Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

… debug_latex_ocr_branch

cmake/external/openvino.cmake

@@ -103,7 +103,7 @@ file(TO_NATIVE_PATH ${PADDLE_SOURCE_DIR}/patches/openvino/convert.patch
      native_convert)
 
 set(OPENVINO_PATCH_COMMAND
-    git checkout -- . && git fetch --depth=1 origin <OPENVINO_COMMIT> && git
+    git checkout -- . && git fetch --depth=1 origin ${OPENVINO_COMMIT} && git
     checkout ${OPENVINO_COMMIT} && patch -Np1 -d ${SOURCE_DIR} <
     ${native_convert} || true)
 

paddle/cinn/backends/compiler.cc

@@ -430,7 +430,6 @@ void Compiler::CompileCudaModule(const Module& module,
     device_fn_name_.emplace_back(kernel_fn_name);
   }
   engine_->Link<CodeGenGpuHost>(host_module);
-
 #else
   CINN_NOT_IMPLEMENTED
 #endif

paddle/cinn/backends/llvm/execution_engine.cc

@@ -171,8 +171,10 @@ std::unique_ptr<llvm::MemoryBuffer> NaiveObjectCache::getObject(
 
 template <typename CodeGenT>
 void ExecutionEngine::Link(const ir::Module &module) {
+  if (module.functions().size() == 0) {
+    return;
+  }
   utils::RecordEvent("ExecutionEngine Link", utils::EventType::kOrdinary);
-
   auto ir_emitter = std::make_unique<CodeGenT>(m.get(), b.get());
   VLOG(3) << "ir_emitter->Compile(module) Begin";
   ir_emitter->Compile(module);
@@ -211,6 +213,16 @@ void ExecutionEngine::Link(const ir::Module &module) {
   }
 }
 
+template <>
+void ExecutionEngine::Link<CodeGenGpuHost>(const ir::Module &module) {
+  if (module.functions().size() == 0) {
+    return;
+  }
+  utils::RecordEvent("ExecutionEngine Link", utils::EventType::kOrdinary);
+  auto ir_emitter = std::make_unique<CodeGenGpuHost>(m.get(), b.get());
+  ir_emitter->Compile(module);
+}
+
 bool ExecutionEngine::AddModule(std::unique_ptr<llvm::Module> module,
                                 std::unique_ptr<llvm::LLVMContext> context) {
   utils::RecordEvent("ExecutionEngine AddModule", utils::EventType::kOrdinary);

paddle/fluid/eager/accumulation/accumulation_node.h

@@ -19,6 +19,8 @@
 #include "paddle/fluid/eager/hooks.h"
 #include "paddle/utils/test_macros.h"
 
+COMMON_DECLARE_int32(call_stack_level);
+
 namespace egr {
 
 class TEST_API GradNodeAccumulation : public GradNodeBase {
@@ -30,6 +32,10 @@ class TEST_API GradNodeAccumulation : public GradNodeBase {
       weak_grad_ = meta->WeakGrad();
     }
 
+    if (FLAGS_call_stack_level == 3) {
+      this->SetForwardTrace(egr::Controller::Instance().GetPythonStack());
+    }
+
     SetDefaultGradInOutMeta();
   }
 

paddle/fluid/eager/auto_code_generator/generator/eager_gen.py

@@ -485,7 +485,7 @@ class {} : public egr::GradNodeBase {{
     // Node Construction
 {}
     // Set for forward trace
-  if (FLAGS_check_nan_inf) {{
+  if (FLAGS_check_nan_inf || FLAGS_call_stack_level == 3) {{
     grad_node->SetForwardTrace(egr::Controller::Instance().GetPythonStack());
   }}
     // SetAttributes if needed
@@ -590,6 +590,7 @@ class {} : public egr::GradNodeBase {{
 #include "paddle/fluid/imperative/amp_utils.h"
 
 COMMON_DECLARE_bool(check_nan_inf);
+COMMON_DECLARE_int32(call_stack_level);
 COMMON_DECLARE_string(tensor_operants_mode);
 COMMON_DECLARE_bool(use_stride_kernel);
 {}