Merge pull request #15840 from jacquesqiao/revert-15684-revert-15661-fix-cpu-broadcast

fix cpu broadcast

paddle/fluid/framework/details/build_strategy.cc

@@ -135,12 +135,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
   void AppendMultiDevPass(const BuildStrategy &strategy) {
     ir::Pass *multi_devices_pass;
     if (strategy_.is_distribution_) {
+      VLOG(3) << "multi device parameter server mode";
       multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
     } else {
       if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
+        VLOG(3) << "multi devices collective mode with allreduce";
         multi_devices_pass =
             AppendPass("allreduce_mode_multi_devices_pass").get();
       } else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
+        VLOG(3) << "multi deivces collective mode with reduce";
         multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
       } else {
         PADDLE_THROW("Unknown reduce strategy.");

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -937,9 +937,21 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
 }
 
 void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
-  if (need_broadcast_var_ ||
-      (UseGPU() &&
-       strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce)) {
+  // broad cast received parameters when training in parameter server mode.
+  if (need_broadcast_var_) {
+    // There are 4 conditions:
+    // 1. GPU && Reduce: Reduce gradient then broadcast gradient to other GPUS.
+    // Need to broadcast received parameters to other GPU.
+    // 2. GPU && AllReduce: AllReduce all graident to each GPU. Need to
+    // broadcast received parameters to other GPU.
+    // 3. CPU && AllReduce: AllReduce all gradient to each thread. Need to
+    // broadcast received parameters to other scope.
+    // 4. CPU && Reduce: because all parameters share the same memory, did not
+    // broadcast received parameters.
+    if (!UseGPU() &&
+        strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
+      return;
+    }
     if (strategy_.fuse_broadcast_op_) {
       CreateFusedBroadcastOp(result, bcast_var_name_set_);
     } else {

python/paddle/fluid/compiler.py

@@ -19,6 +19,7 @@ import sys
 from .. import compat as cpt
 
 from . import core
+from . import framework
 
 __all__ = ['CompiledProgram', 'ExecutionStrategy', 'BuildStrategy']
 
@@ -110,6 +111,8 @@ class CompiledProgram(object):
             self._exec_strategy = ExecutionStrategy()
         if self._build_strategy is None:
             self._build_strategy = BuildStrategy()
+        self._build_strategy.is_distribution = framework.is_pserver_mode(
+            self._program)
         return self
 
     def with_inference_optimize(self, config):

python/paddle/fluid/framework.py

@@ -87,6 +87,15 @@ def _current_expected_place():
     return _imperative_current_expected_place_
 
 
+def is_pserver_mode(main_program):
+    main = main_program if main_program \
+        else default_main_program()
+    for op in main.global_block().ops:
+        if op.type in ["send", "recv"]:
+            return True
+    return False
+
+
 class NameScope(object):
     def __init__(self, name="", parent=None):
         self._children = dict()

python/paddle/fluid/parallel_executor.py

@@ -29,15 +29,6 @@ ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy
 BuildStrategy = core.ParallelExecutor.BuildStrategy
 
 
-def _is_pserver_mode(main_program):
-    main = main_program if main_program \
-        else framework.default_main_program()
-    for op in main.global_block().ops:
-        if op.type in ["send", "recv"]:
-            return True
-    return False
-
-
 class ParallelExecutor(object):
     """
     ParallelExecutor is designed for data parallelism, which focuses on distributing
@@ -140,7 +131,7 @@ class ParallelExecutor(object):
         # FIXME(zcd): is_distribution_ is a temporary field, because in pserver mode,
         # num_trainers is 1, so the current fields of build_strategy doesn't tell if
         # it's distributed model.
-        build_strategy.is_distribution = _is_pserver_mode(
+        build_strategy.is_distribution = framework.is_pserver_mode(
             main_program) or num_trainers > 1
 
         # step4: get main_program, scope, local_scopes