fix graph output address set in the one time memory application scenarios

mindspore/ccsrc/backend/session/gpu_session.cc

@@ -481,7 +481,9 @@ void GPUSession::UpdateOutputTensors(const VectorRef *outputs,
 
         // When the device address of graph cnode output is set in tensor, the graph output need be set new device
         // address, to avoid that the device address context of tensor be rewritten in the next step or next loop.
-        if (node->isa<CNode>()) {
+        // But one time memory application scenarios need to be skipped, because the memory is not allocated next step:
+        // 1. Non cnode 2. Communication kernel.
+        if (node->isa<CNode>() && !AnfAlgo::IsCommunicationOp(node)) {
           auto new_address = std::make_shared<device::gpu::GPUDeviceAddress>(nullptr, address->GetSize());
           AnfAlgo::SetOutputAddr(new_address, output_index, node.get());
           if (context::GraphKernelFlags::GetInstance().IsEnableGraphKernel()) {

mindspore/ccsrc/runtime/device/gpu/kernel_info_setter.h

@@ -25,7 +25,6 @@
 #include "ir/anf.h"
 #include "ir/dtype.h"
 #include "utils/utils.h"
-#include "frontend/operator/ops.h"
 #include "backend/kernel_compiler/kernel.h"
 #include "backend/session/kernel_graph.h"