From dea3795930ef61a60d6fd00277b59f9bdff42014 Mon Sep 17 00:00:00 2001
From: DefTruth <31974251+DefTruth@users.noreply.github.com>
Date: Sun, 25 Jun 2023 19:27:36 +0800
Subject: [PATCH] [Backend] Add backward compatible for paddle inference 2.4.x
 (#2062)

* [Backend] Add backward compatiable for paddle infernence 2.4.x

* [Backend] Add backward compatiable for paddle infernence 2.4.x
---
 cmake/paddle_inference.cmake                  | 16 ++++++++
 .../runtime/backends/paddle/paddle_backend.cc | 37 ++++++++++++++++++-
 2 files changed, 51 insertions(+), 2 deletions(-)

diff --git a/cmake/paddle_inference.cmake b/cmake/paddle_inference.cmake
index bf79ad820..637256971 100755
--- a/cmake/paddle_inference.cmake
+++ b/cmake/paddle_inference.cmake
@@ -22,6 +22,7 @@ endif()
 
 # Custom options for Paddle Inference backend
 option(PADDLEINFERENCE_DIRECTORY "Directory of custom Paddle Inference library" OFF)
+option(PADDLEINFERENCE_API_COMPAT_2_4_x "Whether using Paddle Inference 2.4.x" OFF)
 
 set(PADDLEINFERENCE_PROJECT "extern_paddle_inference")
 set(PADDLEINFERENCE_PREFIX_DIR ${THIRD_PARTY_PATH}/paddle_inference)
@@ -247,3 +248,18 @@ function(set_paddle_encrypt_auth_link_policy LIBRARY_NAME)
     endif()
   endif()
 endfunction()
+
+# Backward compatible for 2.4.x
+string(FIND ${PADDLEINFERENCE_VERSION} "2.4" PADDLEINFERENCE_USE_2_4_x)
+string(FIND ${PADDLEINFERENCE_VERSION} "2.5" PADDLEINFERENCE_USE_2_5_x)
+string(FIND ${PADDLEINFERENCE_VERSION} "0.0.0" PADDLEINFERENCE_USE_DEV)
+
+if((NOT (PADDLEINFERENCE_USE_2_4_x EQUAL -1)) 
+    AND (PADDLEINFERENCE_USE_2_5_x EQUAL -1) AND (PADDLEINFERENCE_USE_DEV EQUAL -1))
+  set(PADDLEINFERENCE_API_COMPAT_2_4_x ON CACHE BOOL "" FORCE)
+  message(WARNING "You are using PADDLEINFERENCE_USE_2_4_x:${PADDLEINFERENCE_VERSION}, force PADDLEINFERENCE_API_COMPAT_2_4_x=ON")
+endif()
+
+if(PADDLEINFERENCE_API_COMPAT_2_4_x)
+  add_definitions(-DPADDLEINFERENCE_API_COMPAT_2_4_x)
+endif()
diff --git a/fastdeploy/runtime/backends/paddle/paddle_backend.cc b/fastdeploy/runtime/backends/paddle/paddle_backend.cc
index 81d8a1f37..9965bae73 100644
--- a/fastdeploy/runtime/backends/paddle/paddle_backend.cc
+++ b/fastdeploy/runtime/backends/paddle/paddle_backend.cc
@@ -48,9 +48,14 @@ void PaddleBackend::BuildOption(const PaddleBackendOption& option) {
       FDINFO << "Will Enable ir_debug for Paddle Backend." << std::endl;
       config_.SwitchIrDebug();
     }
-    if (option_.enable_inference_cutlass){
+    if (option_.enable_inference_cutlass) {
+#ifdef PADDLEINFERENCE_API_COMPAT_2_4_x
+      FDWARNING << "Your are using Paddle infernence 2.4.x, cutlass is not supported!" 
+                << std::endl;
+#else    
       FDINFO << "Will enable_inference_cutlass" << std::endl;
       config_.Exp_EnableUseCutlass();
+#endif
     }
     if (option_.external_stream_) {
       FDINFO << "Will use external stream for Paddle Backend." << std::endl;
@@ -291,9 +296,37 @@ bool PaddleBackend::InitFromPaddle(const std::string& model,
   auto input_names = predictor_->GetInputNames();
   auto output_names = predictor_->GetOutputNames();
   auto input_dtypes = predictor_->GetInputTypes();
-  auto output_dtypes = predictor_->GetOutputTypes();
+
+#ifdef PADDLEINFERENCE_API_COMPAT_2_4_x
+  // Note: GetInputTensorShape, GetOutputTensorShape and GetOutputTypes
+  // are not supported when Paddle Inference API version is 2.4.x.
+  std::map<std::string, std::vector<int64_t>> input_shapes;
+  std::map<std::string, std::vector<int64_t>> output_shapes;
+  std::map<std::string, paddle_infer::DataType> output_dtypes;
+  // Get the all the input shape info.
+  for (size_t i = 0; i < input_names.size(); ++i) {
+    std::vector<int64_t> shape;
+    auto handle = predictor_->GetInputHandle(input_names[i]);
+    for (int j = 0; j < handle->shape().size(); ++j) {
+      shape.push_back(static_cast<int64_t>(handle->shape()[j])); // int32 -> int64
+    }
+    input_shapes[input_names[i]] = shape;
+  }
+  // Get the all the output shape and dtype info.
+  for (size_t i = 0; i < output_names.size(); ++i) {
+    std::vector<int64_t> shape;
+    auto handle = predictor_->GetOutputHandle(output_names[i]);
+    for (int j = 0; j < handle->shape().size(); ++j) {
+      shape.push_back(static_cast<int64_t>(handle->shape()[j])); // int32 -> int64
+    }
+    output_shapes[output_names[i]] = shape;
+    output_dtypes[output_names[i]] = handle->type();
+  }
+#else
   auto input_shapes = predictor_->GetInputTensorShape();
   auto output_shapes = predictor_->GetOutputTensorShape();
+  auto output_dtypes = predictor_->GetOutputTypes();
+#endif
 
   inputs_desc_.resize(input_names.size());
   for (int i = 0; i < input_names.size(); ++i) {