diff --git a/vllm/config.py b/vllm/config.py
index fbd61a332af61..7e0b75eceae5b 100644
--- a/vllm/config.py
+++ b/vllm/config.py
@@ -347,10 +347,10 @@ def verify_async_output_proc(self, parallel_config, speculative_config,
             self.use_async_output_proc = False
             return
 
-        if device_config.device_type != "cuda":
+        if device_config.device_type not in ("cuda", "tpu"):
             logger.warning(
-                "Async output processing is only supported for CUDA."
-                " Disabling it for other platforms.")
+                "Async output processing is only supported for CUDA or TPU. "
+                "Disabling it for other platforms.")
             self.use_async_output_proc = False
             return
 
diff --git a/vllm/worker/tpu_model_runner.py b/vllm/worker/tpu_model_runner.py
index ebb4b89cb4727..a0498315516b8 100644
--- a/vllm/worker/tpu_model_runner.py
+++ b/vllm/worker/tpu_model_runner.py
@@ -1,6 +1,7 @@
 import time
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type, Union
+from typing import (TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple,
+                    Type, Union)
 from unittest.mock import patch
 
 import numpy as np
@@ -51,6 +52,7 @@ class ModelInputForTPU(ModelRunnerInputBase):
     best_of: List[int]
     seq_groups: List[List[int]]
     virtual_engine: int = 0
+    async_callback: Optional[Callable] = None
 
     def as_broadcastable_tensor_dict(
             self) -> Dict[str, Union[int, torch.Tensor]]:
@@ -562,6 +564,8 @@ def _execute_model(*args):
                     model_input.attn_metadata, model_input.input_lens[i:i + 1],
                     model_input.t[i:i + 1], model_input.p[i:i + 1],
                     model_input.num_samples, kv_caches)
+                if i == 0 and model_input.async_callback is not None:
+                    model_input.async_callback()
                 # Retrieve the outputs to CPU.
                 next_token_ids += output_token_ids.cpu().tolist()
                 start_idx = end_idx
@@ -572,6 +576,8 @@ def _execute_model(*args):
                 model_input.attn_metadata, model_input.input_lens,
                 model_input.t, model_input.p, model_input.num_samples,
                 kv_caches)
+            if model_input.async_callback is not None:
+                model_input.async_callback()
             # Retrieve the outputs to CPU.
             next_token_ids = output_token_ids.cpu().tolist()