single pserver workable version

paddle/operators/recv_op.cc

@@ -69,6 +69,9 @@ class RecvOp : public framework::OperatorBase {
     auto param_list = Attr<std::vector<std::string>>("ParamList");
     auto grad_list = Attr<std::vector<std::string>>("GradList");
     size_t param_count = param_list.size();
+    // TODO(typhoonzero): change this to a while_op for every cluster-batch.
+    while (true) {
+      // TODO(typhoonzero): get from multiple trainers.
       for (size_t i = 0; i < param_count; ++i) {
         // blocking get one var from client.
         const detail::TensorWithName &v = rpc_service_->Get();
@@ -106,6 +109,7 @@ class RecvOp : public framework::OperatorBase {
         out.second = out_var->Get<framework::LoDTensor>();
         rpc_service_->Push(out);
       }
+    }  // while(true)
   }
 
  protected:

python/paddle/v2/fluid/executor.py

@@ -93,7 +93,7 @@ class Executor(object):
             dtype=var.dtype,
             type=var.type,
             lod_level=var.lod_level,
-            persistable=True)
+            persistable=var.persistable)
 
     def _optimize_distributed(self, optimize_ops, program, params_and_grads,
                               **kwargs):