diff --git a/vllm/worker/cpu_model_runner.py b/vllm/worker/cpu_model_runner.py
index cebb0f36a2b28..534d167d994fe 100644
--- a/vllm/worker/cpu_model_runner.py
+++ b/vllm/worker/cpu_model_runner.py
@@ -133,7 +133,7 @@ def build(self) -> ModelInputForCPU:
             (input_tokens, input_positions,
              attn_metadata) = self._prepare_decode(
                  self.seq_group_metadata_list)
-            seq_lens = []
+            seq_lens = None
 
         return self.model_input_cls(
             input_tokens=input_tokens,