fix: Raise error when registering Packet-keyed converter

- Packet-keyed converters are too error-prone, since new overloads could
be added or existing overloads may have been overlooked

Author: gs-olive

py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py

@@ -24,7 +24,7 @@ def args_bounds_check(
     return args[i] if len(args) > i else replacement
 
 
-@dynamo_tensorrt_converter(torch.ops.aten.batch_norm)  # type: ignore[misc]
+@dynamo_tensorrt_converter(torch.ops.aten.batch_norm.default)  # type: ignore[misc]
 def aten_ops_batch_norm(
     network: TRTNetwork,
     target: Target,
@@ -296,7 +296,7 @@ def aten_ops_hard_sigmoid(
     )
 
 
-@dynamo_tensorrt_converter(torch.ops.aten.matmul)  # type: ignore[misc]
+@dynamo_tensorrt_converter(torch.ops.aten.matmul.default)  # type: ignore[misc]
 @dynamo_tensorrt_converter(torch.ops.aten.mm.default)  # type: ignore[misc]
 @dynamo_tensorrt_converter(torch.ops.aten.mv.default)  # type: ignore[misc]
 def aten_ops_matmul(

py/torch_tensorrt/dynamo/conversion/converter_registry.py

@@ -16,6 +16,7 @@
     cast,
 )
 
+from torch._ops import OpOverloadPacket
 from torch.fx.node import Argument, Node, Target, _get_qualified_name
 from torch_tensorrt.fx.converter_registry import CONVERTERS
 from torch_tensorrt.fx.types import TRTNetwork, TRTTensor
@@ -101,6 +102,14 @@ def register_converter(converter: ConverterImplSignature) -> ConverterImplSignat
                 capability_validator=capability_validator,
             )
 
+        if isinstance(key, OpOverloadPacket):
+            raise AssertionError(
+                f"Detected converter for OpOverloadPacket {key}. "
+                "We do not support OpOverloadPacket-keyed converters. Make sure to "
+                "explicitly specify each converter overload. For instance "
+                "aten.mean is not a valid key, but aten.mean.default is."
+            )
+
         # If a converter for this operator already exists, append the new converter to the list
         # Otherwise, start a new list
         if key in DYNAMO_ATEN_CONVERTERS: