[CI failed]Revert "[Config Refactor] Derive deploy override fields from stage config" (#3287)

Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>

tests/e2e/offline_inference/test_bagel_text2img.py

@@ -67,7 +67,7 @@ if current_omni_platform.is_rocm():
     ]
 
 # Maximum allowed difference per color channel
-PIXEL_TOLERANCE = 5
+PIXEL_TOLERANCE = 10
 
 # Default test prompt
 DEFAULT_PROMPT = "<|im_start|>A cute cat<|im_end|>"

tests/e2e/online_serving/test_mimo_audio.py

@@ -83,6 +83,7 @@ def get_max_batch_size(size_type="few"):
 @pytest.mark.omni
 @hardware_test(res={"cuda": "L4", "rocm": "MI325"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
+@pytest.mark.skip(reason="CI failed 8571")
 def test_audio_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Test audio and text input processing and text/audio output generation via OpenAI API.

tests/entrypoints/test_async_omni_diffusion_config.py

@@ -4,7 +4,6 @@
 import pytest
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 
-from vllm_omni.config.stage_config import deploy_override_field_names
 from vllm_omni.engine.async_omni_engine import AsyncOmniEngine
 from vllm_omni.entrypoints.cli.serve import OmniServeCommand, _create_default_diffusion_stage_cfg
 
@@ -31,15 +30,6 @@ def test_default_stage_config_includes_cache_backend():
     assert engine_args["model_stage"] == "diffusion"
 
 
-def test_default_stage_config_ignores_none_deploy_overrides():
-    """Ensure nullified deploy override defaults do not alter diffusion defaults."""
-    baseline = AsyncOmniEngine._create_default_diffusion_stage_cfg({})[0]
-    nullified_overrides = {name: None for name in deploy_override_field_names()}
-    stage_cfg = AsyncOmniEngine._create_default_diffusion_stage_cfg(nullified_overrides)[0]
-
-    assert stage_cfg == baseline
-
-
 def test_default_cache_config_used_when_missing():
     """Ensure default cache_config is synthesized when only backend is given."""
     stage_cfg = AsyncOmniEngine._create_default_diffusion_stage_cfg(

tests/helpers/stage_config.py

@@ -494,7 +494,6 @@ _CI_OVERLAYS: dict[str, dict[str, Any]] = {
                 "max_num_seqs": 1,
                 "gpu_memory_utilization": 0.9,
                 "enforce_eager": True,
-                "enable_prefix_caching": False,
                 "max_num_batched_tokens": 16384,
                 "max_model_len": 16384,
                 "skip_mm_profiling": True,

tests/test_arg_utils.py

@@ -369,8 +369,8 @@ def _build_full_serve_parser():
 def test_nullify_stage_engine_defaults_resets_inherited_defaults():
     import argparse
 
-    from vllm_omni.config.stage_config import deploy_override_field_names
     from vllm_omni.engine.arg_utils import (
+        deploy_override_field_names,
         nullify_stage_engine_defaults,
     )
 

tests/test_config_factory.py

@@ -93,7 +93,6 @@ class TestStageConfig:
         assert omega_config.engine_args.worker_type == "ar"
         assert omega_config.final_output is True
         assert omega_config.final_output_type == "text"
-        assert "max_num_seqs" not in omega_config.engine_args
         # Legacy field name for backward compatibility
         assert omega_config.engine_input_source == []
 
@@ -147,24 +146,6 @@ class TestStageConfig:
         omega_config = config.to_omegaconf()
         assert omega_config.engine_args.max_num_seqs == 32
 
-    def test_to_omegaconf_omits_none_deploy_overrides_for_engine_args(self):
-        """None deploy overrides must fall through to EngineArgs defaults."""
-        from vllm_omni.config.stage_config import deploy_override_field_names
-
-        config = StageConfig(
-            stage_id=0,
-            model_stage="thinker",
-            runtime_overrides={name: None for name in deploy_override_field_names()},
-        )
-
-        omega_config = config.to_omegaconf()
-        engine_args = dict(omega_config.engine_args)
-
-        assert "devices" not in engine_args
-        assert "max_batch_size" not in engine_args
-        for name in deploy_override_field_names() - {"devices"}:
-            assert name not in engine_args
-
 
 class TestModelPipeline:
     """Tests for ModelPipeline class."""
@@ -821,40 +802,6 @@ class TestPipelineRegistry:
 
 
 class TestDeployConfigLoading:
-    def test_deploy_override_fields_include_deploy_schema_fields(self):
-        from vllm_omni.config.stage_config import deploy_override_field_names
-
-        expected_fields = {
-            "async_chunk",
-            "async_scheduling",
-            "config_format",
-            "data_parallel_size",
-            "devices",
-            "disable_hybrid_kv_cache_manager",
-            "distributed_executor_backend",
-            "dtype",
-            "enable_chunked_prefill",
-            "enable_flashinfer_autotune",
-            "enable_prefix_caching",
-            "enforce_eager",
-            "gpu_memory_utilization",
-            "load_format",
-            "max_model_len",
-            "max_num_batched_tokens",
-            "max_num_seqs",
-            "mm_processor_cache_gb",
-            "pipeline_parallel_size",
-            "profiler_config",
-            "quantization",
-            "skip_mm_profiling",
-            "subtalker_sampling_params",
-            "tensor_parallel_size",
-            "tokenizer_mode",
-            "trust_remote_code",
-        }
-
-        assert expected_fields == deploy_override_field_names()
-
     def test_load_deploy_config(self):
         from pathlib import Path
 
@@ -870,17 +817,6 @@ class TestDeployConfigLoading:
         assert deploy.connectors is not None
         assert deploy.platforms is not None
 
-        voxtral_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "voxtral_tts.yaml"
-        if voxtral_path.exists():
-            voxtral_deploy = load_deploy_config(voxtral_path)
-            assert voxtral_deploy.stages[0].config_format == "mistral"
-            assert voxtral_deploy.stages[0].load_format == "mistral"
-            assert voxtral_deploy.stages[0].tokenizer_mode == "mistral"
-            assert not any(
-                name in voxtral_deploy.stages[0].engine_extras
-                for name in ("config_format", "load_format", "tokenizer_mode")
-            )
-
     def test_merge_pipeline_deploy(self):
         from pathlib import Path
 
@@ -1075,8 +1011,7 @@ class TestBaseConfigInheritance:
         deploy = load_deploy_config(ci_path)
         assert len(deploy.stages) == 3
         # CI overrides
-        assert deploy.stages[0].load_format == "dummy"
-        assert "load_format" not in deploy.stages[0].engine_extras
+        assert deploy.stages[0].engine_extras.get("load_format") == "dummy"
         assert deploy.stages[0].max_num_seqs == 5
         # Inherited from base
         assert deploy.stages[0].gpu_memory_utilization == 0.9
@@ -1281,7 +1216,7 @@ class TestSentinelDefaultPrecedence:
     def test_none_value_skipped_yaml_wins(self):
         stages = self._stages({"max_num_seqs": None})
         assert stages[2].runtime_overrides.get("max_num_seqs") is None
-        assert "max_num_seqs" not in stages[2].yaml_engine_args
+        assert stages[2].yaml_engine_args.get("max_num_seqs") == 1
 
     def test_empty_kwargs_yaml_only(self):
         stages = self._stages({})

vllm_omni/config/stage_config.py

@@ -402,11 +402,11 @@ class StageDeployConfig:
     """
 
     stage_id: int
-    max_num_seqs: int | None = None
-    gpu_memory_utilization: float | None = None
-    tensor_parallel_size: int | None = None
-    enforce_eager: bool | None = None
-    max_num_batched_tokens: int | None = None
+    max_num_seqs: int = 64
+    gpu_memory_utilization: float = 0.9
+    tensor_parallel_size: int = 1
+    enforce_eager: bool = False
+    max_num_batched_tokens: int = 32768
     max_model_len: int | None = None
     async_scheduling: bool | None = None
     devices: str = "0"
@@ -414,14 +414,6 @@ class StageDeployConfig:
     input_connectors: dict[str, str] | None = None
     default_sampling_params: dict[str, Any] | None = None
     subtalker_sampling_params: dict[str, Any] | None = None
-    profiler_config: dict[str, Any] | None = None
-    disable_hybrid_kv_cache_manager: bool | None = None
-    mm_processor_cache_gb: float | None = None
-    skip_mm_profiling: bool | None = None
-    enable_flashinfer_autotune: bool | None = None
-    config_format: str | None = None
-    load_format: str | None = None
-    tokenizer_mode: str | None = None
     engine_extras: dict[str, Any] = field(default_factory=dict)
 
 
@@ -446,14 +438,14 @@ class DeployConfig:
     pipeline: str | None = None
 
     # === Pipeline-wide engine settings (applied uniformly to every stage) ===
-    trust_remote_code: bool | None = None
+    trust_remote_code: bool = True
     distributed_executor_backend: str | None = None
     dtype: str | None = None
     quantization: str | None = None
-    enable_prefix_caching: bool | None = None
+    enable_prefix_caching: bool = False
     enable_chunked_prefill: bool | None = None
-    data_parallel_size: int | None = None
-    pipeline_parallel_size: int | None = None
+    data_parallel_size: int = 1
+    pipeline_parallel_size: int = 1
 
 
 _STAGE_NON_ENGINE_KEYS = frozenset(
@@ -697,18 +689,6 @@ _PIPELINE_WIDE_ENGINE_FIELDS: tuple[str, ...] = (
 )
 
 
-def deploy_override_field_names() -> frozenset[str]:
-    """Return deploy-schema fields whose CLI defaults must not override YAML."""
-    return (
-        frozenset(_STAGE_DEPLOY_FIELDS)
-        | frozenset(_PIPELINE_WIDE_ENGINE_FIELDS)
-        | {
-            "async_chunk",
-            "devices",
-        }
-    )
-
-
 def _build_engine_args(
     ps: StagePipelineConfig,
     ds: StageDeployConfig | None,
@@ -881,15 +861,13 @@ class StageConfig:
 
         # CLI overrides take precedence over YAML defaults
         for key, value in self.runtime_overrides.items():
-            if value is None:
-                continue
             if key not in ("devices", "max_batch_size"):
                 engine_args[key] = value
 
         # Build runtime config from YAML defaults + CLI overrides
         runtime: dict[str, Any] = dict(self.yaml_runtime)
         runtime.setdefault("process", True)
-        if self.runtime_overrides.get("devices") is not None:
+        if "devices" in self.runtime_overrides:
             runtime["devices"] = self.runtime_overrides["devices"]
 
         # Legacy compat: migrate runtime.max_batch_size → engine_args.max_num_seqs
@@ -905,6 +883,8 @@ class StageConfig:
             effective_mbs = int(cli_mbs or legacy_mbs or 1)
             engine_args.setdefault("max_num_seqs", effective_mbs)
 
+        engine_args.setdefault("max_num_seqs", 1)
+
         # Build full config dict
         config_dict: dict[str, Any] = {
             "stage_id": self.stage_id,

vllm_omni/deploy/cosyvoice3.yaml

@@ -40,7 +40,7 @@ stages:
       # near-identity repetition penalty forces vLLM to track
       # output_token_ids for RAS (stop-token logit logsumexp).
       repetition_penalty: 1.0001
-    enable_prefix_caching: false
+    disable_hybrid_kv_cache_manager: true
     mm_processor_cache_gb: 0
     skip_mm_profiling: true
 
@@ -54,5 +54,5 @@ stages:
       from_stage_0: connector_of_shared_memory
     default_sampling_params:
       max_tokens: 2048
-    enable_prefix_caching: false
+    disable_hybrid_kv_cache_manager: true
     skip_mm_profiling: true

vllm_omni/deploy/mimo_audio.yaml

@@ -42,7 +42,6 @@ stages:
     max_num_seqs: 1
     gpu_memory_utilization: 0.2
     enforce_eager: true
-    enable_prefix_caching: false
     async_scheduling: false
     max_num_batched_tokens: 8192
     max_model_len: 8192

vllm_omni/deploy/qwen2_5_omni.yaml

@@ -22,7 +22,6 @@ stages:
     max_num_seqs: 1
     gpu_memory_utilization: 0.8
     enforce_eager: true
-    enable_prefix_caching: false
     mm_processor_cache_gb: 0
     devices: "0"
     default_sampling_params:
@@ -50,7 +49,6 @@ stages:
     max_num_seqs: 1
     gpu_memory_utilization: 0.15
     enforce_eager: true
-    enable_prefix_caching: false
     enable_flashinfer_autotune: false
     async_scheduling: false
     devices: "0"

vllm_omni/deploy/qwen3_omni_moe.yaml

@@ -23,7 +23,6 @@ connectors:
 stages:
   - stage_id: 0
     gpu_memory_utilization: 0.9
-    enable_prefix_caching: false
     devices: "0"
     default_sampling_params:
       temperature: 0.4
@@ -48,7 +47,6 @@ stages:
   - stage_id: 2
     gpu_memory_utilization: 0.1
     enforce_eager: true
-    enable_prefix_caching: false
     async_scheduling: false
     max_num_batched_tokens: 51200
     devices: "1"

vllm_omni/deploy/qwen3_tts.yaml

@@ -53,7 +53,6 @@ stages:
     max_num_seqs: 1
     gpu_memory_utilization: 0.3
     enforce_eager: true
-    enable_prefix_caching: false
     async_scheduling: true
     # Must be divisible by num_code_groups and cover (left_context + chunk).
     # Prefill length is Q * num_frames (e.g. 16 * 2148 = 34368); keep

vllm_omni/deploy/voxcpm2.yaml

@@ -20,7 +20,6 @@ stages:
     max_num_batched_tokens: 4096
     max_model_len: 4096
     devices: "0"
-    trust_remote_code: true
     default_sampling_params:
       temperature: 0.0
       top_p: 1.0

vllm_omni/deploy/voxtral_tts.yaml

@@ -24,7 +24,6 @@ stages:
     max_num_seqs: 32
     gpu_memory_utilization: 0.8
     enforce_eager: false
-    enable_prefix_caching: false
     async_scheduling: true
     max_model_len: 4096
     devices: "0"
@@ -49,7 +48,6 @@ stages:
     max_num_seqs: 32
     gpu_memory_utilization: 0.1
     enforce_eager: true
-    enable_prefix_caching: false
     async_scheduling: false
     max_num_batched_tokens: 65536
     max_model_len: 65536

vllm_omni/engine/arg_utils.py

@@ -456,12 +456,50 @@ SHARED_FIELDS: frozenset[str] = frozenset(
     }
 )
 
+_DEPLOY_ENGINE_ARG_OVERRIDE_FIELDS: frozenset[str] = frozenset(
+    {
+        # Capacity / scheduling.
+        "async_scheduling",
+        "max_model_len",
+        "max_num_batched_tokens",
+        "max_num_seqs",
+        # Memory / parallelism.
+        "data_parallel_size",
+        "gpu_memory_utilization",
+        "pipeline_parallel_size",
+        "tensor_parallel_size",
+        # Execution / loading.
+        "enforce_eager",
+        "distributed_executor_backend",
+        "dtype",
+        "quantization",
+        "trust_remote_code",
+        # Caching / chunking.
+        "async_chunk",
+        "enable_prefix_caching",
+        "enable_chunked_prefill",
+        # Model-specific engine extras.
+        "subtalker_sampling_params",
+    }
+)
+
+_DEPLOY_RUNTIME_OVERRIDE_FIELDS: frozenset[str] = frozenset(
+    {
+        "devices",
+    }
+)
+
 
 def orchestrator_field_names() -> frozenset[str]:
     """Return the names of every field on OrchestratorArgs."""
     return frozenset(f.name for f in fields(OrchestratorArgs))
 
 
+def deploy_override_field_names() -> frozenset[str]:
+    """Return kwargs whose parser defaults must not override deploy YAML."""
+    return _DEPLOY_ENGINE_ARG_OVERRIDE_FIELDS | _DEPLOY_RUNTIME_OVERRIDE_FIELDS
+
+
 def internal_blacklist_keys() -> frozenset[str]:
     """Return the set of CLI keys that must never be forwarded as per-stage
     engine overrides.
@@ -615,8 +653,6 @@ def nullify_stage_engine_defaults(parser: argparse.ArgumentParser) -> None:
     """Reset stage-level engine flag defaults to ``None``; preserve real
     default in help text. Only deploy-YAML override fields are touched.
     Idempotent."""
-    from vllm_omni.config.stage_config import deploy_override_field_names
-
     override_dests = deploy_override_field_names()
 
     for action in parser._actions:

vllm_omni/engine/async_omni_engine.py

@@ -1324,9 +1324,6 @@ class AsyncOmniEngine:
 
         num_devices = max(1, int(parallel_config.world_size))
         devices = ",".join(str(i) for i in range(num_devices))
-        enforce_eager = kwargs.get("enforce_eager")
-        if enforce_eager is None:
-            enforce_eager = False
 
         stage_engine_args = {
             "max_num_seqs": 1,
@@ -1340,7 +1337,7 @@ class AsyncOmniEngine:
             "enable_cache_dit_summary": kwargs.get("enable_cache_dit_summary", False),
             "enable_cpu_offload": kwargs.get("enable_cpu_offload", False),
             "enable_layerwise_offload": kwargs.get("enable_layerwise_offload", False),
-            "enforce_eager": enforce_eager,
+            "enforce_eager": False if kwargs.get("enforce_eager") is None else kwargs.get("enforce_eager"),
             "boundary_ratio": kwargs.get("boundary_ratio", None),
             "flow_shift": kwargs.get("flow_shift", None),
             "diffusion_load_format": kwargs.get("diffusion_load_format", "default"),

vllm_omni/entrypoints/omni_base.py

@@ -100,7 +100,9 @@ class OmniBase(PDDisaggregationMixin):
         kwargs: dict[str, Any] = {k: v for k, v in vars(args).items() if not k.startswith("_")}
 
         if parser is not None and not getattr(parser, "_omni_nullified", False):
-            from vllm_omni.config.stage_config import deploy_override_field_names
+            from vllm_omni.engine.arg_utils import (
+                deploy_override_field_names,
+            )
             from vllm_omni.entrypoints.utils import detect_explicit_cli_keys
 
             explicit = detect_explicit_cli_keys(sys.argv[1:], parser) or set()