Convert to mixed precision support serialize params if a origin model doesn't have params. (#52994)

paddle/fluid/inference/analysis/passes/CMakeLists.txt

@@ -13,7 +13,8 @@ cc_library(
 cc_library(
   convert_to_mixed_precision
   SRCS convert_to_mixed_precision.cc
-  DEPS analysis_pass ir_graph_build_pass auto_mixed_precision_pass)
+  DEPS analysis_pass ir_graph_build_pass auto_mixed_precision_pass
+       constant_folding_pass)
 cc_library(
   ir_params_sync_among_devices_pass
   SRCS ir_params_sync_among_devices_pass.cc

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

@@ -16,6 +16,7 @@
 
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/ir/auto_mixed_precision_pass.h"
+#include "paddle/fluid/framework/ir/constant_folding_pass.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 #include "paddle/fluid/inference/io.h"
 #include "paddle/phi/common/backend.h"
@@ -71,8 +72,13 @@ ConvertToMixedPrecisionPass::ConvertToMixedPrecisionPass(
 
 void ConvertToMixedPrecisionPass::LoadModel() {
   framework::Executor exe{platform::CPUPlace{}};
-
-  auto program_desc = inference::Load(&exe, &scope_, model_file_, params_file_);
+  // If we did not find the provided weight path,
+  // we assume that the model to be converted only has a model file and no
+  // params file, we believe this situation is reasonable. In this case, weight
+  // data may not be loaded.
+  bool load_params = !params_file_.empty();
+  auto program_desc =
+      inference::Load(&exe, &scope_, model_file_, params_file_, load_params);
   main_graph_ = std::unique_ptr<framework::ir::Graph>(
       new framework::ir::Graph(*program_desc));
   main_graph_->SetNotOwned(framework::ir::kParamScopeAttr, &scope_);
@@ -81,6 +87,8 @@ void ConvertToMixedPrecisionPass::LoadModel() {
 void ConvertToMixedPrecisionPass::Run() {
   LoadModel();
 
+  framework::ir::ConstantFoldingPass constant_folding_pass;
+  constant_folding_pass.Apply(main_graph_.get());
   framework::ir::AutoMixedPrecisionPass pass;
   pass.Set("mixed_precision_mode", new int{static_cast<int>(mixed_precision_)});
   if (backend_ == phi::Backend::GPU) {
@@ -117,6 +125,7 @@ void ConvertToMixedPrecisionPass::SaveMixedModel() {
 
     const auto& global_block = mixed_program_desc.Block(0);
     std::vector<std::string> save_var_list;
+    bool has_persistable_var = false;
     for (framework::VarDesc* var : global_block.AllVars()) {
       if (IsPersistable(var)) {
         framework::VarDesc* new_var = save_block->Var(var->Name());
@@ -127,13 +136,35 @@ void ConvertToMixedPrecisionPass::SaveMixedModel() {
         new_var->SetPersistable(true);
 
         save_var_list.push_back(new_var->Name());
+        has_persistable_var = true;
+      }
+    }
+    std::string save_params_path = path;
+    if (save_params_path.empty() && has_persistable_var) {
+      LOG(WARNING)
+          << "The [SerializeParams] function did not find the provided weight "
+             "path, "
+             "so we assume that the model to be converted only has a model "
+             "file and no params file, "
+             "we believe this situation is reasonable. After constant folding, "
+             "a weight file will be generated, which is saved in the same "
+             "level file directory "
+             "as the model file by default and ends in pdiparams.";
+      save_params_path = mixed_model_file_;
+      std::string::size_type pos = save_params_path.rfind(".pdmodel");
+      if (pos != std::string::npos) {
+        save_params_path.replace(pos, 8, ".pdiparams");
+        LOG(WARNING) << " The storage path of the converted mixed-precision "
+                        "params has been created: ["
+                     << save_params_path << "]";
       }
     }
+
     std::sort(save_var_list.begin(), save_var_list.end());
     auto* op = save_block->AppendOp();
     op->SetType("save_combine");
     op->SetInput("X", save_var_list);
-    op->SetAttr("file_path", path);
+    op->SetAttr("file_path", save_params_path);
     op->CheckAttrs();
 
     framework::Executor exe(platform::CPUPlace{});

python/paddle/inference/wrapper.py

@@ -94,9 +94,14 @@ def convert_to_mixed_precision(
         black_list: Operators that do not convert precision.
     '''
     mixed_model_dirname = os.path.dirname(mixed_model_file)
-    mixed_params_dirname = os.path.dirname(mixed_params_file)
-    if not os.path.exists(mixed_model_dirname):
-        os.makedirs(mixed_model_dirname)
+    # Support mixed_params_file is empty, because some models don't have params, but convert_to_mixed_precision will call
+    # constant_folding_pass, it will generate a new params file to save persistable vars, which is saved in the same
+    # level file directory as the model file by default and ends in pdiparams.
+    mixed_params_dirname = (
+        os.path.dirname(mixed_params_file)
+        if len(mixed_params_file) != 0
+        else mixed_model_dirname
+    )
     if not os.path.exists(mixed_params_dirname):
         os.makedirs(mixed_params_dirname)
     convert_to_mixed_precision_bind(