[Feature] Support Qwen Omni online batch inference (vllm-project#438)

Signed-off-by: ZeldaHuang <[email protected]>
Signed-off-by: wangyu31577 <[email protected]>

Author: ZeldaHuang

tests/e2e/online_serving/test_qwen3_omni.py

@@ -4,6 +4,7 @@
 E2E Online tests for Qwen3-Omni model with video input and audio output.
 """
 
+import concurrent.futures
 import os
 import socket
 import subprocess
@@ -167,40 +168,55 @@ def dummy_messages_from_video_data(
 
 
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
-def test_video_to_audio(
+def test_video_to_audio_concurrent(
     client: openai.OpenAI,
     omni_server,
     base64_encoded_video: str,
 ) -> None:
-    """Test processing video, generating audio output via OpenAI API."""
+    """Test processing video with multiple concurrent completions, generating audio output via OpenAI API."""
     # Create data URL for the base64 encoded video
     video_data_url = f"data:video/mp4;base64,{base64_encoded_video}"
 
     messages = dummy_messages_from_video_data(video_data_url)
 
-    # Test single completion
-    chat_completion = client.chat.completions.create(
-        model=omni_server.model,
-        messages=messages,
-    )
-
-    assert len(chat_completion.choices) == 2  # 1 for text output, 1 for audio output
-
-    # Verify text output
-    text_choice = chat_completion.choices[0]
-    assert text_choice.finish_reason == "length"
-
-    # Verify we got a response
-    text_message = text_choice.message
-    assert text_message.content is not None and len(text_message.content) >= 10
-    assert text_message.role == "assistant"
-
-    # Verify audio output
-    audio_choice = chat_completion.choices[1]
-    assert audio_choice.finish_reason == "stop"
-    audio_message = audio_choice.message
-
-    # Check if audio was generated
-    if hasattr(audio_message, "audio") and audio_message.audio:
-        assert audio_message.audio.data is not None
-        assert len(audio_message.audio.data) > 0
+    # Test multiple concurrent completions
+    num_concurrent_requests = 5
+
+    with concurrent.futures.ThreadPoolExecutor(max_workers=num_concurrent_requests) as executor:
+        # Submit multiple completion requests concurrently
+        futures = [
+            executor.submit(
+                client.chat.completions.create,
+                model=omni_server.model,
+                messages=messages,
+            )
+            for _ in range(num_concurrent_requests)
+        ]
+
+        # Wait for all requests to complete and collect results
+        chat_completions = [future.result() for future in concurrent.futures.as_completed(futures)]
+
+    # Verify all completions succeeded
+    assert len(chat_completions) == num_concurrent_requests
+
+    for chat_completion in chat_completions:
+        assert len(chat_completion.choices) == 2  # 1 for text output, 1 for audio output
+
+        # Verify text output
+        text_choice = chat_completion.choices[0]
+        assert text_choice.finish_reason == "length"
+
+        # Verify we got a response
+        text_message = text_choice.message
+        assert text_message.content is not None and len(text_message.content) >= 10
+        assert text_message.role == "assistant"
+
+        # Verify audio output
+        audio_choice = chat_completion.choices[1]
+        assert audio_choice.finish_reason == "stop"
+        audio_message = audio_choice.message
+
+        # Check if audio was generated
+        if hasattr(audio_message, "audio") and audio_message.audio:
+            assert audio_message.audio.data is not None
+            assert len(audio_message.audio.data) > 0

vllm_omni/entrypoints/omni_stage.py

@@ -16,6 +16,7 @@
 import logging
 import multiprocessing as mp
 import os
+import queue
 import sys
 import traceback
 from typing import Any
@@ -1135,19 +1136,14 @@ def filter(self, record: _logging.LogRecord) -> bool:
         )
     except Exception as e:
         _logging.getLogger(__name__).warning("[Stage-%s] Failed to send stage ready signal: %s", stage_id, e)
-
+    generation_out_q = asyncio.Queue()
     # Batch processing loop
-    while True:
-        task = in_q.get()
-        _recv_dequeue_ts = _time.time()
-        if task is None:
-            _logging.getLogger(__name__).debug("[Stage-%s] Received shutdown signal", stage_id)
-            break
-
-        _rx_bytes_by_rid: dict[Any, int] = {}
-        _rx_decode_ms_by_rid: dict[Any, float] = {}
-        _in_flight_ms_by_rid: dict[Any, float] = {}
+    _rx_bytes_by_rid: dict[Any, int] = {}
+    _rx_decode_ms_by_rid: dict[Any, float] = {}
+    _in_flight_ms_by_rid: dict[Any, float] = {}
 
+    async def generation_single_request(task: dict[str, Any]):
+        _recv_dequeue_ts = _time.time()
         rid = task["request_id"]
         try:
             sent_ts = float(task.get("sent_ts", None)) if isinstance(task, dict) else None
@@ -1157,62 +1153,101 @@ def filter(self, record: _logging.LogRecord) -> bool:
                 _in_flight_ms_by_rid[rid] = 0.0
         except Exception:
             _in_flight_ms_by_rid[rid] = 0.0
-        ein, _rx_metrics = try_recv_via_connector(
-            task=task,
-            connectors=connectors,
-            stage_id=stage_id,
-        )
-        if ein is None or _rx_metrics is None:
-            raise RuntimeError(
-                f"[Stage-{stage_id}] Missing connector payload for request {rid}. "
-                "Ensure connectors are configured for all incoming edges."
-            )
-        _rx_decode_ms_by_rid[rid] = float(_rx_metrics.get("rx_decode_time_ms", 0.0))
-        _rx_bytes_by_rid[rid] = int(_rx_metrics.get("rx_transfer_bytes", 0))
-
-        sampling_params = task["sampling_params"]
-        _logging.getLogger(__name__).debug("[Stage-%s] Received batch size=1, request_ids=%s", stage_id, rid)
-        print("--------------------------------", flush=True)
-        print(f"[Stage-{stage_id}] Received batch size=1, request_ids={rid}", flush=True)
-        print("--------------------------------", flush=True)
         try:
-            _batch_seq += 1
+            ein, _rx_metrics = try_recv_via_connector(
+                task=task,
+                connectors=connectors,
+                stage_id=stage_id,
+            )
+            if ein is None or _rx_metrics is None:
+                raise RuntimeError(
+                    f"[Stage-{stage_id}] Missing connector payload for request {rid}. "
+                    "Ensure connectors are configured for all incoming edges."
+                )
+            _rx_decode_ms_by_rid[rid] = float(_rx_metrics.get("rx_decode_time_ms", 0.0))
+            _rx_bytes_by_rid[rid] = int(_rx_metrics.get("rx_transfer_bytes", 0))
+
+            sampling_params = task["sampling_params"]
+            _logging.getLogger(__name__).debug("[Stage-%s] Received batch size=1, request_ids=%s", stage_id, rid)
+            print("--------------------------------", flush=True)
+            print(f"[Stage-{stage_id}] Received batch size=1, request_ids={rid}", flush=True)
+            print("--------------------------------", flush=True)
             _gen_t0 = _time.time()
             if isinstance(ein, list):
                 ein = ein[0]
-
             async for res in stage_engine.generate(ein, sampling_params, rid):
                 gen_output = res
             _gen_t1 = _time.time()
             _gen_ms = (_gen_t1 - _gen_t0) * 1000.0
+            await generation_out_q.put((rid, gen_output, _gen_ms))
+        except Exception as e:
+            _logging.getLogger(__name__).exception("[Stage-%s] Failed on request %s: %s", stage_id, rid, e)
+            out_q.put(
+                {
+                    "request_id": rid,
+                    "stage_id": stage_id,
+                    "error": str(e),
+                }
+            )
 
-            r_outputs = [gen_output]
-            _num_tokens = count_tokens_from_outputs(r_outputs)
-            _agg_total_tokens += _num_tokens
-            _agg_total_gen_time_ms += _gen_ms
-
-            if _stats_file:
-                _avg_tokens_per_s = (
-                    (_agg_total_tokens * 1000.0 / _agg_total_gen_time_ms) if _agg_total_gen_time_ms > 0 else 0.0
-                )
-                log_stage_running_avg(
-                    _stats_file,
-                    stage_id,
-                    int(_agg_total_tokens),
-                    float(_agg_total_gen_time_ms),
-                    float(_avg_tokens_per_s),
-                )
+    _batch_gen_t0 = _time.time()
+    while True:
+        try:
+            task = in_q.get_nowait()
+            if task is None:
+                _logging.getLogger(__name__).debug("[Stage-%s] Received shutdown signal", stage_id)
+                break
+            asyncio.create_task(generation_single_request(task))
+        except queue.Empty:
+            await asyncio.sleep(0.001)
+        batch_request_outputs: list[Any] = []
+        batch_request_ids: list[Any] = []
+        _gen_ms_list = []
+        while True:
+            try:
+                rids, gen_output, _gen_ms = generation_out_q.get_nowait()
+                _num_tokens = count_tokens_from_outputs([gen_output])
+                batch_request_outputs.append(gen_output)
+                _gen_ms_list.append(_gen_ms)
+                batch_request_ids.append(rids)
+                _agg_total_tokens += _num_tokens
+            except asyncio.QueueEmpty:
+                await asyncio.sleep(0.001)
+                break
+
+        if not batch_request_outputs:
+            continue
+        _batch_seq += 1
+        if _stats_file:
+            _batch_gen_t1 = _time.time()
+            _agg_total_gen_time_ms += (_batch_gen_t1 - _batch_gen_t0) * 1000
+            _batch_gen_t0 = _batch_gen_t1
+            _avg_tokens_per_s = (
+                (_agg_total_tokens * 1000.0 / _agg_total_gen_time_ms) if _agg_total_gen_time_ms > 0 else 0.0
+            )
+            log_stage_running_avg(
+                _stats_file,
+                stage_id,
+                int(_agg_total_tokens),
+                float(_agg_total_gen_time_ms),
+                float(_avg_tokens_per_s),
+            )
+            logger.info("[Stage-%s] Running avg: %s tokens/s", stage_id, _avg_tokens_per_s)
+            for rid, _gen_ms in zip(batch_request_ids, _gen_ms_list):
                 log_stage_batch_stats(_stats_file, stage_id, 1, float(_gen_ms), [rid])
 
+        logger.info("[Stage-%s] Sending outputs to main process", stage_id)
+        for rid, output, _gen_ms in zip(batch_request_ids, batch_request_outputs, _gen_ms_list):
             try:
+                r_outputs = [output]
                 use_shm, payload = maybe_dump_to_shm(r_outputs, shm_threshold_bytes)
                 _metrics = {
                     "num_tokens_out": int(count_tokens_from_outputs(r_outputs)),
                     "stage_gen_time_ms": _gen_ms,
                     "batch_id": int(_batch_seq),
-                    "rx_decode_time_ms": float(_rx_decode_ms_by_rid.get(rid, 0.0)),
-                    "rx_transfer_bytes": int(_rx_bytes_by_rid.get(rid, 0)),
-                    "rx_in_flight_time_ms": float(_in_flight_ms_by_rid.get(rid, 0.0)),
+                    "rx_decode_time_ms": float(_rx_decode_ms_by_rid.pop(rid, 0.0)),
+                    "rx_transfer_bytes": int(_rx_bytes_by_rid.pop(rid, 0)),
+                    "rx_in_flight_time_ms": float(_in_flight_ms_by_rid.pop(rid, 0.0)),
                 }
                 if _stats_file:
                     compute_and_log_stage_request_stats(
@@ -1266,23 +1301,14 @@ def filter(self, record: _logging.LogRecord) -> bool:
                         "metrics": {
                             "num_tokens_out": int(count_tokens_from_outputs(r_outputs)),
                             "stage_gen_time_ms": _gen_ms,
-                            "rx_decode_time_ms": float(_rx_decode_ms_by_rid.get(rid, 0.0)),
-                            "rx_transfer_bytes": int(_rx_bytes_by_rid.get(rid, 0)),
-                            "rx_in_flight_time_ms": float(_in_flight_ms_by_rid.get(rid, 0.0)),
+                            "rx_decode_time_ms": float(_rx_decode_ms_by_rid.pop(rid, 0.0)),
+                            "rx_transfer_bytes": int(_rx_bytes_by_rid.pop(rid, 0)),
+                            "rx_in_flight_time_ms": float(_in_flight_ms_by_rid.pop(rid, 0.0)),
                         },
                     }
                 )
             _logging.getLogger(__name__).debug("[Stage-%s] Enqueued result for request %s to downstream", stage_id, rid)
 
-        except Exception as e:
-            _logging.getLogger(__name__).exception("[Stage-%s] Failed on request %s: %s", stage_id, rid, e)
-            out_q.put(
-                {
-                    "request_id": rid,
-                    "stage_id": stage_id,
-                    "error": str(e),
-                }
-            )
     print("--------------------------------", flush=True)
     print(f"[Stage-{stage_id}] Stage worker exiting", flush=True)
     print("--------------------------------", flush=True)