Accelerate network interpretation by 15x; fixed redundant code in TRT Interpreter (#3396)

Author: cehongwang

py/torch_tensorrt/dynamo/conversion/_TRTInterpreter.py

@@ -33,7 +33,9 @@
 from torch_tensorrt.dynamo.conversion._ConverterRegistry import (
     DYNAMO_CONVERTERS as CONVERTERS,
 )
-from torch_tensorrt.dynamo.conversion._ConverterRegistry import CallingConvention
+from torch_tensorrt.dynamo.conversion._ConverterRegistry import (
+    CallingConvention,
+)
 from torch_tensorrt.dynamo.conversion._TRTBuilderMonitor import TRTBulderMonitor
 from torch_tensorrt.dynamo.conversion.converter_utils import (
     get_node_io,
@@ -740,10 +742,6 @@ def run(
     def run_node(self, n: torch.fx.Node) -> torch.fx.Node:
         self._cur_node_name = get_node_name(n)
         self._cur_node = n
-        # add "_itensor_to_tensor_meta"
-        kwargs = dict(n.kwargs)
-        kwargs["_itensor_to_tensor_meta"] = self._itensor_to_tensor_meta
-        n.kwargs = kwargs
 
         if _LOGGER.isEnabledFor(logging.DEBUG):
             _LOGGER.debug(
@@ -759,11 +757,6 @@ def run_node(self, n: torch.fx.Node) -> torch.fx.Node:
             f"Converted node {self._cur_node_name} [{n.target}] ({get_node_io(n, self.const_mapping)})"
         )
 
-        # remove "_itensor_to_tensor_meta"
-        kwargs = dict(n.kwargs)
-        del kwargs["_itensor_to_tensor_meta"]
-        n.kwargs = kwargs
-
         if isinstance(trt_node, trt.ITensor):
             self._itensor_to_tensor_meta[trt_node] = n.meta.get("tensor_meta")