update, test=develop

paddle/fluid/framework/distributed_strategy.proto

@@ -33,7 +33,7 @@ message ShardingConfig {
   optional bool hybrid_dp = 2 [ default = false ];
   optional int32 sharding_group_size = 3 [ default = 8 ];
   optional bool as_outer_parallelism = 4 [ default = false ];
-  optional int32 inner_parallelism_size = 5 [ default = 8 ];
+  optional int32 parallelism = 5 [ default = 1 ];
   optional bool use_pipeline = 6 [ default = false ];
 }
 
@@ -47,6 +47,8 @@ message AMPConfig {
   repeated string custom_white_list = 7;
   repeated string custom_black_list = 8;
   repeated string custom_black_varnames = 9;
+  optional bool use_pure_fp16 = 10 [ default = false ];
+  optional bool use_fp16_guard = 11 [ default = true ];
 }
 
 message LocalSGDConfig {
@@ -145,7 +147,7 @@ message DistributedStrategy {
   optional int32 fuse_grad_size_in_MB = 19 [ default = 32 ];
   optional float fuse_grad_size_in_TFLOPS = 20 [ default = 50 ];
   optional bool cudnn_exhaustive_search = 21 [ default = true ];
-  optional int32 conv_workspace_size_limit = 22 [ default = 4000 ];
+  optional int32 conv_workspace_size_limit = 22 [ default = 512 ];
   optional bool cudnn_batchnorm_spatial_persistent = 23 [ default = true ];
   optional bool adaptive_localsgd = 24 [ default = false ];
   optional bool fp16_allreduce = 25 [ default = false ];

paddle/fluid/framework/section_worker.cc

@@ -11,25 +11,16 @@ limitations under the License. */
 
 #if defined(PADDLE_WITH_NCCL)
 #include <float.h>
+#include "paddle/fluid/framework/device_worker.h"
 #include "paddle/fluid/framework/executor_gc_helper.h"
-#include "paddle/fluid/framework/garbage_collector.h"
-#include "paddle/fluid/framework/program_desc.h"
-
-#include "google/protobuf/io/zero_copy_stream_impl.h"
-#include "google/protobuf/message.h"
-#include "google/protobuf/text_format.h"
 
-#include "paddle/fluid/framework/device_worker.h"
-#include "paddle/fluid/framework/fleet/box_wrapper.h"
-#include "paddle/fluid/framework/tensor_util.h"
-#include "paddle/fluid/framework/trainer_desc.pb.h"
-#include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/device_context.h"
-#include "paddle/fluid/platform/lodtensor_printer.h"
 
 namespace paddle {
 namespace framework {
 
+class TrainerDesc;
+
 uint64_t SectionWorker::batch_id_(0);
 
 void SectionWorker::Initialize(const TrainerDesc& desc) {