diff --git a/vllm/compilation/backends.py b/vllm/compilation/backends.py
index 0c7bbfe599b0..826d1744d88a 100644
--- a/vllm/compilation/backends.py
+++ b/vllm/compilation/backends.py
@@ -141,14 +141,14 @@ def produce_guards_expression(self, *args, **kwargs):
         return ""
 
 
-def wrap_inductor(graph,
+def wrap_inductor(graph: fx.GraphModule,
                   example_inputs,
                   additional_inductor_config,
                   compilation_config: CompilationConfig,
                   graph_index: int = 0,
                   num_graphs: int = 1,
                   runtime_shape: Optional[int] = None,
-                  use_inductor: bool = True):
+                  use_inductor: bool = True) -> Any:
     if graph_index == 0:
         # before compiling the first graph, record the start time
         global compilation_start_time
@@ -209,7 +209,7 @@ def wrap_inductor(graph,
         returns_tuple = graph_returns_tuple(graph)
 
         # this is the graph we return to Dynamo to run
-        def compiled_graph(*args):
+        def compiled_graph(*args) -> Optional[fx.CompiledFxGraph]:
             # convert args to list
             list_args = list(args)
             graph_output = inductor_compiled_graph(list_args)
@@ -247,7 +247,7 @@ def _check_can_cache(*args, **kwargs):
             # see https://github.com/pytorch/pytorch/blob/9f5ebf3fc609105a74eab4ccc24932d6353ff566/torch/_inductor/codecache.py#L1221 # noqa
             return
 
-        def _get_shape_env():
+        def _get_shape_env() -> AlwaysHitShapeEnv:
             return AlwaysHitShapeEnv()
 
         with patch(# for hijacking the hash of the compiled graph
@@ -537,7 +537,7 @@ def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable:
             example_inputs[x].clone() for x in self.sym_tensor_indices
         ]
 
-        def copy_and_call(*args):
+        def copy_and_call(*args) -> fx.GraphModule:
             list_args = list(args)
             for i, index in enumerate(self.sym_tensor_indices):
                 runtime_tensor = list_args[index]
diff --git a/vllm/compilation/multi_output_match.py b/vllm/compilation/multi_output_match.py
index 0ad648abfbb3..b6bcecdc89e2 100644
--- a/vllm/compilation/multi_output_match.py
+++ b/vllm/compilation/multi_output_match.py
@@ -7,6 +7,7 @@
 from torch._higher_order_ops.auto_functionalize import auto_functionalized
 from torch._inductor import pattern_matcher as pm
 from torch._ops import OpOverload
+from torch.fx import Node
 
 from vllm.compilation.fx_utils import find_auto_fn
 
@@ -97,7 +98,7 @@ def insert_getitems(self, tuple_node: fx.Node,
                 self.graph.call_function(operator.getitem, (tuple_node, idx))
                 for idx in indices)
 
-    def insert_auto_fn(self, op: OpOverload, kwargs):
+    def insert_auto_fn(self, op: OpOverload, kwargs) -> Node:
         """
         Insert an auto_functionalized node with the given op and kwargs.
         """
diff --git a/vllm/compilation/pass_manager.py b/vllm/compilation/pass_manager.py
index fb522ae053e9..34f5f355798b 100644
--- a/vllm/compilation/pass_manager.py
+++ b/vllm/compilation/pass_manager.py
@@ -1,4 +1,4 @@
-from typing import List
+from typing import Any, Dict, List
 
 from torch import fx as fx
 
@@ -53,7 +53,7 @@ def add(self, pass_: InductorPass):
         assert isinstance(pass_, InductorPass)
         self.passes.append(pass_)
 
-    def __getstate__(self):
+    def __getstate__(self) -> Dict[str, List[Any]]:
         """
         Custom pickling for the pass manager, as some passes cannot be pickled.
         Pickling occurs because the pass manager is set as the value of