Support inference compilation in training package (#46008)

* merge python lib
* Update third_party.cmake
* Update CMakeLists.txt

CMakeLists.txt

@@ -249,7 +249,7 @@ option(WITH_SYSTEM_BLAS "Use system blas library" OFF)
 option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
 option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
 option(ON_INFER "Turn on inference optimization and inference-lib generation"
-       OFF)
+       ON)
 ################################ Internal Configurations #######################################
 option(WITH_NV_JETSON "Compile PaddlePaddle with NV JETSON" OFF)
 option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler and gperftools"

cmake/third_party.cmake

@@ -236,7 +236,7 @@ endif()
 if(WIN32
    OR APPLE
    OR NOT WITH_GPU
-   OR ON_INFER)
+   OR (ON_INFER AND NOT WITH_PYTHON))
   set(WITH_DGC OFF)
 endif()
 

paddle/fluid/framework/new_executor/new_executor_defs.cc

@@ -21,17 +21,8 @@
 
 #include "paddle/phi/core/utils/rw_lock.h"
 
-// When in inference scenario, the scopes will not be written by two threads in
-// a mean time, but a scope may be read by multiple threads concurrently, and
-// the mutex will cause serious performance issue.
-// So the mutex is disabled when `ON_INFER`.
-#ifdef PADDLE_ON_INFERENCE
-#define SCOPE_VARS_READER_LOCK
-#define SCOPE_VARS_WRITER_LOCK
-#else
 #define SCOPE_VARS_READER_LOCK AutoRDLock auto_lock(&vars_lock_);
 #define SCOPE_VARS_WRITER_LOCK AutoWRLock auto_lock(&vars_lock_);
-#endif
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/new_executor/new_executor_defs.h

@@ -24,17 +24,8 @@
 #include "paddle/fluid/platform/event.h"
 #include "paddle/phi/core/utils/rw_lock.h"
 
-// When in inference scenario, the scopes will not be written by two threads in
-// a mean time, but a scope may be read by multiple threads concurrently, and
-// the mutex will cause serious performance issue.
-// So the mutex is disabled when `ON_INFER`.
-#ifdef PADDLE_ON_INFERENCE
-#define SCOPE_VARS_READER_LOCK
-#define SCOPE_VARS_WRITER_LOCK
-#else
 #define SCOPE_VARS_READER_LOCK AutoRDLock auto_lock(&vars_lock_);
 #define SCOPE_VARS_WRITER_LOCK AutoWRLock auto_lock(&vars_lock_);
-#endif
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/scope.cc

@@ -25,21 +25,10 @@ PADDLE_DEFINE_EXPORTED_bool(
     "Delete local scope eagerly. It will reduce GPU memory usage but "
     "slow down the destruction of variables.(around 1% performance harm)");
 
-// When in inference scenario, the scopes will not be written by two threads in
-// a mean time, but a scope may be read by multiple threads concurrently, and
-// the mutex will cause serious performance issue.
-// So the mutex is disabled when `ON_INFER`.
-#ifdef PADDLE_ON_INFERENCE
-#define SCOPE_KIDS_READER_LOCK
-#define SCOPE_KIDS_WRITER_LOCK
-#define SCOPE_VARS_READER_LOCK
-#define SCOPE_VARS_WRITER_LOCK
-#else
 #define SCOPE_KIDS_READER_LOCK phi::AutoRDLock auto_lock(&kids_lock_);
 #define SCOPE_KIDS_WRITER_LOCK phi::AutoWRLock auto_lock(&kids_lock_);
 #define SCOPE_VARS_READER_LOCK phi::AutoRDLock auto_lock(&vars_lock_);
 #define SCOPE_VARS_WRITER_LOCK phi::AutoWRLock auto_lock(&vars_lock_);
-#endif
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/scope.h

@@ -179,12 +179,9 @@ class Scope : public ScopeBase {
 
   DISABLE_COPY_AND_ASSIGN(Scope);
 
-#ifndef PADDLE_ON_INFERENCE
-
  private:
   mutable phi::RWLock kids_lock_;
   mutable phi::RWLock vars_lock_;
-#endif
 };
 
 // Generate some debug string about the inherience structure of scope, quite

paddle/fluid/platform/flags.cc

@@ -388,11 +388,7 @@ PADDLE_DEFINE_EXPORTED_int32(
  *       enable garbage collection strategy when training large networks.
  */
 // Disable gc by default when inference library is built
-#ifdef PADDLE_ON_INFERENCE
-static const double kDefaultEagerDeleteTensorGB = -1;
-#else
 static const double kDefaultEagerDeleteTensorGB = 0;
-#endif
 
 PADDLE_DEFINE_EXPORTED_double(
     eager_delete_tensor_gb,
@@ -663,11 +659,7 @@ PADDLE_DEFINE_EXPORTED_bool(use_mkldnn, false, "Use MKLDNN to run");
  * If FLAGS_call_stack_level == 2, the python stack, c++ stack, and error
  * message summary will be shown.
  */
-#ifdef PADDLE_ON_INFERENCE
-static const int32_t kDefaultCallStackLevel = 2;
-#else
 static const int32_t kDefaultCallStackLevel = 1;
-#endif
 
 PADDLE_DEFINE_EXPORTED_int32(
     call_stack_level,

paddle/fluid/pybind/CMakeLists.txt

@@ -140,7 +140,7 @@ if(WITH_CUSTOM_DEVICE)
   set(PYBIND_DEPS ${PYBIND_DEPS} custom_device_common_op_registry)
 endif()
 
-if(NOT ON_INFER)
+if(WITH_PYTHON)
   set(PYBIND_DEPS ${PYBIND_DEPS} processgroup eager_reducer)
   if(WITH_NCCL OR WITH_RCCL)
     set(PYBIND_DEPS ${PYBIND_DEPS} processgroup_nccl)

paddle/fluid/pybind/pybind.cc

@@ -2547,7 +2547,7 @@ All parameter, weight, gradient are variables in Paddle.
   BindCompatible(&m);
   BindDataset(&m);
   BindGenerator(&m);
-#ifndef PADDLE_ON_INFERENCE
+#ifndef PADDLE_NO_PYTHON
   BindDistributed(&m);
 #endif
 #ifdef PADDLE_WITH_ASCEND

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -799,7 +799,7 @@ py_test_modules(
 # it is found that windows CI will run all the training unittests with the ON_INFER option turned on,
 # which will not appear in other CIs. The calculation behavior of some ops in inference mode is
 # inconsistent with that in non-inference mode.
-if(NOT ON_INFER)
+if(WITH_PYTHON)
   py_test_modules(test_parallel_executor_seresnext_base_cpu MODULES
                   test_parallel_executor_seresnext_base_cpu)
   py_test_modules(test_parallel_executor_seresnext_with_reduce_cpu MODULES

python/paddle/fluid/tests/unittests/dygraph_to_static/CMakeLists.txt

@@ -34,7 +34,7 @@ set(TEST_EAGER_OPS
 list(REMOVE_ITEM TEST_OPS test_lac)
 # NOTE(Aurelius84): In case of Windows CI, if open ON_INFER, RWLOCK of Scope
 # will be removed and will cause some random failed in multi-thread.
-if(NOT ON_INFER)
+if(WITH_PYTHON)
   py_test_modules(test_lac MODULES test_lac ENVS FLAGS_enable_eager_mode=1)
   set_tests_properties(test_lac PROPERTIES TIMEOUT 120)
 endif()