[Refactor] Migrate Step-Audio2 online test to deploy config (#5309)

Signed-off-by: zwhzzz0821 <2831474076@qq.com>

tests/e2e/online_serving/stage_configs/step_audio2_ci.yaml

@@ -1,81 +0,0 @@
-# Stage config for Step-Audio2 CI tests
-# Stage 0: Thinker (audio understanding → text + audio tokens)
-# Stage 1: Token2Wav (audio tokens → waveform)
-
-# The following config has been verified on 2x 24GB GPU (L4/RTX3090/RTX4090).
-stage_args:
-  - stage_id: 0
-    runtime:
-      process: true
-      devices: "0"
-      max_batch_size: 1
-    engine_args:
-      model_stage: thinker
-      model_arch: StepAudio2ThinkerForConditionalGeneration
-      tensor_parallel_size: 1
-      hf_overrides:
-          architectures: ["StepAudio2ThinkerForConditionalGeneration"]
-
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      max_model_len: 1024
-      max_num_batched_tokens: 1024
-      max_num_seqs: 1
-      gpu_memory_utilization: 0.7
-      skip_mm_profiling: true
-      enforce_eager: true
-      trust_remote_code: true
-      enable_prefix_caching: false
-      engine_output_type: text  # Output text + audio tokens
-    is_comprehension: true
-    final_output: true
-    final_output_type: text
-    default_sampling_params:
-      temperature: 0.7
-      top_p: 0.9
-      top_k: -1
-      max_tokens: 256
-      seed: 42
-      detokenize: true
-      repetition_penalty: 1.05
-
-  - stage_id: 1
-    runtime:
-      process: true
-      devices: "0"
-      max_batch_size: 1
-    engine_args:
-
-      model_stage: token2wav
-      model_arch: StepAudio2Token2WavModel
-      worker_cls: vllm_omni.worker.gpu_generation_worker.GPUGenerationWorker
-      scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler
-      hf_overrides:
-          architectures: ["StepAudio2Token2WavModel"]
-      gpu_memory_utilization: 0.2
-      enforce_eager: true
-      trust_remote_code: true
-      enable_prefix_caching: false
-      max_num_batched_tokens: 1024
-      engine_output_type: audio  # Output waveform
-    engine_input_source: [0]
-    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.step_audio2.thinker2token2wav
-    final_output: true
-    final_output_type: audio
-    default_sampling_params:
-      temperature: 0.0
-      top_p: 1.0
-      top_k: -1
-      max_tokens: 1
-      seed: 42
-      detokenize: false
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 1
-  edges:
-    - from: 0
-      to: 1
-      window_size: -1

tests/e2e/online_serving/test_step_audio2_expansion.py

@@ -3,34 +3,32 @@
 """
 E2E expansion tests for Step-Audio2 online serving (weekly CI).
 
-Tests speech-to-speech translation via /v1/chat/completions.
+Tests audio-to-text generation via /v1/chat/completions.
 Request shape follows examples/online_serving/step_audio2/openai_chat_completion_client.py.
 """
 
 import os
-from pathlib import Path
 
 import pytest
 
 from tests.helpers.mark import hardware_test
 from tests.helpers.media import generate_synthetic_audio
 from tests.helpers.runtime import OmniServerParams
+from tests.helpers.stage_config import get_deploy_config_path
 
 os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
 
 MODEL = "stepfun-ai/Step-Audio-2-mini"
-STAGE_CONFIG = str(Path(__file__).parent / "stage_configs" / "step_audio2_ci.yaml")
+DEPLOY_CONFIG = get_deploy_config_path("step_audio2_ci.yaml")
 SAMPLE_RATE = 16000
 SEED = 42
-# pyttsx3 synthetic input phrase; Step-Audio2 text output includes codec tokens
-# (<audio_NNNN>) so we validate audio only, not text↔audio similarity.
 SYNTH_PHRASE = "how are you"
 
 TEST_PARAMS = [
     pytest.param(
         OmniServerParams(
             model=MODEL,
-            stage_config_path=STAGE_CONFIG,
+            server_args=["--deploy-config", DEPLOY_CONFIG, "--trust-remote-code"],
             env_dict={"VLLM_IMAGE_FETCH_TIMEOUT": "60"},
         ),
         id="step_audio2",
@@ -64,19 +62,11 @@ def _default_sampling_params_list() -> list[dict]:
             "detokenize": True,
             "repetition_penalty": 1.05,
         },
-        {
-            "temperature": 0.0,
-            "top_p": 1.0,
-            "top_k": -1,
-            "max_tokens": 1,
-            "seed": SEED,
-            "detokenize": False,
-        },
     ]
 
 
-def _build_s2st_request_config(omni_server, audio_base64: str) -> dict:
-    """Build audio-to-audio (S2ST) chat completion request config."""
+def _build_audio_to_text_request_config(omni_server, audio_base64: str) -> dict:
+    """Build an audio-to-text chat completion request config."""
     audio_url = f"data:audio/wav;base64,{audio_base64}"
     return {
         "model": omni_server.model,
@@ -86,7 +76,7 @@ def _build_s2st_request_config(omni_server, audio_base64: str) -> dict:
                 "content": [
                     {
                         "type": "text",
-                        "text": "You are an audio processing assistant. Listen and repeat the audio content.",
+                        "text": "You are a speech recognition assistant. Transcribe the audio accurately.",
                     }
                 ],
             },
@@ -94,26 +84,19 @@ def _build_s2st_request_config(omni_server, audio_base64: str) -> dict:
                 "role": "user",
                 "content": [
                     {"type": "audio_url", "audio_url": {"url": audio_url}},
-                    {"type": "text", "text": "Please listen to this audio and repeat its content."},
+                    {"type": "text", "text": "Please transcribe the audio content."},
                 ],
             },
-            {"role": "assistant", "content": "<tts_start>"},
         ],
-        # Thinker text includes undetokenized audio codec placeholders; assert audio only.
-        "modalities": ["audio"],
-        "audio_ref_text": SYNTH_PHRASE,
+        "modalities": ["text"],
         "sampling_params_list": _default_sampling_params_list(),
-        "extra_body": {
-            "continue_final_message": True,
-            "add_generation_prompt": False,
-        },
     }
 
 
 @hardware_test(res={"cuda": "H100"}, num_cards=1)
-def test_single_s2st_request(omni_server, openai_client) -> None:
-    """Test single speech-to-speech request via chat completions API."""
-    request_config = _build_s2st_request_config(
+def test_single_audio_to_text_request(omni_server, openai_client) -> None:
+    """Test a single audio-to-text request via the chat completions API."""
+    request_config = _build_audio_to_text_request_config(
         omni_server,
         _synthetic_audio_base64(duration_sec=2),
     )
@@ -122,9 +105,9 @@ def test_single_s2st_request(omni_server, openai_client) -> None:
 
 @hardware_test(res={"cuda": "H100"}, num_cards=1)
 @pytest.mark.parametrize("request_num", [2, 4])
-def test_concurrent_s2st_requests(omni_server, openai_client, request_num: int) -> None:
-    """Test concurrent speech-to-speech requests."""
-    request_config = _build_s2st_request_config(
+def test_concurrent_audio_to_text_requests(omni_server, openai_client, request_num: int) -> None:
+    """Test concurrent audio-to-text requests."""
+    request_config = _build_audio_to_text_request_config(
         omni_server,
         _synthetic_audio_base64(duration_sec=2),
     )

tests/test_config_factory.py

@@ -1174,6 +1174,7 @@ stages:
             ("step_audio_2.yaml", "step_audio_2", 2, "audio", None),
             ("step_audio_2_asr.yaml", "step_audio_2_asr", 1, "text", "step_audio_2_asr"),
             ("step_audio_2_async_chunk.yaml", "step_audio_2", 2, "audio", None),
+            ("step_audio2_ci.yaml", "step_audio_2_asr", 1, "text", "step_audio_2_asr"),
             ("hunyuan_video_15.yaml", "hunyuan_video_15", 1, "video", None),
             ("wan2_2_ti2v.yaml", "wan2_2_ti2v", 1, "video", None),
         ],
@@ -1241,6 +1242,33 @@ stages:
         assert deploy.stages[0].tensor_parallel_size == 2
         assert deploy.stages[1].devices == "1"
 
+    def test_step_audio2_online_ci_deploy_is_single_thinker(self):
+        deploy = load_deploy_config(Path(get_deploy_config_path("step_audio2_ci.yaml")))
+
+        assert deploy.pipeline == "step_audio_2_asr"
+        assert deploy.async_chunk is False
+        assert deploy.trust_remote_code is True
+        assert deploy.enable_prefix_caching is False
+        assert len(deploy.stages) == 1
+
+        stage = deploy.stages[0]
+        assert stage.devices == "0"
+        assert stage.tensor_parallel_size == 1
+        assert stage.max_model_len == 1024
+        assert stage.max_num_batched_tokens == 1024
+        assert stage.max_num_seqs == 1
+        assert stage.gpu_memory_utilization == 0.7
+        assert stage.skip_mm_profiling is True
+        assert stage.enforce_eager is True
+        assert stage.async_scheduling is False
+        assert stage.engine_extras["hf_overrides"]["architectures"] == ["StepAudio2ThinkerForConditionalGeneration"]
+
+        stages = merge_pipeline_deploy(OMNI_PIPELINES[deploy.pipeline], deploy)
+        assert len(stages) == 1
+        assert stages[0].final_output is True
+        assert stages[0].final_output_type == "text"
+        assert stages[0].yaml_engine_args["engine_output_type"] == "text"
+
     def test_step_audio2_dispatches_sync_and_async_chunk_processors(self):
         pipeline = resolve_pipeline_config("step_audio_2")
         assert isinstance(pipeline, PipelineConfig)

vllm_omni/deploy/step_audio2_ci.yaml

@@ -0,0 +1,31 @@
+# Step-Audio2 online CI deploy: thinker only (audio/text -> text).
+#
+# The two-stage topology can deadlock when an audio-to-text request skips the
+# Token2Wav stage, so this config explicitly selects the ASR pipeline.
+# Keeps the legacy 1x 24GB CI sizing; validated here on 1x H100-80GB.
+pipeline: step_audio_2_asr
+async_chunk: false
+trust_remote_code: true
+enable_prefix_caching: false
+
+stages:
+  - stage_id: 0
+    devices: "0"
+    tensor_parallel_size: 1
+    max_model_len: 1024
+    max_num_batched_tokens: 1024
+    max_num_seqs: 1
+    gpu_memory_utilization: 0.7
+    skip_mm_profiling: true
+    enforce_eager: true
+    async_scheduling: false
+    hf_overrides:
+      architectures: ["StepAudio2ThinkerForConditionalGeneration"]
+    default_sampling_params:
+      temperature: 0.7
+      top_p: 0.9
+      top_k: -1
+      max_tokens: 256
+      seed: 42
+      detokenize: true
+      repetition_penalty: 1.05