[inference] Use output var name to mark the NVTX flag (#49825)

* add outvar name for nvtx mark

* nly network created with kEXPLICIT_BATCH can setsetMaxBatchSize

paddle/fluid/framework/naive_executor.cc

@@ -62,7 +62,8 @@ void NaiveExecutor::Run() {
             << op->DebugStringEx(scope_) << " on scope " << scope_;
     op->SetIsCalledByExecutor(false);
 #ifdef PADDLE_WITH_INFERENCE_NVTX
-    platform::CudaNvtxRangePush(op->Type(), platform::NvtxRangeColor::Green);
+    platform::CudaNvtxRangePush(op->Type() + "|" + op->OutputVars(true).front(),
+                                platform::NvtxRangeColor::Green);
 #endif
 
     // According to reuse table, we share the out tensor's holder.

paddle/fluid/inference/tensorrt/engine.cc

@@ -148,7 +148,9 @@ void TensorRTEngine::FreezeNetwork() {
                           platform::errors::InvalidArgument(
                               "Call InitNetwork first to initialize network."));
   // build engine.
-  infer_builder_->setMaxBatchSize(max_batch_);
+  if (!with_dynamic_shape_) {
+    infer_builder_->setMaxBatchSize(max_batch_);
+  }
 #if IS_TRT_VERSION_GE(8300)
   infer_builder_config_->setMemoryPoolLimit(
       nvinfer1::MemoryPoolType::kWORKSPACE, max_workspace_);