNode>* vec_func_list,
                         VariableScope* var_scope, bool use_local_scope = true);
 

paddle/fluid/framework/new_executor/standalone_executor.cc

@@ -55,7 +55,8 @@ StandaloneExecutor::StandaloneExecutor(const platform::Place& place,
     // No need to use_local_scope for startup_program, its variables are
     // persistable
     paddle::framework::interpreter::build_op_func_list(
-        place_, startup_prog.Block(0), &vec_func_list, &global_scope_, false);
+        place_, startup_prog.Block(0), {}, &vec_func_list, &global_scope_,
+        false);
   }
 }
 
@@ -126,18 +127,13 @@ std::shared_ptr<InterpreterCore> StandaloneExecutor::GetInterpreterCore(
             << place_;
     VLOG(3) << "add fetch op: " << add_fetch_op;
     std::shared_ptr<InterpreterCore> core = nullptr;
+
     if (add_fetch_op) {
-      // NOTE(Aurelius84): `add_fetch` will modify BlockDesc, so we should copy
-      // a
-      // new program.
-      auto new_prog = std::make_shared<framework::ProgramDesc>(main_prog_);
-      auto* block = new_prog->MutableBlock(0);
-      interpreter::add_fetch(fetch_names, block);
-
-      core = std::make_shared<InterpreterCore>(place_, *block, &global_scope_);
-      core->SetCopyProgram(new_prog);
+      core = CreateInterpreterCore(place_, main_prog_, &global_scope_,
+                                   fetch_names);
     } else {
-      core = std::make_shared<InterpreterCore>(place_, main_prog_.Block(0),
+      core = std::make_shared<InterpreterCore>(
+          place_, main_prog_.Block(0), /*skip_gc_vars=*/std::set<std::string>(),
           &global_scope_);
     }
     interpretercores_.emplace(oss.str(), core);

paddle/fluid/framework/new_executor/standalone_executor_test.cc

@@ -145,9 +145,7 @@ TEST(StandaloneExecutor, run) {
   exec.Run({}, {}, {});
   auto start = std::chrono::steady_clock::now();
 
-  // ProfilerStart("new_executor.prof");
-
-  for (size_t i = 0; i < 2320; ++i) {
+  for (size_t i = 0; i < 10; ++i) {
     if (i % 200 == 0) {
       std::cout << i << std::endl;
     }
@@ -155,13 +153,71 @@ TEST(StandaloneExecutor, run) {
     exec.Run({}, {}, {});
   }
 
-  // ProfilerStop();
-
   auto end = std::chrono::steady_clock::now();
   std::chrono::duration<double> diff = end - start;
 
   std::cout << "time cost " << diff.count() << std::endl;
-  // ASSERT_LT(diff.count(), 30);
+}
+
+TEST(StandaloneExecutor, skip_gc_vars) {
+  FLAGS_eager_delete_tensor_gb = 0;
+
+  int64_t batch_size = 20;
+
+  auto place = platform::CUDAPlace(0);
+  auto startup_prog = load_from_file("lm_startup_program");
+  auto main_prog = load_from_file("lm_main_program");
+
+  auto& global_block = main_prog.Block(0);
+
+  auto& op1 = global_block.AllOps()[1];
+  auto shape1 = BOOST_GET_CONST(std::vector<int64_t>, op1->GetAttr("shape"));
+  shape1[0] = batch_size * 20;
+  op1->SetAttr("shape", shape1);
+
+  auto& op2 = global_block.AllOps()[2];
+  auto shape2 = BOOST_GET_CONST(std::vector<int64_t>, op2->GetAttr("shape"));
+  shape2[0] = batch_size;
+  op2->SetAttr("shape", shape2);
+
+  auto& op3 = global_block.AllOps()[3];
+  auto shape3 = BOOST_GET_CONST(std::vector<int64_t>, op3->GetAttr("shape"));
+  shape3[0] = batch_size;
+  op3->SetAttr("shape", shape3);
+
+  Scope scope;
+
+  VariableScope startup_scope(&scope);
+  std::shared_ptr<InterpreterCore> startup_core =
+      CreateInterpreterCore(place, startup_prog, &startup_scope);
+  startup_core->Run({}, {});
+
+  std::set<std::string> skip_gc_vars = {"uniform_0.tmp_0", "transpose_0.tmp_0",
+                                        "embedding_0.tmp_0", "slice_0.tmp_0",
+                                        "split_1.tmp_2"};
+  std::set<std::string> gc_vars = {"uniform_1.tmp_0", "matmul_0.tmp_0",
+                                   "split_0.tmp_0", "elementwise_add_0.tmp_0",
+                                   "tmp_0"};
+  auto check_gc_result = [](VariableScope& scope, std::set<std::string>& vars,
+                            bool is_skip_gc) {
+    for (const std::string& var_name : vars) {
+      ASSERT_EQ(
+          scope.FindVar(var_name)->GetMutable<LoDTensor>()->IsInitialized(),
+          is_skip_gc);
+    }
+  };
+
+  VariableScope main_scope(&scope);
+  std::shared_ptr<InterpreterCore> main_core =
+      CreateInterpreterCore(place, main_prog, &main_scope, {}, skip_gc_vars);
+
+  main_core->Run({}, {});
+  check_gc_result(main_scope, skip_gc_vars, true);
+  check_gc_result(main_scope, gc_vars, false);
+
+  main_core->Run({}, {});
+  check_gc_result(main_scope, skip_gc_vars, true);
+  check_gc_result(main_scope, gc_vars, false);
 }
 
 }  // namespace framework