Fix instant variable oom in paddle2cinn (#56662)

FisherWY · web-flow · commit df9d9c5994d1 · 2023-08-29T19:52:46.000+08:00
When using paddle2cinn, CompilationContext.with_instantiate_variables should be set to false, otherwise CINN will instant and manage variables memory, this leads to double the memory usage, which eventually leads to out of memory error.
This PR will set CompilationContext.with_instantiate_variables to false before context pass to constructing the graph compiler.
diff --git a/paddle/fluid/framework/paddle2cinn/cinn_compiler.cc b/paddle/fluid/framework/paddle2cinn/cinn_compiler.cc
@@ -322,11 +322,11 @@ std::unique_ptr<CinnCompiledObject> CinnCompiler::CompileGraph(
 
   auto scope = BuildScope(target, cinn_graph);
   CompilationContext context(cinn_graph, scope, target);
-  auto graph_compiler = std::make_unique<GraphCompiler>(context);
   context.with_instantiate_variables = false;
   if (!FLAGS_enable_pe_launch_cinn) {
     context.with_buffer_handle_instruction_inserted = true;
   }
+  auto graph_compiler = std::make_unique<GraphCompiler>(context);
   std::unique_ptr<AutoTuner> auto_tuner;
   if (FLAGS_enable_cinn_auto_tune) {
     VLOG(4) << "Compile with auto-tune";
@@ -339,11 +339,11 @@ std::unique_ptr<CinnCompiledObject> CinnCompiler::CompileGraph(
   }
   context.fetch_var_ids = std::move(fetch_ids);
   context.stream = stream;
-  auto compiled_res = graph_compiler->Build();
+  auto compiled_res = graph_compiler->Build(&context);
   auto compiled_obj = std::make_unique<CinnCompiledObject>();
   *compiled_obj = {std::move(graph_compiler),
                    std::move(auto_tuner),
-                   std::move(compiled_res),
+                   std::move(compiled_res.runtime_program),
                    scope,
                    symbol.var_model_to_program_map()};
   compiled_obj->cached_index = compiled_num;
