Align paddle inference configs with those of FD (#2436)

paddlex/inference/components/paddle_predictor/predictor.py

@@ -73,10 +73,44 @@ class BasePaddlePredictor(BaseComponent):
         params_file = (self.model_dir / f"{self.model_prefix}.pdiparams").as_posix()
         config = Config(model_file, params_file)
 
+        config.enable_memory_optim()
         if self.option.device in ("gpu", "dcu"):
-            config.enable_use_gpu(200, self.option.device_id)
-            if hasattr(config, "enable_new_ir"):
-                config.enable_new_ir(self.option.enable_new_ir)
+            if self.option.device == "gpu":
+                config.exp_disable_mixed_precision_ops({"feed", "fetch"})
+            config.enable_use_gpu(100, self.option.device_id)
+            if self.option.device == "gpu":
+                # NOTE: The pptrt settings are not aligned with those of FD.
+                precision_map = {
+                    "trt_int8": Config.Precision.Int8,
+                    "trt_fp32": Config.Precision.Float32,
+                    "trt_fp16": Config.Precision.Half,
+                }
+                if self.option.run_mode in precision_map.keys():
+                    config.enable_tensorrt_engine(
+                        workspace_size=(1 << 25) * self.option.batch_size,
+                        max_batch_size=self.option.batch_size,
+                        min_subgraph_size=self.option.min_subgraph_size,
+                        precision_mode=precision_map[self.option.run_mode],
+                        use_static=self.option.trt_use_static,
+                        use_calib_mode=self.option.trt_calib_mode,
+                    )
+
+                    if self.option.shape_info_filename is not None:
+                        if not os.path.exists(self.option.shape_info_filename):
+                            config.collect_shape_range_info(
+                                self.option.shape_info_filename
+                            )
+                            logging.info(
+                                f"Dynamic shape info is collected into: {self.option.shape_info_filename}"
+                            )
+                        else:
+                            logging.info(
+                                f"A dynamic shape info file ( {self.option.shape_info_filename} ) already exists. \
+        No need to generate again."
+                            )
+                        config.enable_tuned_tensorrt_dynamic_shape(
+                            self.option.shape_info_filename, True
+                        )
         elif self.option.device == "npu":
             config.enable_custom_device("npu")
         elif self.option.device == "xpu":
@@ -86,53 +120,32 @@ class BasePaddlePredictor(BaseComponent):
         else:
             assert self.option.device == "cpu"
             config.disable_gpu()
-            if hasattr(config, "enable_new_ir"):
-                config.enable_new_ir(self.option.enable_new_ir)
-            if hasattr(config, "enable_new_executor"):
-                config.enable_new_executor(True)
             if "mkldnn" in self.option.run_mode:
                 try:
                     config.enable_mkldnn()
-                    config.set_cpu_math_library_num_threads(self.option.cpu_threads)
                     if "bf16" in self.option.run_mode:
                         config.enable_mkldnn_bfloat16()
                 except Exception as e:
                     logging.warning(
                         "MKL-DNN is not available. We will disable MKL-DNN."
                     )
-
-        precision_map = {
-            "trt_int8": Config.Precision.Int8,
-            "trt_fp32": Config.Precision.Float32,
-            "trt_fp16": Config.Precision.Half,
-        }
-        if self.option.run_mode in precision_map.keys():
-            config.enable_tensorrt_engine(
-                workspace_size=(1 << 25) * self.option.batch_size,
-                max_batch_size=self.option.batch_size,
-                min_subgraph_size=self.option.min_subgraph_size,
-                precision_mode=precision_map[self.option.run_mode],
-                use_static=self.option.trt_use_static,
-                use_calib_mode=self.option.trt_calib_mode,
-            )
-
-            if self.option.shape_info_filename is not None:
-                if not os.path.exists(self.option.shape_info_filename):
-                    config.collect_shape_range_info(self.option.shape_info_filename)
-                    logging.info(
-                        f"Dynamic shape info is collected into: {self.option.shape_info_filename}"
-                    )
-                else:
-                    logging.info(
-                        f"A dynamic shape info file ( {self.option.shape_info_filename} ) already exists. \
-No need to generate again."
-                    )
-                config.enable_tuned_tensorrt_dynamic_shape(
-                    self.option.shape_info_filename, True
-                )
+                config.set_mkldnn_cache_capacity(-1)
+            else:
+                if hasattr(config, "disable_mkldnn"):
+                    config.disable_mkldnn()
 
         # Disable paddle inference logging
         config.disable_glog_info()
+
+        config.set_cpu_math_library_num_threads(self.option.cpu_threads)
+
+        if not (self.option.device == "gpu" and self.option.run_mode.startswith("trt")):
+            if hasattr(config, "enable_new_ir"):
+                config.enable_new_ir(self.option.enable_new_ir)
+            if hasattr(config, "enable_new_executor"):
+                config.enable_new_executor()
+            config.set_optimization_level(3)
+
         for del_p in self.option.delete_pass:
             config.delete_pass(del_p)
 
@@ -142,11 +155,6 @@ No need to generate again."
                 config.delete_pass("conv2d_add_act_fuse_pass")
                 config.delete_pass("conv2d_add_fuse_pass")
 
-        # Enable shared memory
-        config.enable_memory_optim()
-        config.switch_ir_optim(True)
-        # Disable feed, fetch OP, needed by zero_copy_run
-        config.switch_use_feed_fetch_ops(False)
         predictor = create_predictor(config)
 
         # Get input and output handlers

paddlex/inference/utils/pp_option.py

@@ -22,6 +22,7 @@ from .new_ir_blacklist import NEWIR_BLOCKLIST
 class PaddlePredictorOption(object):
     """Paddle Inference Engine Option"""
 
+    # NOTE: TRT modes start with `trt_`
     SUPPORT_RUN_MODE = (
         "paddle",
         "trt_fp32",
@@ -69,7 +70,7 @@ class PaddlePredictorOption(object):
             "min_subgraph_size": 3,
             "shape_info_filename": None,
             "trt_calib_mode": False,
-            "cpu_threads": 1,
+            "cpu_threads": 8,
             "trt_use_static": False,
             "delete_pass": [],
             "enable_new_ir": True if self.model_name not in NEWIR_BLOCKLIST else False,