[Bugfix] fix benchmark audio timing error and add benchmark test (#1109)

Signed-off-by: wangyu31577 <wangyu31577@hundsun.com>
Co-authored-by: wangyu31577 <wangyu31577@hundsun.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

.buildkite/pipeline.yml

@@ -178,24 +178,60 @@ steps:
           volumes:
             - "/fsx/hf_cache:/fsx/hf_cache"
 
+
+  - label: "Benchmark Test"
+    timeout_in_minutes: 15
+    depends_on: image-build
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v tests/benchmarks/test_serve_cli.py
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
+
   - label: "Omni Model Test"
     timeout_in_minutes: 15
     depends_on: image-build
     commands:
-      - export VLLM_LOGGING_LEVEL=DEBUG
-      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-      - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
     agents:
       queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
     plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
+    - docker#v5.2.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+        always-pull: true
+        propagate-environment: true
+        environment:
+        - "HF_HOME=/fsx/hf_cache"
+        volumes:
+        - "/fsx/hf_cache:/fsx/hf_cache"
 
   - label: "Omni Model Test with H100"
     timeout_in_minutes: 30

docs/cli/bench/serve.md

@@ -128,7 +128,7 @@ Specify to save benchmark results to a json file
   Second item: videos with resolution 720x1280 and 16 frames
   Third item: audios with 1s duration and 5 channels w.p. 0.1
   OBS.: If the probabilities do not sum to 1, they are normalized.
-  Only the random-mm mode supports this parameter.
+  Only the random-mm mode supports this parameter
 
 ## Usage Examples
 

tests/benchmarks/test_serve_cli.py

@@ -0,0 +1,60 @@
+import subprocess
+from pathlib import Path
+
+import pytest
+
+from tests.conftest import OmniServer
+
+models = ["Qwen/Qwen3-Omni-30B-A3B-Instruct"]
+stage_configs = [str(Path(__file__).parent.parent / "e2e" / "stage_configs" / "qwen3_omni_ci.yaml")]
+
+# Create parameter combinations for model and stage config
+test_params = [(model, stage_config) for model in models for stage_config in stage_configs]
+
+
+@pytest.fixture(scope="module")
+def omni_server(request):
+    """Start vLLM-Omni server as a subprocess with actual model weights.
+    Uses session scope so the server starts only once for the entire test session.
+    Multi-stage initialization can take 10-20+ minutes.
+    """
+    model, stage_config_path = request.param
+
+    print(f"Starting OmniServer with model: {model}")
+    print("This may take 10-20+ minutes for initialization...")
+
+    with OmniServer(model, ["--stage-configs-path", stage_config_path, "--stage-init-timeout", "120"]) as server:
+        print("OmniServer started successfully")
+        yield server
+        print("OmniServer stopped")
+
+
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_bench_serve_chat(omni_server):
+    command = [
+        "vllm",
+        "bench",
+        "serve",
+        "--omni",
+        "--model",
+        omni_server.model,
+        "--port",
+        str(omni_server.port),
+        "--dataset-name",
+        "random",
+        "--random-input-len",
+        "32",
+        "--random-output-len",
+        "4",
+        "--num-prompts",
+        "5",
+        "--endpoint",
+        "/v1/chat/completions",
+        "--backend",
+        "openai-chat-omni",
+    ]
+    result = subprocess.run(command, capture_output=True, text=True)
+    print(result.stdout)
+    print(result.stderr)
+
+    assert result.returncode == 0, f"Benchmark failed: {result.stderr}"

vllm_omni/benchmarks/patch/patch.py

@@ -108,7 +108,7 @@ async def async_request_openai_chat_omni_completions(
     output.prompt_len = request_func_input.prompt_len
 
     generated_text = ""
-    generated_audio = ""
+    generated_audio = None
     ttft = 0.0
     st = time.perf_counter()
     output.start_time = st
@@ -160,7 +160,7 @@ async def async_request_openai_chat_omni_completions(
                                             if generated_audio is None:
                                                 generated_audio = seg
                                             else:
-                                                generated_audio = seg + generated_audio
+                                                generated_audio = generated_audio + seg
 
                             elif usage := data.get("usage"):
                                 output.output_tokens = usage.get("completion_tokens")