[Inference] fix Predictor support pir model save(.json) (#65595)

lizexu123 · yuanlehome · web-flow · commit be4adff1dd59 · 2024-07-02T02:49:18.000-07:00
* kk

* 解决json和pdmodel保存路径不一致,pdmodel没保存下来优化后的模型

* fix

* 删去了pdmodel跑pir的时候,跑了旧ir的pass

* fix

* fix

* update

---------

Co-authored-by: yuanlehome &lt;yuanlehome@163.com&gt;
diff --git a/paddle/fluid/inference/api/analysis_config.cc b/paddle/fluid/inference/api/analysis_config.cc
@@ -1256,13 +1256,6 @@ void AnalysisConfig::DisableGlogInfo() {
   Update();
 }
 
-void AnalysisConfig::PartiallyRelease() {
-  prog_file_.clear();
-  prog_file_.shrink_to_fit();
-  params_file_.clear();
-  params_file_.shrink_to_fit();
-}
-
 void AnalysisConfig::EnableGpuMultiStream() { thread_local_stream_ = true; }
 
 std::string AnalysisConfig::Summary() {
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -441,11 +441,12 @@ bool AnalysisPredictor::Init(
   std::string model_path = config_.prog_file();
   load_pir_model_ =
       model_path.substr(model_path.find_last_of(".") + 1) == "json";
+
   // Use Optimized model to inference
   if (config_.use_optimized_model_) {
     std::string optimized_model_path = GetOptimizedModelPath();
     std::string optimized_model;
-    if (load_pir_model_) {
+    if (config_.new_ir_enabled()) {
       optimized_model = optimized_model_path + "/" + "_optimized.json";
     } else {
       optimized_model = optimized_model_path + "/" + "_optimized.pdmodel";
@@ -454,6 +455,9 @@ bool AnalysisPredictor::Init(
         optimized_model_path + "/" + "_optimized.pdiparams";
     if (FileExists(optimized_model) && FileExists(optimized_params)) {
       config_.SetModel(optimized_model, optimized_params);
+      if (config_.new_ir_enabled()) {
+        load_pir_model_ = true;
+      }
       LOG(INFO) << "Load Optimized model from " << optimized_model
                 << " and Load Optimized optimized_params from "
                 << optimized_params;
@@ -470,7 +474,6 @@ bool AnalysisPredictor::Init(
   if (!PrepareScope(parent_scope)) {
     return false;
   }
-
   InitPlace();
 
   if (!CreateExecutor()) {
@@ -652,6 +655,7 @@ void AnalysisPredictor::ClearExtraParams() {
       }
     }
   }
+
   scope_->EraseVars(extra_params);
   VLOG(1) << "Clear " << extra_params.size() << " extra params.";
 }
@@ -1063,23 +1067,23 @@ bool AnalysisPredictor::SaveOrLoadPirParameters(bool for_save) {
   for (size_t i = 0; i < len; ++i) {
     auto *var = sub_scope_->FindVar(param_names[i]);
     pir::Value value = vars[i];
+
     if (var == nullptr) {
       if (value && value.type().isa<pir::DenseTensorType>()) {
         var = sub_scope_->Var(param_names[i]);
-        auto *tensor_temp = var->GetMutable<phi::DenseTensor>();
-        tensor_temp->Resize(common::make_ddim(pir::GetShapeFromValue(value)));
-        phi::DeviceContextPool &pool = phi::DeviceContextPool::Instance();
-        const phi::DeviceContext *dev_ctx = nullptr;
-        dev_ctx = pool.Get(place_);
-        pir::Type type_ = pir::GetDataTypeFromValue(value);
-        phi::DataType type_data = paddle::dialect::TransToPhiDataType(type_);
-        dev_ctx->Alloc(tensor_temp, type_data);
       } else {
         PADDLE_THROW(platform::errors::Unavailable(
             "Only support parameter data of type DenseTensor."));
       }
     }
     auto *tensor_temp = var->GetMutable<phi::DenseTensor>();
+    tensor_temp->Resize(common::make_ddim(pir::GetShapeFromValue(value)));
+
+    phi::DeviceContextPool &pool = phi::DeviceContextPool::Instance();
+    const phi::DeviceContext *dev_ctx = pool.Get(place_);
+    pir::Type type_ = pir::GetDataTypeFromValue(value);
+    phi::DataType type_data = paddle::dialect::TransToPhiDataType(type_);
+    dev_ctx->Alloc(tensor_temp, type_data);
     tensor_out.push_back(tensor_temp);
   }
 
@@ -1108,7 +1112,6 @@ bool AnalysisPredictor::PreparePirProgram() {
       platform::errors::Fatal("Here, pir_program must be a nullptr!"));
 
   pir_program_ = std::make_shared<pir::Program>(pir::IrContext::Instance());
-
   pir::ReadModule(config_.prog_file(), pir_program_.get(), 1 /*pir_version*/);
   if (!SaveOrLoadPirParameters(false)) {
     return false;
@@ -1238,7 +1241,7 @@ bool AnalysisPredictor::PrepareExecutor() {
 #endif
   }
 
-  if (load_pir_model_) {
+  if (config_.new_ir_enabled()) {
     executor_->Prepare(sub_scope_);
   } else {
     DisablePrepareDataOpt(inference_program_, 0, false);
@@ -1259,7 +1262,6 @@ bool AnalysisPredictor::PrepareExecutor() {
     execution_config.skip_gc_vars.insert(output_names.begin(),
                                          output_names.end());
 
-    VLOG(1) << "predictor program " << *pir_program_;
     if (config_.new_ir_enabled()) {
       executor_->PrepareInterpreterCore(
           sub_scope_, *pir_program_, execution_config);
@@ -2276,13 +2278,12 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
 #endif
         delete prog;
       });
-  // The config and argument take a lot of storage,
-  // when the predictor settings are complete, we release these stores.
-  config_.PartiallyRelease();
+
 #if defined(PADDLE_WITH_TESTING)
   fusion_statis_ = *argument_->fusion_statis_ptr();
 #endif
-
+  // The argument take a lot of storage,
+  // when the predictor settings are complete, we release these stores.
 #if defined(_WIN32)
   argument_->PartiallyRelease();
 #else
diff --git a/paddle/fluid/inference/api/paddle_analysis_config.h b/paddle/fluid/inference/api/paddle_analysis_config.h
@@ -1120,7 +1120,6 @@ struct PD_INFER_DECL AnalysisConfig {
   /// stream to the thread, and this behavior may be changed in the future.
   ///
   void EnableGpuMultiStream();
-  void PartiallyRelease();
 
   ///
   /// \brief Print the summary of config.
diff --git a/paddle/fluid/inference/capi_exp/pd_config.cc b/paddle/fluid/inference/capi_exp/pd_config.cc
@@ -487,10 +487,7 @@ void PD_ConfigSetExecStream(__pd_keep PD_Config* pd_config, void* stream) {
   CHECK_AND_CONVERT_PD_CONFIG;
   return config->SetExecStream(stream);
 }
-void PD_ConfigPartiallyRelease(__pd_keep PD_Config* pd_config) {
-  CHECK_AND_CONVERT_PD_CONFIG;
-  config->PartiallyRelease();
-}
+
 void PD_ConfigDeletePass(__pd_keep PD_Config* pd_config, const char* pass) {
   CHECK_AND_CONVERT_PD_CONFIG;
   config->pass_builder()->DeletePass(pass);
diff --git a/paddle/fluid/inference/capi_exp/pd_config.h b/paddle/fluid/inference/capi_exp/pd_config.h
@@ -721,13 +721,6 @@ PADDLE_CAPI_EXPORT extern void PD_ConfigSetInvalid(
 PADDLE_CAPI_EXPORT extern PD_Bool PD_ConfigIsValid(
     __pd_keep PD_Config* pd_config);
 ///
-/// \brief Partially release the memory
-///
-/// \param[in] pd_config config
-///
-PADDLE_CAPI_EXPORT extern void PD_ConfigPartiallyRelease(
-    __pd_keep PD_Config* pd_config);
-///
 /// \brief Delete all passes that has a certain type 'pass'.
 ///
 /// \param[in] pass the certain pass type to be deleted.
diff --git a/test/cpp/inference/api/analyzer_capi_exp_pd_config_tester.cc b/test/cpp/inference/api/analyzer_capi_exp_pd_config_tester.cc
@@ -112,7 +112,6 @@ TEST(PD_Config, interface) {
   bool is_valid = PD_ConfigIsValid(config);
   EXPECT_FALSE(is_valid);
 
-  PD_ConfigPartiallyRelease(config);
   PD_ConfigDestroy(config);
 }
 

Original file line number	Diff line number	Diff line change
`@@ -112,7 +112,6 @@ TEST(PD_Config, interface) {`
`112`	`112`	`bool is_valid = PD_ConfigIsValid(config);`
`113`	`113`	`EXPECT_FALSE(is_valid);`
`114`	`114`
`115`		`- PD_ConfigPartiallyRelease(config);`
`116`	`115`	`PD_ConfigDestroy(config);`
`117`	`116`	`}`
`118`	`117`