diff --git a/examples/offline_inference_vision_language.py b/examples/offline_inference_vision_language.py
index a2e35ac365a2c..17ca7ca37398d 100644
--- a/examples/offline_inference_vision_language.py
+++ b/examples/offline_inference_vision_language.py
@@ -6,6 +6,7 @@
 on HuggingFace model repository.
 """
 import time
+
 from transformers import AutoTokenizer
 
 from vllm import LLM, SamplingParams
@@ -24,8 +25,7 @@ def run_llava(question: str, modality: str):
 
     prompt = f"USER: <image>\n{question}\nASSISTANT:"
 
-    llm = LLM(model="llava-hf/llava-1.5-7b-hf",
-              max_model_len=4096)
+    llm = LLM(model="llava-hf/llava-1.5-7b-hf", max_model_len=4096)
 
     stop_token_ids = None
     return llm, prompt, stop_token_ids
diff --git a/vllm/inputs/data.py b/vllm/inputs/data.py
index ce5895ca4bb49..266bd32d86946 100644
--- a/vllm/inputs/data.py
+++ b/vllm/inputs/data.py
@@ -7,7 +7,8 @@
 from typing_extensions import NotRequired, TypedDict, TypeVar, assert_never
 
 if TYPE_CHECKING:
-    from vllm.multimodal import MultiModalDataDict, MultiModalKwargs, MultiModalPlaceholderDict
+    from vllm.multimodal import (MultiModalDataDict, MultiModalKwargs,
+                                 MultiModalPlaceholderDict)
     from vllm.multimodal.inputs import MultiModalInputsV2
 
 
@@ -265,14 +266,14 @@ def multi_modal_data(self) -> "MultiModalDataDict":
         assert_never(inputs)
 
     @cached_property
-    def multi_modal_inputs(self) -> "MultiModalKwargs":
+    def multi_modal_inputs(self) -> Union[Dict, "MultiModalKwargs"]:
         inputs = self.inputs
 
         if inputs["type"] == "token":
             return inputs.get("multi_modal_inputs", {})
 
         if inputs["type"] == "multimodal":
-            return inputs.get("mm_inputs", {})
+            return inputs.get("mm_kwargs", {})
 
         assert_never(inputs)
 
diff --git a/vllm/v1/engine/__init__.py b/vllm/v1/engine/__init__.py
index 02b35cc273257..083141d811c53 100644
--- a/vllm/v1/engine/__init__.py
+++ b/vllm/v1/engine/__init__.py
@@ -1,12 +1,11 @@
 import enum
 from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Union
+from typing import List, Optional, Union
 
 import msgspec
 
 from vllm.lora.request import LoRARequest
-from vllm.multimodal import (MultiModalDataDict, MultiModalKwargs,
-                             MultiModalPlaceholderDict)
+from vllm.multimodal import MultiModalKwargs, MultiModalPlaceholderDict
 from vllm.sampling_params import RequestOutputKind, SamplingParams
 
 
diff --git a/vllm/v1/engine/core.py b/vllm/v1/engine/core.py
index b3be92470ed6e..31773b3eed5f7 100644
--- a/vllm/v1/engine/core.py
+++ b/vllm/v1/engine/core.py
@@ -14,7 +14,6 @@
 from vllm.config import CacheConfig, VllmConfig
 from vllm.logger import init_logger
 from vllm.usage.usage_lib import UsageContext
-from vllm.multimodal import MultiModalDataDict, MultiModalKwargs
 from vllm.v1.core.scheduler import Scheduler
 from vllm.v1.engine import (EngineCoreOutput, EngineCoreOutputs,
                             EngineCoreRequest, EngineCoreRequestType)