PaddlePaddle
diff --git a/‎paddle/fluid/framework/compiled_program.cc
Copy file name to clipboard
-10Lines changed: 0 additions & 10 deletions b/‎paddle/fluid/framework/compiled_program.cc
Copy file name to clipboard
-10Lines changed: 0 additions & 10 deletions
diff --git a/‎paddle/fluid/framework/details/CMakeLists.txt
Copy file name to clipboard
-6Lines changed: 0 additions & 6 deletions b/‎paddle/fluid/framework/details/CMakeLists.txt
Copy file name to clipboard
-6Lines changed: 0 additions & 6 deletions
diff --git a/‎paddle/fluid/framework/details/build_strategy.cc
Copy file name to clipboard
100755100644
+11-49Lines changed: 11 additions & 49 deletions b/‎paddle/fluid/framework/details/build_strategy.cc
Copy file name to clipboard
100755100644
+11-49Lines changed: 11 additions & 49 deletions
diff --git a/‎paddle/fluid/framework/details/build_strategy.h
Copy file name to clipboard
-10Lines changed: 0 additions & 10 deletions b/‎paddle/fluid/framework/details/build_strategy.h
Copy file name to clipboard
-10Lines changed: 0 additions & 10 deletions
diff --git a/‎paddle/fluid/framework/details/fetch_barrier_op_handle.cc
Copy file name to clipboard
-63Lines changed: 0 additions & 63 deletions b/‎paddle/fluid/framework/details/fetch_barrier_op_handle.cc
Copy file name to clipboard
-63Lines changed: 0 additions & 63 deletions
diff --git a/‎paddle/fluid/framework/details/fetch_barrier_op_handle.h
Copy file name to clipboard
-77Lines changed: 0 additions & 77 deletions b/‎paddle/fluid/framework/details/fetch_barrier_op_handle.h
Copy file name to clipboard
-77Lines changed: 0 additions & 77 deletions
@@ -75,15 +75,6 @@ class CompiledProgramPrivate {
 
   ir::Graph *ApplyMemoryOptimizePass(ir::Graph *graph);
 
-  inline bool HasGarbageCollectors() const { return !gcs_.empty(); }
-
-  void ApplyFixOpRunOrderPass(ir::Graph *graph) {
-    if (build_strategy_.fix_op_run_order_) {
-      auto pass = ir::PassRegistry::Instance().Get("fix_op_run_order_pass");
-      pass->Apply(graph);
-    }
-  }
-
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
   void InitNCCLCtxs(framework::Scope *scope, const BuildStrategy &bst) {
     VLOG(1) << "nccl comm num:" << bst.nccl_comm_num_ << ", nranks:" << nranks_
@@ -593,7 +584,6 @@ CompiledProgram::CompiledProgram(const std::vector<platform::Place> &places,
   // ncclOp
   std::vector<ir::Graph *> async_graphs =
       CompileGraphWithBuildStrategy(graph, &graphs, loss_var_name);
-  // PrepareForCUDAGraphCapture(graph);
   graph = member_->ApplyMemoryOptimizePass(graph);
 }
 
 
@@ -22,11 +22,7 @@ set(op_handle_srcs
     share_tensor_buffer_functor.cc
     computation_op_handle.cc
     share_tensor_buffer_op_handle.cc
-    rpc_op_handle.cc
-    fetch_barrier_op_handle.cc
-    multi_devices_helper.cc
     variable_visitor.cc
-    gather_op_handle.cc
     eager_deletion_op_handle.cc)
 
 if(WITH_DGC)
@@ -81,7 +77,6 @@ set(IR_PASS_DEPS
     multi_batch_merge_pass
     fuse_relu_depthwise_conv_pass
     lock_free_optimize_pass
-    sequential_execution_pass
     add_reader_dependency_pass
     modify_op_lock_and_record_event_pass
     coalesce_grad_tensor_pass
@@ -91,7 +86,6 @@ set(IR_PASS_DEPS
     sync_batch_norm_pass
     runtime_context_cache_pass
     graph_to_program_pass
-    fix_op_run_order_pass
     fuse_gemm_epilogue_pass
     fused_attention_pass
     fuse_adamw_op_pass
 
@@ -30,16 +30,6 @@ PD_DECLARE_bool(use_cinn);
 
 namespace paddle::framework::details {
 
-static inline bool SeqOnlyAllReduceOps(const BuildStrategy &strategy) {
-  // Should fix the allreduce op order if scheduling
-  // them in multiple threads or processes to avoid hang.
-  // NOTE: ParallelGraph would execute this pass on each graph, so
-  // don't need to append it here.
-  return (!strategy.enable_sequential_execution_ &&
-          strategy.num_trainers_ > 1) &&
-         !strategy.enable_parallel_graph_;
-}
-
 static inline void ConvertDefaultValue(paddle::optional<bool> *default_value) {
   if (*default_value == paddle::none) {
     *default_value = true;
@@ -62,8 +52,6 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
     }
 #endif
 
-    AppendPassWithCheck(strategy_.enable_sequential_execution_,
-                        "sequential_execution_pass");
     AppendPassWithCheck(strategy_.sync_batch_norm_, "sync_batch_norm_pass");
 
     AppendOpFusePasses();
@@ -95,16 +83,6 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
              "parallel_graph.";
       strategy_.fuse_all_reduce_ops_ = false;
     }
-    if (strategy_.is_distribution_) {
-      LOG_IF(WARNING, strategy_.fuse_all_optimizer_ops_ == true)
-          << "Currently, fuse_all_optimizer_ops only works under "
-             "Non-distributed mode.";
-      strategy_.fuse_all_optimizer_ops_ = false;
-      LOG_IF(WARNING, strategy_.fuse_all_reduce_ops_ == true)
-          << "Currently, fuse_all_reduce_ops_ only works under "
-             "Non-distributed mode.";
-      strategy_.fuse_all_reduce_ops_ = false;
-    }
     if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
       LOG_IF(WARNING, strategy_.fuse_all_optimizer_ops_ == true)
           << "Currently, fuse_all_optimizer_ops only works under AllReduce "
@@ -227,27 +205,17 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
   // Convert graph to run on multi-devices.
   void AppendMultiDevPass() {
     ir::Pass *multi_devices_pass = nullptr;
-    if (strategy_.async_mode_) {
-      multi_devices_pass = AppendPass("async_multi_devices_pass").get();
-    } else if (strategy_.is_distribution_) {
-      multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
-    } else {
-      switch (strategy_.reduce_) {
-        case BuildStrategy::ReduceStrategy::kAllReduce:
-          multi_devices_pass =
-              AppendPass("all_reduce_mode_multi_devices_pass").get();
-          break;
-        case BuildStrategy::ReduceStrategy::kReduce:
-          multi_devices_pass =
-              AppendPass("reduce_mode_multi_devices_pass").get();
-          break;
-        case BuildStrategy::ReduceStrategy::kNoReduce:
-          multi_devices_pass = AppendPass("no_reduce_multi_devices_pass").get();
-          break;
-        default:
-          PADDLE_THROW(
-              platform::errors::Unimplemented("Unknown reduce strategy."));
-      }
+    switch (strategy_.reduce_) {
+      case BuildStrategy::ReduceStrategy::kAllReduce:
+        multi_devices_pass =
+            AppendPass("all_reduce_mode_multi_devices_pass").get();
+        break;
+      case BuildStrategy::ReduceStrategy::kReduce:
+        multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
+        break;
+      default:
+        PADDLE_THROW(
+            platform::errors::Unimplemented("Unknown reduce strategy."));
     }
     multi_devices_pass->SetNotOwned<const BuildStrategy>("strategy",
                                                          &strategy_);
@@ -367,9 +335,6 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
     } else if (pass->Type() == "coalesce_grad_tensor_pass") {
       pass->Erase(kNRanks);
       pass->Set<size_t>(kNRanks, new size_t(nranks));
-    } else if (pass->Type() == "sequential_execution_pass") {
-      LOG(INFO) << "set enable_sequential_execution:"
-                << enable_sequential_execution_;
     } else if (pass->Type() == "fuse_relu_depthwise_conv_pass") {
       if (use_device != p::kCUDA) {
         VLOG(1) << "fuse_relu_depthwise_conv_pass is only supported on "
@@ -422,11 +387,8 @@ USE_PASS(fuse_bn_act_pass);
 USE_PASS(fuse_bn_add_act_pass);
 USE_PASS(graph_viz_pass);
 USE_PASS(multi_batch_merge_pass);
-USE_PASS(no_reduce_multi_devices_pass);
 USE_PASS(reduce_mode_multi_devices_pass);
 USE_PASS(all_reduce_mode_multi_devices_pass);
-USE_PASS(dist_multi_devices_pass);
-USE_PASS(sequential_execution_pass);
 USE_PASS(modify_op_lock_and_record_event_pass);
 USE_PASS(lock_free_optimize_pass);
 USE_PASS(coalesce_grad_tensor_pass);
 
@@ -91,18 +91,11 @@ struct BuildStrategy {
   // all the backward ops are finished before running the optimization
   // ops. It might make the training speed of data parallelism faster.
   bool enable_backward_optimizer_op_deps_{true};
-  // TODO(dev-paddle): enable_sequential_execution depends on
-  // kStaleProgramOpDescs, it is not appropriate, because kStaleProgramOpDescs
-  // will be removed in the near future.
-  bool enable_sequential_execution_{false};
   bool remove_unnecessary_lock_{true};
   // TODO(dev-paddle): cache_runtime_context may cause some models to hang up
   // while running.
   bool cache_runtime_context_{false};
 
-  // Fix the op run order.
-  bool fix_op_run_order_{false};
-
   // Lowering sub-graph into cinn ops.
   bool build_cinn_pass_{false};
 
@@ -251,13 +244,10 @@ inline std::ostream &operator<<(std::ostream &os,
   os << "debug_graphviz_path_: " << strategy.debug_graphviz_path_ << std::endl;
   os << "enable_backward_optimizer_op_deps_: "
      << strategy.enable_backward_optimizer_op_deps_ << std::endl;
-  os << "enable_sequential_execution_: "
-     << strategy.enable_sequential_execution_ << std::endl;
   os << "remove_unnecessary_lock_: " << strategy.remove_unnecessary_lock_
      << std::endl;
   os << "cache_runtime_context_: " << strategy.cache_runtime_context_
      << std::endl;
-  os << "fix_op_run_order_: " << strategy.fix_op_run_order_ << std::endl;
   os << "fuse_bn_act_ops_: " << strategy.fuse_bn_act_ops_ << std::endl;
   os << "fuse_bn_add_act_ops_: " << strategy.fuse_bn_add_act_ops_ << std::endl;
   os << "fuse_elewise_add_act_ops_: " << strategy.fuse_elewise_add_act_ops_