PaddlePaddle
diff --git a/‎paddle/cinn/hlir/dialect/operator/transforms/cinn_group_cluster_pass.cc
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions b/‎paddle/cinn/hlir/dialect/operator/transforms/cinn_group_cluster_pass.cc
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions
diff --git a/‎paddle/common/flags.cc
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions b/‎paddle/common/flags.cc
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions
diff --git a/‎paddle/fluid/framework/new_executor/instruction/tensorrt_engine_instruction.cc
Copy file name to clipboard
+24-2Lines changed: 24 additions & 2 deletions b/‎paddle/fluid/framework/new_executor/instruction/tensorrt_engine_instruction.cc
Copy file name to clipboard
+24-2Lines changed: 24 additions & 2 deletions
diff --git a/‎paddle/fluid/framework/new_executor/instruction/tensorrt_engine_instruction.h
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions b/‎paddle/fluid/framework/new_executor/instruction/tensorrt_engine_instruction.h
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions
diff --git a/‎paddle/fluid/pir/dialect/operator/ir/op_attribute.h
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions b/‎paddle/fluid/pir/dialect/operator/ir/op_attribute.h
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions
diff --git a/‎paddle/fluid/pir/transforms/CMakeLists.txt
Copy file name to clipboard
+8-4Lines changed: 8 additions & 4 deletions b/‎paddle/fluid/pir/transforms/CMakeLists.txt
Copy file name to clipboard
+8-4Lines changed: 8 additions & 4 deletions
diff --git a/‎paddle/fluid/pir/transforms/passes.h
Copy file name to clipboard
+5Lines changed: 5 additions & 0 deletions b/‎paddle/fluid/pir/transforms/passes.h
Copy file name to clipboard
+5Lines changed: 5 additions & 0 deletions
diff --git a/‎paddle/fluid/pir/transforms/sub_graph_detector.cc
Copy file name to clipboard
+21-2Lines changed: 21 additions & 2 deletions b/‎paddle/fluid/pir/transforms/sub_graph_detector.cc
Copy file name to clipboard
+21-2Lines changed: 21 additions & 2 deletions
diff --git a/‎paddle/fluid/pir/transforms/sub_graph_detector.h
Copy file name to clipboard
+2Lines changed: 2 additions & 0 deletions b/‎paddle/fluid/pir/transforms/sub_graph_detector.h
Copy file name to clipboard
+2Lines changed: 2 additions & 0 deletions
@@ -31,6 +31,7 @@
 #include "paddle/cinn/hlir/dialect/operator/ir/attribute_storage.h"
 #include "paddle/cinn/hlir/dialect/operator/ir/cinn_op.h"
 #include "paddle/cinn/hlir/dialect/operator/ir/manual_op.h"
+#include "paddle/cinn/hlir/dialect/operator/ir/op_dialect.h"
 #include "paddle/cinn/hlir/dialect/operator/transforms/group_merge/op_with_group_merge_util.h"
 #include "paddle/cinn/hlir/framework/pir/utils.h"
 #include "paddle/cinn/operator_fusion/cluster_interface.h"
 
@@ -1892,3 +1892,9 @@ PHI_DEFINE_EXPORTED_bool(
     pinned_memory_as_cpu_backend,
     false,
     "Whether use CPU backend, when tensor is pinned_memory.");
+
+PHI_DEFINE_EXPORTED_int32(
+    trt_min_group_size,
+    3,
+    "when the trt subgraph size is not larger than `trt_min_group_size`, the "
+    "group will fallback to original graph.");
@@ -34,7 +34,7 @@ TensorRTEngineInstruction::TensorRTEngineInstruction(
   auto op_attributes = op->attributes();
 
   VLOG(6) << "Start Build engine";
-  auto engine_serialized_data = op_attributes.at("engine_serialized_data")
+  auto engine_serialized_path = op_attributes.at("engine_serialized_data")
                                     .dyn_cast<pir::StrAttribute>()
                                     .AsString();
   workspace_size_ =
@@ -155,7 +155,8 @@ TensorRTEngineInstruction::TensorRTEngineInstruction(
              "===============";
   trt_engine_ = std::make_unique<paddle::platform::TensorRTEngine>(
       params, paddle::platform::NaiveLogger::Global());
-  trt_engine_->Deserialize(engine_serialized_data);
+  auto engine_data = ReadBinaryFileToString(engine_serialized_path);
+  trt_engine_->Deserialize(engine_data);
 
   VLOG(6) << "Finish build engine for: " << op_name_;
 
@@ -759,5 +760,26 @@ void TensorRTEngineInstruction::Run() {
   RunTrt();
 }
 
+std::string TensorRTEngineInstruction::ReadBinaryFileToString(
+    const std::string &filePath) {
+  std::ifstream inputFile(filePath, std::ios::binary);
+
+  if (!inputFile) {
+    throw std::runtime_error("Failed to open file: " + filePath);
+  }
+
+  inputFile.seekg(0, std::ios::end);
+  std::streamsize fileSize = inputFile.tellg();
+  inputFile.seekg(0, std::ios::beg);
+
+  std::string fileContent(static_cast<size_t>(fileSize), '\0');
+
+  if (!inputFile.read(&fileContent[0], fileSize)) {
+    throw std::runtime_error("Failed to read file: " + filePath);
+  }
+
+  return fileContent;
+}
+
 }  // namespace framework
 }  // namespace paddle
@@ -42,6 +42,7 @@ class TensorRTEngineInstruction : public InstructionBase {
   const std::string& Name() const override { return op_name_; }
 
  private:
+  std::string ReadBinaryFileToString(const std::string& filePath);
   void PrepareDynamicShape();
   void RunTrt();
   void BindInputTensor(const std::string& input_name,
 
@@ -25,6 +25,7 @@ namespace paddle {
 namespace dialect {
 // __force_backend__ in ["gpu","gpudnn","cpu",""]
 inline const char kForceBackendAttr[] = "__force_backend__";
+inline const char kCanRunTrtAttr[] = "__l_trt__";
 
 class IntArrayAttribute : public pir::Attribute {
  public:
 
@@ -1,9 +1,8 @@
 file(GLOB_RECURSE transforms_srcs "*.cc")
 if(NOT WITH_CINN)
-  list(
-    REMOVE_ITEM transforms_srcs ${CMAKE_CURRENT_SOURCE_DIR}/build_cinn_pass.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/sub_graph_extract_pass.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/sub_graph_detector.cc)
+  list(REMOVE_ITEM transforms_srcs
+       ${CMAKE_CURRENT_SOURCE_DIR}/build_cinn_pass.cc
+       ${CMAKE_CURRENT_SOURCE_DIR}/sub_graph_extract_pass.cc)
 endif()
 
 if(NOT WITH_ONEDNN)
@@ -16,6 +15,11 @@ if(NOT WITH_XPU)
   list(REMOVE_ITEM transforms_srcs ${xpu_srcs})
 endif()
 
+if(NOT TENSORRT_FOUND)
+  file(GLOB_RECURSE trt_srcs "tensorrt/*.cc")
+  list(REMOVE_ITEM transforms_srcs ${trt_srcs})
+endif()
+
 set(transforms_deps
     drr
     op_dialect
 
@@ -93,3 +93,8 @@ USE_PIR_PASS(fc_xpu_fuse_pass);
 #ifdef PADDLE_WITH_CINN
 USE_PIR_PASS(convert_MEA_to_FA);
 #endif
+
+#ifdef PADDLE_WITH_TENSORRT
+USE_PIR_PASS(trt_op_marker_pass);
+USE_PIR_PASS(trt_sub_graph_extract_pass);
+#endif
@@ -23,8 +23,12 @@
 #include <string>
 #include <unordered_map>
 
+#ifdef PADDLE_WITH_CINN
 #include "paddle/cinn/hlir/dialect/operator/ir/manual_op.h"
 #include "paddle/cinn/hlir/dialect/operator/ir/op_dialect.h"
+#include "paddle/cinn/hlir/framework/pir/utils.h"
+#endif
+
 #include "paddle/fluid/pir/dialect/operator/ir/pd_op.h"
 #include "paddle/fluid/pir/utils/general_functions.h"
 #include "paddle/pir/include/core/builder.h"
@@ -34,7 +38,6 @@
 #include "paddle/pir/include/pass/pass.h"
 #include "paddle/pir/include/pass/pass_registry.h"
 
-#include "paddle/cinn/hlir/framework/pir/utils.h"
 #include "paddle/common/flags.h"
 
 #ifdef PADDLE_WITH_DNNL
@@ -609,7 +612,9 @@ pir::Operation* FindInsertPoint(const GroupOpsVec& group_ops,
 void ReplaceWithGroupOp(pir::Block* block,
                         const GroupOpsVec& group_ops) {  // NOLINT
   ::pir::IrContext* ctx = ::pir::IrContext::Instance();
+#ifdef PADDLE_WITH_CINN
   ctx->GetOrRegisterDialect<cinn::dialect::OperatorDialect>();
+#endif
 #ifdef PADDLE_WITH_DNNL
   ctx->GetOrRegisterDialect<paddle::dialect::OneDNNOperatorDialect>();
 #endif
@@ -622,7 +627,9 @@ void ReplaceWithGroupOp(pir::Block* block,
   builder.set_insertion_point(insert_point);
   VLOG(6) << "Insert GroupOp after " << insert_point->name();
 
-  // step 2: Replace the old op with GroupOp.
+// step 2: Replace the old op with GroupOp.
+#ifdef PADDLE_WITH_CINN
+
   auto new_group_op = [&]() -> cinn::dialect::GroupOp {
     std::vector<pir::Type> output_types;
     for (auto& value : outputs) output_types.emplace_back(value.type());
@@ -633,6 +640,18 @@ void ReplaceWithGroupOp(pir::Block* block,
     }
     return group_op;
   }();
+#else
+  auto new_group_op = [&]() -> pir::GroupOp {
+    std::vector<pir::Type> output_types;
+    for (auto& value : outputs) output_types.emplace_back(value.type());
+
+    auto group_op = builder.Build<pir::GroupOp>(output_types);
+    for (auto op : group_ops) {
+      op->MoveTo(group_op.block(), group_op.block()->end());
+    }
+    return group_op;
+  }();
+#endif
 
   // step 3: Replace outputs of inner ops
   const std::vector<pir::Value> group_outs = new_group_op->results();
 
@@ -23,7 +23,9 @@
 #include <unordered_map>
 #include <unordered_set>
 
+#ifdef PADDLE_WITH_CINN
 #include "paddle/cinn/hlir/dialect/operator/ir/op_dialect.h"
+#endif
 #include "paddle/pir/include/core/builder.h"
 
 namespace pir {