[Bugfix] Keep resumable duplex prompts out of async-chunk prewarm (#7992)

Signed-off-by: Tianyao Wu <rayroy31@gmail.com>

tests/engine/test_duplex_orchestrator.py

@@ -307,6 +307,24 @@ async def test_append_submits_the_resumable_stage0_request_and_counts_it_running
     assert counter.value == 0
 
 
+@pytest.mark.asyncio
+async def test_resumable_append_prompt_does_not_reach_prewarmed_stages() -> None:
+    # A resumable append carries Stage0's duplex buffer. The async-chunk
+    # prewarm copies ``request_state.prompt`` into downstream placeholders,
+    # so that buffer must not end up there (#7962).
+    orchestrator, clients, rpc_q, _ = _build(stages=2)
+    orchestrator.async_chunk = True
+    orchestrator._stage_receives_async_chunks = lambda stage_id: stage_id > 0  # type: ignore[method-assign]
+    await _open(orchestrator, rpc_q)
+
+    await _submit(orchestrator, _append_audio())
+
+    assert clients[0].add_request_calls[0][0].model_intermediate_buffer
+    assert len(clients[1].add_request_calls) == 1
+    assert clients[1].add_request_calls[0][0].model_intermediate_buffer is None
+    await orchestrator.session_manager.shutdown()
+
+
 @pytest.mark.asyncio
 async def test_session_update_refreshes_the_next_append_sampling_params() -> None:
     orchestrator, clients, rpc_q, _ = _build()

vllm_omni/engine/duplex_orchestrator.py

@@ -357,9 +357,6 @@ class DuplexOrchestrator(Orchestrator, DuplexStagePort):
         if not isinstance(request_state, DuplexOrchestratorRequestState):
             raise RuntimeError(f"duplex request was not preregistered: {context.request_id}")
         request_state.streaming.enabled = submission.resumable
-        # Keep raw Stage0 prompt (additional_information / multi_modal_data) for
-        # stage input processors via process_engine_inputs.
-        request_state.prompt = dict(submission.prompt)
         if submission.resumable:
             request = build_engine_core_request_from_tokens(
                 request_id=context.request_id,
@@ -369,6 +366,11 @@ class DuplexOrchestrator(Orchestrator, DuplexStagePort):
                 resumable=True,
             )
         else:
+            # Keep raw Stage0 prompt (additional_information / multi_modal_data) for
+            # stage input processors via process_engine_inputs. Resumable requests
+            # leave it unset: their prompt is one append with Stage0's duplex
+            # buffer, which the async-chunk prewarm would copy downstream.
+            request_state.prompt = dict(submission.prompt)
             # Use the ordinary multimodal input processor for turn-model plugins.
             # Its CPU preprocessing runs off the session/orchestrator event loop.
             request = await asyncio.to_thread(