diff --git a/vllm/_custom_ops.py b/vllm/_custom_ops.py
index 19f31b8ec419d..aeacf5dda5761 100644
--- a/vllm/_custom_ops.py
+++ b/vllm/_custom_ops.py
@@ -1,5 +1,4 @@
 import contextlib
-import functools
 import importlib
 from typing import TYPE_CHECKING, List, Optional, Tuple, Union
 
@@ -36,34 +35,6 @@ def register_fake(fn):
         from torch.library import impl_abstract as register_fake
 
 
-def hint_on_error(fn):
-
-    @functools.wraps(fn)
-    def wrapper(*args, **kwargs):
-        try:
-            return fn(*args, **kwargs)
-
-        except NotImplementedError as e:
-            msg = (
-                "Error in calling custom op %s: %s\n"
-                "Not implemented or built, mostly likely because the current current device "
-                "does not support this kernel (less likely TORCH_CUDA_ARCH_LIST was set "
-                "incorrectly while building)")
-            logger.error(msg, fn.__name__, e)
-            raise NotImplementedError(msg % (fn.__name__, e)) from e
-        except AttributeError as e:
-            msg = (
-                "Error in calling custom op %s: %s\n"
-                "Possibly you have built or installed an obsolete version of vllm.\n"
-                "Please try a clean build and install of vllm,"
-                "or remove old built files such as vllm/*cpython*.so and build/ ."
-            )
-            logger.error(msg, fn.__name__, e)
-            raise e
-
-    return wrapper
-
-
 # activation ops
 def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
     torch.ops._C.silu_and_mul(out, x)
@@ -1101,25 +1072,3 @@ def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
 def register_graph_buffers(fa: int, handles: List[List[int]],
                            offsets: List[List[int]]) -> None:
     torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)
-
-
-# temporary fix for https://github.com/vllm-project/vllm/issues/5456
-# TODO: remove this in v0.6.0
-names_and_values = globals()
-names_and_values_to_update = {}
-# prepare variables to avoid dict size change during iteration
-k, v, arg = None, None, None
-fn_type = type(lambda x: x)
-for k, v in names_and_values.items():
-    # find functions that are defined in this file and have torch.Tensor
-    # in their annotations. `arg == "torch.Tensor"` is used to handle
-    # the case when users use `import __annotations__` to turn type
-    # hints into strings.
-    if isinstance(v, fn_type) \
-        and v.__code__.co_filename == __file__ \
-        and any(arg is torch.Tensor or arg == "torch.Tensor"
-                for arg in v.__annotations__.values()):
-        names_and_values_to_update[k] = hint_on_error(v)
-
-names_and_values.update(names_and_values_to_update)
-del names_and_values_to_update, names_and_values, v, k, fn_type