diff --git a/vllm/worker/tpu_model_runner.py b/vllm/worker/tpu_model_runner.py
index 12e4215038d74..dd122f9f1272b 100644
--- a/vllm/worker/tpu_model_runner.py
+++ b/vllm/worker/tpu_model_runner.py
@@ -543,7 +543,8 @@ def execute_model(
                         seq_group_metadata_list=ctx.seq_group_metadata_list,
                         scheduler_outputs=ctx.scheduler_outputs,
                         is_async=False,
-                        is_last_step=False)
+                        is_last_step=False,
+                        is_first_step_output=i == 0)
                     model_input.async_callback()
             if use_async_out_proc:
                 return [sampler_outputs[-1]]