[Serving]ppcls preprocessor support gpu (#615)

* serving ppcls support gpu

* serving ppcls preprocessor use cpu

examples/vision/classification/paddleclas/README.md

@@ -54,3 +54,4 @@ PaddleClas模型导出，请参考其文档说明[模型导出](https://github.c
 
 - [Python部署](python)
 - [C++部署](cpp)
+- [服务化部署](serving)

examples/vision/classification/paddleclas/serving/models/preprocess/1/model.py

@@ -68,6 +68,9 @@ class TritonPythonModel:
             __file__)) + "/inference_cls.yaml"
         self.preprocess_ = fd.vision.classification.PaddleClasPreprocessor(
             yaml_path)
+        if args['model_instance_kind'] == 'GPU':
+            device_id = int(args['model_instance_device_id'])
+            self.preprocess_.use_gpu(device_id)
 
     def execute(self, requests):
         """`execute` must be implemented in every Python model. `execute`

examples/vision/classification/paddleclas/serving/models/preprocess/config.pbtxt

@@ -20,7 +20,11 @@ output [
 
 instance_group [
   {
-      count: 1
+    # The number of instances is 1
-      kind: KIND_CPU
+    count: 1
+    # Use CPU, GPU inference option is:KIND_GPU
+    kind: KIND_CPU
+    # The instance is deployed on the 0th GPU card
+    # gpus: [0]
   }
 ]

fastdeploy/pybind/fd_tensor.cc

@@ -151,8 +151,7 @@ pybind11::capsule FDTensorToDLPack(FDTensor& fd_tensor) {
 
   dlpack_tensor->dl_tensor.dtype = FDToDlpackType(fd_tensor.dtype);
 
-  // TODO(liqi): FDTensor add device_id
+  dlpack_tensor->dl_tensor.device.device_id = fd_tensor.device_id;
-  dlpack_tensor->dl_tensor.device.device_id = 0;
   if(fd_tensor.device == Device::GPU) {
     if (fd_tensor.is_pinned_memory) {
       dlpack_tensor->dl_tensor.device.device_type = DLDeviceType::kDLCUDAHost;

fastdeploy/vision/classification/ppcls/ppcls_pybind.cc

@@ -27,8 +27,10 @@ void BindPaddleClas(pybind11::module& m) {
         if (!self.Run(&images, &outputs)) {
           pybind11::eval("raise Exception('Failed to preprocess the input data in PaddleClasPreprocessor.')");
         }
-        for (size_t i = 0; i < outputs.size(); ++i) {
+        if (!self.WithGpu()) {
-          outputs[i].StopSharing();
+          for (size_t i = 0; i < outputs.size(); ++i) {
+            outputs[i].StopSharing();
+          }
         }
         return outputs;
       })

fastdeploy/vision/classification/ppcls/preprocessor.h

@@ -44,6 +44,8 @@ class FASTDEPLOY_DECL PaddleClasPreprocessor {
    */
   void UseGpu(int gpu_id = -1);
 
+  bool WithGpu() { return use_cuda_; }
+
  private:
   bool BuildPreprocessPipelineFromConfig(const std::string& config_file);
   std::vector<std::shared_ptr<Processor>> processors_;