[Inference] save_optimized_model_pass support gpu (#55551)

* fix cudnn 8.7+ bug on cudnnConvolutionBiasActivationForward

* save_optimized_model_pass support gpu

paddle/fluid/inference/analysis/passes/save_optimized_model_pass.cc

@@ -55,6 +55,9 @@ void SaveOptimizedModelPass::SaveOptimizedModel(Argument* argument) {
   auto* graph = argument->main_graph_ptr();
 
   framework::ProgramDesc optimized_program_desc;
+  // NOTE(liuyuanle): If the following line of code is not added, an error
+  // [SegmentFault] may occur!
+  optimized_program_desc.CopyFrom(*argument->main_program().Proto());
   framework::ir::GraphToProgram(*graph, &optimized_program_desc);
 
   auto IsPersistable = [](const framework::VarDesc* var) {
@@ -130,7 +133,9 @@ void SaveOptimizedModelPass::SaveOptimizedModel(Argument* argument) {
 }
 
 void SaveOptimizedModelPass::RunImpl(Argument* argument) {
-  if (argument->use_xpu_valid()) {
+  // TODO(inference): Support trt.
+  if (argument->use_xpu() ||
+      (argument->use_gpu() && !argument->use_tensorrt())) {
     SaveOptimizedModel(argument);
   }
 }

paddle/fluid/inference/api/analysis_predictor.cc

@@ -1379,6 +1379,7 @@ void AnalysisPredictor::PrepareArgument() {
   argument_->SetOptimInputShape(config_.optim_input_shape_);
   argument_->SetTensorRtTunedDynamicShape(
       config_.tuned_tensorrt_dynamic_shape());
+  argument_->SetUseTensorRT(false);
   if (config_.use_gpu() && config_.tensorrt_engine_enabled()) {
     LOG(INFO) << "TensorRT subgraph engine is enabled";
     argument_->SetUseTensorRT(true);

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -181,8 +181,6 @@ class PD_INFER_DECL PassStrategy : public PaddlePassBuilder {
   bool use_ipu_{false};
   bool use_mkldnn_{false};
   bool use_custom_device_{false};
-
-  bool use_gpu_low_precision_{false};
   /// \endcond
 };