merge from origin

core/predictor/common/utils.h

@@ -13,8 +13,10 @@
 // limitations under the License.
 
 #pragma once
-#include <string>
+#include <algorithm>
+#include <cctype>
 #include <fstream>
+#include <string>
 #include "core/predictor/common/inner_common.h"
 #include "core/predictor/common/macros.h"
 
@@ -26,6 +28,38 @@ namespace predictor {
 namespace butil = base;
 #endif
 
+enum class Precision {
+  kUnk = -1,     // unknown type
+  kFloat32 = 0,  // fp32
+  kInt8,         // int8
+  kHalf,         // fp16
+  kBfloat16,     // bf16
+};
+
+static std::string PrecisionTypeString(const Precision data_type) {
+  switch (data_type) {
+    case Precision::kFloat32:
+      return "kFloat32";
+    case Precision::kInt8:
+      return "kInt8";
+    case Precision::kHalf:
+      return "kHalf";
+    case Precision::kBfloat16:
+      return "kBloat16";
+    default:
+      return "unUnk";
+  }
+}
+
+static std::string ToLower(const std::string& data) {
+  std::string result = data;
+  std::transform(
+      result.begin(), result.end(), result.begin(), [](unsigned char c) {
+        return tolower(c);
+      });
+  return result;
+}
+
 class TimerFlow {
  public:
   static const int MAX_SIZE = 1024;

python/paddle_serving_app/local_predict.py

@@ -27,6 +27,12 @@ logging.basicConfig(format="%(asctime)s - %(levelname)s - %(message)s")
 logger = logging.getLogger("LocalPredictor")
 logger.setLevel(logging.INFO)
 
+precision_map = {
+    'int8': paddle_infer.PrecisionType.Int8,
+    'fp32': paddle_infer.PrecisionType.Float32,
+    'fp16': paddle_infer.PrecisionType.Half,
+}
+
 
 class LocalPredictor(object):
     """
@@ -56,6 +62,8 @@ class LocalPredictor(object):
                           use_trt=False,
                           use_lite=False,
                           use_xpu=False,
+                          precision="fp32",
+                          use_calib=False,
                           use_feed_fetch_ops=False):
         """
         Load model configs and create the paddle predictor by Paddle Inference API.
@@ -71,6 +79,8 @@ class LocalPredictor(object):
             use_trt: use nvidia TensorRT optimization, False default
             use_lite: use Paddle-Lite Engint, False default
             use_xpu: run predict on Baidu Kunlun, False default
+            precision: precision mode, "fp32" default
+            use_calib: use TensorRT calibration, False default
             use_feed_fetch_ops: use feed/fetch ops, False default.
         """
         client_config = "{}/serving_server_conf.prototxt".format(model_path)
@@ -88,9 +98,11 @@ class LocalPredictor(object):
         logger.info(
             "LocalPredictor load_model_config params: model_path:{}, use_gpu:{},\
             gpu_id:{}, use_profile:{}, thread_num:{}, mem_optim:{}, ir_optim:{},\
-            use_trt:{}, use_lite:{}, use_xpu: {}, use_feed_fetch_ops:{}".format(
-                model_path, use_gpu, gpu_id, use_profile, thread_num, mem_optim,
-                ir_optim, use_trt, use_lite, use_xpu, use_feed_fetch_ops))
+            use_trt:{}, use_lite:{}, use_xpu: {}, precision: {}, use_calib: {},\
+            use_feed_fetch_ops:{}"
+            .format(model_path, use_gpu, gpu_id, use_profile, thread_num,
+                    mem_optim, ir_optim, use_trt, use_lite, use_xpu, precision,
+                    use_calib, use_feed_fetch_ops))
 
         self.feed_names_ = [var.alias_name for var in model_conf.feed_var]
         self.fetch_names_ = [var.alias_name for var in model_conf.fetch_var]
@@ -106,6 +118,9 @@ class LocalPredictor(object):
             self.fetch_names_to_idx_[var.alias_name] = i
             self.fetch_names_to_type_[var.alias_name] = var.fetch_type
 
+        precision_type = paddle_infer.PrecisionType.Float32
+        if precision.lower() in precision_map:
+            precision_type = precision_map[precision.lower()]
         if use_profile:
             config.enable_profile()
         if mem_optim:
@@ -121,6 +136,7 @@ class LocalPredictor(object):
             config.enable_use_gpu(100, gpu_id)
             if use_trt:
                 config.enable_tensorrt_engine(
+                    precision_mode=precision_type,
                     workspace_size=1 << 20,
                     max_batch_size=32,
                     min_subgraph_size=3,
@@ -129,7 +145,7 @@ class LocalPredictor(object):
 
         if use_lite:
             config.enable_lite_engine(
-                precision_mode=paddle_infer.PrecisionType.Float32,
+                precision_mode=precision_type,
                 zero_copy=True,
                 passes_filter=[],
                 ops_filter=[])
@@ -138,6 +154,11 @@ class LocalPredictor(object):
             # 2MB l3 cache
             config.enable_xpu(8 * 1024 * 1024)
 
+        if not use_gpu and not use_lite:
+            if precision_type == paddle_infer.PrecisionType.Int8:
+                config.enable_quantizer()
+            if precision.lower() == "bf16":
+                config.enable_mkldnn_bfloat16()
         self.predictor = paddle_infer.create_predictor(config)
 
     def predict(self, feed=None, fetch=None, batch=False, log_id=0):

python/paddle_serving_server/server.py

@@ -71,6 +71,8 @@ class Server(object):
         self.max_concurrency = 0
         self.num_threads = 2
         self.port = 8080
+        self.precision = "fp32"
+        self.use_calib = False
         self.reload_interval_s = 10
         self.max_body_size = 64 * 1024 * 1024
         self.module_path = os.path.dirname(paddle_serving_server.__file__)
@@ -113,6 +115,12 @@ class Server(object):
     def set_port(self, port):
         self.port = port
 
+    def set_precision(self, precision="fp32"):
+        self.precision = precision
+
+    def set_use_calib(self, use_calib=False):
+        self.use_calib = use_calib
+
     def set_reload_interval(self, interval):
         self.reload_interval_s = interval
 
@@ -186,6 +194,10 @@ class Server(object):
             engine.use_trt = self.use_trt
             engine.use_lite = self.use_lite
             engine.use_xpu = self.use_xpu
+            engine.use_gpu = False
+            if self.device == "gpu":
+                engine.use_gpu = True
+
             if os.path.exists('{}/__params__'.format(model_config_path)):
                 engine.combined_model = True
             else:
@@ -472,6 +484,8 @@ class Server(object):
                       "-max_concurrency {} " \
                       "-num_threads {} " \
                       "-port {} " \
+                      "-precision {} " \
+                      "-use_calib {} " \
                       "-reload_interval_s {} " \
                       "-resource_path {} " \
                       "-resource_file {} " \
@@ -485,6 +499,8 @@ class Server(object):
                           self.max_concurrency,
                           self.num_threads,
                           self.port,
+                          self.precision,
+                          self.use_calib,
                           self.reload_interval_s,
                           self.workdir,
                           self.resource_fn,
@@ -500,6 +516,8 @@ class Server(object):
                       "-max_concurrency {} " \
                       "-num_threads {} " \
                       "-port {} " \
+                      "-precision {} " \
+                      "-use_calib {} " \
                       "-reload_interval_s {} " \
                       "-resource_path {} " \
                       "-resource_file {} " \
@@ -514,6 +532,8 @@ class Server(object):
                           self.max_concurrency,
                           self.num_threads,
                           self.port,
+                          self.precision,
+                          self.use_calib,
                           self.reload_interval_s,
                           self.workdir,
                           self.resource_fn,
@@ -562,6 +582,12 @@ class MultiLangServer(object):
     def set_port(self, port):
         self.gport_ = port
 
+    def set_precision(self, precision="fp32"):
+        self.precision = precision
+
+    def set_use_calib(self, use_calib=False):
+        self.use_calib = use_calib
+
     def set_reload_interval(self, interval):
         self.bserver_.set_reload_interval(interval)
 

python/pipeline/pipeline_server.py

@@ -238,6 +238,8 @@ class PipelineServer(object):
                 "devices": "",
                 "mem_optim": True,
                 "ir_optim": False,
+                "precision": "fp32",
+                "use_calib": False,
             },
         }
         for op in self._used_op:
@@ -394,6 +396,8 @@ class ServerYamlConfChecker(object):
             "devices": "",
             "mem_optim": True,
             "ir_optim": False,
+            "precision": "fp32",
+            "use_calib": False,
         }
         conf_type = {
             "model_config": str,
@@ -403,6 +407,8 @@ class ServerYamlConfChecker(object):
             "devices": str,
             "mem_optim": bool,
             "ir_optim": bool,
+            "precision": str,
+            "use_calib": bool,
         }
         conf_qualification = {"thread_num": (">=", 1), }
         ServerYamlConfChecker.check_conf(conf, default_conf, conf_type,