【TensorRT】Fix the situation where the filter in the convert_conv function is the output of another operator (#72302)

PolaKuma · web-flow · commit e4dbcb1788ed · 2025-04-18T12:23:11.000+08:00
* fix bugs

* fix conv3d

* fix codestyle

* fix

* fix codestyle

* Update converter_utils.py

* Update converter_utils.py

* Update converter_utils.py

* fix codestyle

* Update converter_utils.py

* fix
diff --git a/python/paddle/tensorrt/converter_utils.py b/python/paddle/tensorrt/converter_utils.py
@@ -732,10 +732,12 @@ def convert_conv2d(network, paddle_op, inputs):
     set_layer_name(layer, paddle_op)
     support_fp32_mix_precision(paddle_op.name(), layer)
 
-    filter_param = paddle_op.operands()[1].source()
-    filter_name = filter_param.get_defining_op().attrs()['parameter_name']
-    refit_manager = RefitManager()
-    refit_manager.set_mapping(filter_name, filter_name, RefitRole.CONSTANT)
+    trt_manager = TensorRTConfigManager()
+    if trt_manager.get_refit_params_path():
+        filter_param = paddle_op.operands()[1].source()
+        filter_name = filter_param.get_defining_op().attrs()['parameter_name']
+        refit_manager = RefitManager()
+        refit_manager.set_mapping(filter_name, filter_name, RefitRole.CONSTANT)
 
     return layer.get_output(0)
 
@@ -817,10 +819,12 @@ def convert_conv3d(network, paddle_op, inputs):
 
     layer.dilation_nd = nv_dilations
     set_layer_name(layer, paddle_op)
-    filter_param = paddle_op.operands()[1].source()
-    filter_name = filter_param.get_defining_op().attrs()['parameter_name']
-    refit_manager = RefitManager()
-    refit_manager.set_mapping(filter_name, filter_name, RefitRole.CONSTANT)
+    trt_manager = TensorRTConfigManager()
+    if trt_manager.get_refit_params_path():
+        filter_param = paddle_op.operands()[1].source()
+        filter_name = filter_param.get_defining_op().attrs()['parameter_name']
+        refit_manager = RefitManager()
+        refit_manager.set_mapping(filter_name, filter_name, RefitRole.CONSTANT)
 
     return layer.get_output(0)
 
diff --git a/python/paddle/tensorrt/export.py b/python/paddle/tensorrt/export.py
@@ -330,6 +330,8 @@ def __init__(
         self.workspace_size = workspace_size
         self.use_cuda_graph = use_cuda_graph
         self.refit_params_path = refit_params_path
+        if self.refit_params_path:
+            self.disable_passes.append("constant_folding_pass")
         paddle.framework.set_flags(
             {'FLAGS_trt_min_group_size': min_subgraph_size}
         )
diff --git a/python/paddle/tensorrt/util.py b/python/paddle/tensorrt/util.py
@@ -264,6 +264,11 @@ def get_force_fp32_ops(self):
             return self.trt_config.ops_run_float
         return []
 
+    def get_refit_params_path(self):
+        if self.trt_config and self.trt_config.refit_params_path:
+            return self.trt_config.refit_params_path
+        return None
+
 
 class TensorRTConstantManager:
     _instance = None

Original file line number	Diff line number	Diff line change
`@@ -330,6 +330,8 @@ def __init__(`
`330`	`330`	`self.workspace_size = workspace_size`
`331`	`331`	`self.use_cuda_graph = use_cuda_graph`
`332`	`332`	`self.refit_params_path = refit_params_path`
	`333`	`+ if self.refit_params_path:`
	`334`	`+ self.disable_passes.append("constant_folding_pass")`
`333`	`335`	`paddle.framework.set_flags(`
`334`	`336`	`{'FLAGS_trt_min_group_size': min_subgraph_size}`
`335`	`337`	`)`