Merge pull request #1436 from ShiningZhang/dev

support set xpu id

paddle_inference/paddle/include/paddle_engine.h

@@ -266,6 +266,7 @@ class PaddleInferenceEngine : public EngineCore {
     if (engine_conf.has_use_xpu() && engine_conf.use_xpu()) {
       // 2 MB l3 cache
       config.EnableXpu(2 * 1024 * 1024);
+      config.SetXpuDeviceId(gpu_id);
     }
 
     if (engine_conf.has_enable_memory_optimization() &&

python/paddle_serving_app/local_predict.py

@@ -219,6 +219,7 @@ class LocalPredictor(object):
         if use_xpu:
             # 2MB l3 cache
             config.enable_xpu(8 * 1024 * 1024)
+            config.set_xpu_device_id(gpu_id)
         # set cpu low precision
         if not use_gpu and not use_lite:
             if precision_type == paddle_infer.PrecisionType.Int8:

python/pipeline/local_service_handler.py

@@ -280,6 +280,10 @@ class LocalServiceHandler(object):
                 server.set_gpuid(gpuid)
             # TODO: support arm or arm + xpu later
             server.set_device(self._device_name)
+            if self._use_xpu:
+                server.set_xpu()
+            if self._use_lite:
+                server.set_lite()
 
         server.set_op_sequence(op_seq_maker.get_op_sequence())
         server.set_num_threads(thread_num)