cleanup test=develop (#15347)

paddle/fluid/framework/details/execution_strategy.h

@@ -25,7 +25,10 @@ struct ExecutionStrategy {
   size_t num_threads_{0};
   bool use_cuda_{true};
   bool allow_op_delay_{false};
-  size_t num_iteration_per_drop_scope_{1};
+  // If we set this to 1, we will delete all variables when finish a batch. and
+  // this will loss 15%+ performance.
+  // Please be aware about this parameters.
+  size_t num_iteration_per_drop_scope_{100};
   ExecutorType type_{kDefault};
   bool dry_run_{false};
 };

paddle/fluid/operators/distributed/variable_response.cc

@@ -117,8 +117,9 @@ bool VariableResponse::CopyLodTensorData(
       tensor->mutable_data(ctx.GetPlace(), ToVarType(meta_.data_type()));
 
   VLOG(6) << "Tensor.memory_size = " << tensor->memory_size()
-          << ", Buffer Size = " << length;
-  PADDLE_ENFORCE_EQ(tensor->memory_size(), static_cast<unsigned int>(length));
+          << ", Buffer Size = " << length << ", dims:" << dims
+          << ", numel:" << tensor->numel();
+  PADDLE_ENFORCE_GE(tensor->memory_size(), static_cast<unsigned int>(length));
   return ReadRaw(input, ctx, tensor->place(), tensor_data, length);
 }