diff --git a/vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py b/vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py index d96ef74db09..aed8d19b197 100644 --- a/vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py +++ b/vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py @@ -295,6 +295,7 @@ def process_pending_chunks( ) while len(running_queue) > self.scheduler_max_num_seqs: request = running_queue.pop() + request.status = RequestStatus.PREEMPTED waiting_queue.prepend_requests([request]) def restore_queues(self, waiting_queue: Any, running_queue: list[Request]) -> None: