fix onnxruntime bug (#42095)

paddle/fluid/inference/api/details/zero_copy_tensor.cc

@@ -693,10 +693,9 @@ void Tensor::ORTCopyToCpu(T *data) const {
   if (place_ == PlaceType::kCPU) {
     std::memcpy(static_cast<void *>(data), value.GetTensorData<void *>(), size);
   } else {
-    paddle::memory::Copy(paddle::platform::CPUPlace(),
-                         static_cast<void *>(data),
-                         paddle::platform::CUDAPlace(device_),
-                         value.GetTensorData<void>(), size, nullptr);
+    PADDLE_THROW(paddle::platform::errors::Unavailable(
+        "CopyToCpu error.The current ONNXRuntime backend doesn't support "
+        "GPU."));
   }
 }
 

paddle/fluid/inference/api/onnxruntime_predictor.cc

@@ -279,6 +279,12 @@ bool ONNXRuntimePredictor::Run(const std::vector<PaddleTensor> &inputs,
 
 bool ONNXRuntimePredictor::ZeroCopyRun() {
   try {
+    const char *device_name = place_ == PlaceType::kCPU ? "Cpu" : "Cuda";
+    for (auto output : output_desc_) {
+      Ort::MemoryInfo out_memory_info(device_name, OrtDeviceAllocator,
+                                      place_.GetDeviceId(), OrtMemTypeDefault);
+      binding_->BindOutput(output.name.c_str(), out_memory_info);
+    }
     session_.Run({}, *(binding_.get()));
   } catch (const std::exception &e) {
     LOG(ERROR) << e.what();