diff --git a/src/imitater/model/chat_model.py b/src/imitater/model/chat_model.py
index 02007c4..9b51f50 100644
--- a/src/imitater/model/chat_model.py
+++ b/src/imitater/model/chat_model.py
@@ -74,6 +74,7 @@ def _init_vllm_engine(self) -> None:
             trust_remote_code=True,
             max_model_len=self.config.maxlen,
             tensor_parallel_size=len(self.config.device),
+            dtype="float16",
         )
         self._engine = AsyncLLMEngine.from_engine_args(engine_args)