[BugFix] Async scheduling and PP compatibility with DP (#23770)

Signed-off-by: Nick Hill <nhill@redhat.com>

vllm/v1/executor/multiproc_executor.py

@@ -191,6 +191,10 @@ class MultiprocExecutor(Executor):
                 outputs, self.output_rank)
         return self.kv_output_aggregator.aggregate(outputs, self.output_rank)
 
+    def execute_dummy_batch(self) -> None:
+        self.collective_rpc("execute_dummy_batch",
+                            unique_reply_rank=self.output_rank)
+
     def take_draft_token_ids(self) -> Optional[DraftTokenIds]:
         # OPTIMIZATION: Get output only from a single worker (output_rank)
         outputs = self.collective_rpc("take_draft_token_ids",
@@ -242,12 +246,17 @@ class MultiprocExecutor(Executor):
                 dequeue_timeout = None if deadline is None else (
                     deadline - time.monotonic())
 
-                if non_block:
+                if self.io_thread_pool is not None:
+                    # We must consume worker_response_mq from a single thread.
                     result = self.io_thread_pool.submit(  # type: ignore
                         get_response, w, dequeue_timeout, self.shutdown_event)
-                else:
+                    if not non_block:
+                        result = result.result()
+                elif not non_block:
                     result = get_response(w, dequeue_timeout)
-
+                else:
+                    raise RuntimeError("non_block can only be used when"
+                                       " max_concurrent_batches > 1")
                 responses.append(result)
 
             return responses