diff --git a/inferencer-fluid-gpu/include/fluid_gpu_engine.h b/inferencer-fluid-gpu/include/fluid_gpu_engine.h
index e2abe8e1ed60a82a18f1846cd30dc66a59083efd..64fe14fba0e06adada1ab61e15198bcc10db5457 100644
--- a/inferencer-fluid-gpu/include/fluid_gpu_engine.h
+++ b/inferencer-fluid-gpu/include/fluid_gpu_engine.h
@@ -32,6 +32,8 @@
 #endif
 #include "predictor/framework/infer.h"
 
+DECLARE_int32(gpuid);
+
 namespace baidu {
 namespace paddle_serving {
 namespace fluid_gpu {
@@ -129,7 +131,7 @@ class FluidGpuAnalysisCore : public FluidFamilyCore {
     paddle::AnalysisConfig analysis_config;
     analysis_config.SetParamsFile(data_path + "/__params__");
     analysis_config.SetProgFile(data_path + "/__model__");
-    analysis_config.EnableUseGpu(100, 0);
+    analysis_config.EnableUseGpu(100, FLAGS_gpuid);
     analysis_config.SetCpuMathLibraryNumThreads(1);
     analysis_config.SwitchSpecifyInputNames(true);
     AutoLock lock(GlobalPaddleCreateMutex::instance());
@@ -159,7 +161,7 @@ class FluidGpuNativeCore : public FluidFamilyCore {
     native_config.prog_file = data_path + "/__model__";
     native_config.use_gpu = true;
     native_config.fraction_of_gpu_memory = 0.9;
-    native_config.device = 0;
+    native_config.device = FLAGS_gpuid;
     AutoLock lock(GlobalPaddleCreateMutex::instance());
     _core = paddle::CreatePaddlePredictor<paddle::NativeConfig,
                                           paddle::PaddleEngineKind::kNative>(
@@ -185,7 +187,7 @@ class FluidGpuAnalysisDirCore : public FluidFamilyCore {
 
     paddle::AnalysisConfig analysis_config;
     analysis_config.SetModel(data_path);
-    analysis_config.EnableUseGpu(100, 0);
+    analysis_config.EnableUseGpu(100, FLAGS_gpuid);
     analysis_config.SwitchSpecifyInputNames(true);
     analysis_config.SetCpuMathLibraryNumThreads(1);
     AutoLock lock(GlobalPaddleCreateMutex::instance());
@@ -214,7 +216,7 @@ class FluidGpuNativeDirCore : public FluidFamilyCore {
     native_config.model_dir = data_path;
     native_config.use_gpu = true;
     native_config.fraction_of_gpu_memory = 0.9;
-    native_config.device = 0;
+    native_config.device = FLAGS_gpuid;
     AutoLock lock(GlobalPaddleCreateMutex::instance());
     _core = paddle::CreatePaddlePredictor<paddle::NativeConfig,
                                           paddle::PaddleEngineKind::kNative>(
@@ -464,7 +466,7 @@ class FluidGpuNativeDirWithSigmoidCore : public FluidGpuWithSigmoidCore {
     native_config.model_dir = data_path;
     native_config.use_gpu = true;
     native_config.fraction_of_gpu_memory = 0.9;
-    native_config.device = 0;
+    native_config.device = FLAGS_gpuid;
     AutoLock lock(GlobalPaddleCreateMutex::instance());
     _core->_fluid_core =
         paddle::CreatePaddlePredictor<paddle::NativeConfig,
@@ -491,7 +493,7 @@ class FluidGpuAnalysisDirWithSigmoidCore : public FluidGpuWithSigmoidCore {
 
     paddle::AnalysisConfig analysis_config;
     analysis_config.SetModel(data_path);
-    analysis_config.EnableUseGpu(100, 0);
+    analysis_config.EnableUseGpu(100, FLAGS_gpuid);
     analysis_config.SwitchSpecifyInputNames(true);
     analysis_config.SetCpuMathLibraryNumThreads(1);
     AutoLock lock(GlobalPaddleCreateMutex::instance());
diff --git a/inferencer-fluid-gpu/src/fluid_gpu_engine.cpp b/inferencer-fluid-gpu/src/fluid_gpu_engine.cpp
index 804473d723442af24247a6f06415d7cf61c77d25..70e89b3127724756eae82651a5ac2a9007f00e90 100644
--- a/inferencer-fluid-gpu/src/fluid_gpu_engine.cpp
+++ b/inferencer-fluid-gpu/src/fluid_gpu_engine.cpp
@@ -15,6 +15,8 @@
 #include "inferencer-fluid-gpu/include/fluid_gpu_engine.h"
 #include "predictor/framework/factory.h"
 
+DEFINE_int32(gpuid, 0, "GPU device id to use");
+
 namespace baidu {
 namespace paddle_serving {
 namespace fluid_gpu {