Merge branch 'bert' of https://github.com/MRXLT/Serving into bert

demo-serving/conf/cube.conf

 [{
-    "dict_name": "dict",
+    "dict_name": "test_dict",
     "shard": 2,
     "dup": 1,
     "timeout": 200,

demo-serving/conf/model_toolkit.prototxt

+engines {
+  name: "image_classification_resnet"
+  type: "FLUID_CPU_ANALYSIS_DIR"
+  reloadable_meta: "./data/model/paddle/fluid_time_file"
+  reloadable_type: "timestamp_ne"
+  model_data_path: "./data/model/paddle/fluid/SE_ResNeXt50_32x4d"
+  runtime_thread_num: 0
+  batch_infer_size: 0
+  enable_batch_align: 0  
+  enable_memory_optimization: true
+  static_optimization: false
+  force_update_static_cache: false
+}
+
+engines {
+  name: "text_classification_bow"
+  type: "FLUID_CPU_ANALYSIS_DIR"
+  reloadable_meta: "./data/model/paddle/fluid_time_file"
+  reloadable_type: "timestamp_ne"
+  model_data_path: "./data/model/paddle/fluid/text_classification_lstm"
+  runtime_thread_num: 0
+  batch_infer_size: 0
+  enable_batch_align: 0
+}
+
+engines {
+  name: "ctr_prediction"
+  type: "FLUID_CPU_ANALYSIS_DIR"
+  reloadable_meta: "./data/model/paddle/fluid_time_file"
+  reloadable_type: "timestamp_ne"
+  model_data_path: "./data/model/paddle/fluid/ctr_prediction"
+  runtime_thread_num: 0
+  batch_infer_size: 0
+  enable_batch_align: 0
+  sparse_param_service_type: REMOTE
+  sparse_param_service_table_name: "test_dict"
+}
+
 engines {
   name: "bert"
-  type: "FLUID_GPU_ANALYSIS_DIR"
+  type: "FLUID_CPU_ANALYSIS_DIR"
   reloadable_meta: "./data/model/paddle/fluid_time_file"
   reloadable_type: "timestamp_ne"
   model_data_path: "./data/model/paddle/fluid/bert"

demo-serving/op/bert_service_op.cpp

@@ -31,7 +31,7 @@ using baidu::paddle_serving::predictor::bert_service::Embedding_values;
 extern int64_t MAX_SEQ_LEN = 128;
 const bool POOLING = true;
 const int LAYER_NUM = 12;
-const int EMB_SIZE = 768;
+extern int EMB_SIZE = 768;
 
 int BertServiceOp::inference() {
   timeval op_start;
@@ -49,6 +49,7 @@ int BertServiceOp::inference() {
   }
 
   MAX_SEQ_LEN = req->instances(0).max_seq_len();
+  EMB_SIZE = req->instances(0).emb_size();
 
   paddle::PaddleTensor src_ids;
   paddle::PaddleTensor pos_ids;

demo-serving/op/bert_service_op.h

@@ -25,6 +25,8 @@
 #endif
 #include "demo-serving/bert_service.pb.h"
 
+#include <sys/time.h>
+
 namespace baidu {
 namespace paddle_serving {
 namespace serving {

demo-serving/proto/bert_service.proto

@@ -25,7 +25,8 @@ message BertReqInstance {
   repeated int64 sentence_type_ids = 2;
   repeated int64 position_ids = 3;
   repeated float input_masks = 4;
-  required int64 max_seq_len = 5;
+  optional int64 max_seq_len = 5;
+  optional int64 emb_size = 6;
 };
 
 message Request { repeated BertReqInstance instances = 1; };
@@ -34,7 +35,11 @@ message Embedding_values { repeated float values = 1; };
 
 message BertResInstance { repeated Embedding_values instances = 1; };
 
-message Response { repeated BertResInstance instances = 1; };
+message Response {
+  repeated BertResInstance instances = 1;
+  optional int64 op_time = 2;
+  optional int64 infer_time = 3;
+};
 
 service BertService {
   rpc inference(Request) returns (Response);

sdk-cpp/proto/bert_service.proto

@@ -25,7 +25,8 @@ message BertReqInstance {
   repeated int64 sentence_type_ids = 2;
   repeated int64 position_ids = 3;
   repeated float input_masks = 4;
-  required int64 max_seq_len = 5;
+  optional int64 max_seq_len = 5;
+  optional int64 emb_size = 6;
 };
 
 message Request { repeated BertReqInstance instances = 1; };
@@ -34,7 +35,11 @@ message Embedding_values { repeated float values = 1; };
 
 message BertResInstance { repeated Embedding_values instances = 1; };
 
-message Response { repeated BertResInstance instances = 1; };
+message Response {
+  repeated BertResInstance instances = 1;
+  optional int64 op_time = 2;
+  optional int64 infert_time = 3;
+};
 
 service BertService {
   rpc inference(Request) returns (Response);