fix allReduce bug

Author: chengduoZH

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -107,7 +107,7 @@ void AllReduceOpHandle::RunImpl() {
       auto &trg = *this->local_scopes_[0]
                        ->FindVar(kLocalExecScopeName)
                        ->Get<Scope *>()
-                       ->Var()
+                       ->FindVar(in_var_handles[0]->name_)
                        ->GetMutable<framework::LoDTensor>();
 
       // Reduce All Tensor to trg in CPU

python/paddle/fluid/parallel_executor.py

@@ -121,8 +121,9 @@ def __init__(self,
                 # performance. Worth tunning for other models in the future.
                 exec_strategy.num_threads = len(self._places) * 4
             else:
-                # Currently num_threads must be 1.
-                exec_strategy.num_threads = 1
+                cpu_num = int(
+                    os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+                exec_strategy.num_threads = cpu_num
 
         if build_strategy is None:
             build_strategy = BuildStrategy()