فهرست منبع

fix Paddle-TRT some model within FP16

zhink 8 ماه پیش
والد
کامیت
80cb6da064
2فایلهای تغییر یافته به همراه12 افزوده شده و 3 حذف شده
  1. 1 1
      paddlex/inference/models/common/static_infer.py
  2. 11 2
      paddlex/inference/utils/trt_config.py

+ 1 - 1
paddlex/inference/models/common/static_infer.py

@@ -534,7 +534,7 @@ class StaticInfer(object):
                             self._option.trt_dynamic_shapes,
                             self._option.trt_dynamic_shape_input_data,
                         )
-                    if self._option.model_name in DISABLE_TRT_HALF_OPS_CONFIG:
+                    if self._option.model_name in DISABLE_TRT_HALF_OPS_CONFIG and self._option.run_mode == "trt_fp16":
                         lazy_paddle.inference.InternalUtils.disable_tensorrt_half_ops(
                             config, DISABLE_TRT_HALF_OPS_CONFIG[self._option.model_name]
                         )

+ 11 - 2
paddlex/inference/utils/trt_config.py

@@ -190,8 +190,17 @@ OLD_IR_TRT_CFG_SETTING = {
 }
 
 DISABLE_TRT_HALF_OPS_CONFIG = {
-    # TODO: just for example
-    "model_name": {"layer_norm"}
+    "ConvNeXt_tiny": {"layer_norm"},
+    "ConvNeXt_small": {"layer_norm"},
+    "ConvNeXt_base_224": {"layer_norm"},
+    "ConvNeXt_large_224": {"layer_norm"},
+    "ConvNeXt_base_384": {"layer_norm"},
+    "ConvNeXt_large_384": {"layer_norm"},
+    "PP-HGNetV2-B3": {"softmax"},
+    "MobileNetV1_x0_5": {"fused_conv2d_add_act"},
+    "SeaFormer_small": {"fused_conv2d_add_act"},
+    "SeaFormer_tiny": {"fused_conv2d_add_act"},
+    "PP-OCRv4_mobile_seal_det": {"fused_conv2d_add_act","softmax"},
 }
 
 ############ pir trt ############