Removing empty_like from evaluator and adding it to lowering pass

Author: apbose

py/torch_tensorrt/dynamo/conversion/ops_evaluators.py

@@ -47,14 +47,3 @@ def aten_ops_arange_start_step(
     name: str,
 ) -> Union[TRTTensor, Sequence[TRTTensor]]:
     return np.arange(*args)
-
-
-@dynamo_tensorrt_converter(torch.ops.aten.empty_like.default)
-def aten_ops_empty_like(
-    ctx: ConversionContext,
-    target: Target,
-    args: Tuple[Argument, ...],
-    kwargs: Dict[str, Argument],
-    name: str,
-) -> TRTTensor:
-    return np.empty_like(*args)

py/torch_tensorrt/dynamo/lowering/_decomposition_groups.py

@@ -37,6 +37,7 @@
     aten.elu_backward,
     aten._embedding_bag,
     aten.embedding_dense_backward,
+    aten.empty_like,
     aten._euclidean_dist.default,
     aten.expand_as,
     aten.eye,