[Config Refactor] Derive deploy override fields from stage config (#3162)

Signed-off-by: xiaohajiayou <923390377@qq.com>

tests/entrypoints/test_async_omni_diffusion_config.py

@@ -4,6 +4,7 @@
 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
 
@@ -30,6 +31,15 @@ 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

@@ -494,6 +494,7 @@ _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

@@ -93,6 +93,7 @@ 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 == []
 
@@ -146,6 +147,24 @@ 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."""
@@ -802,6 +821,40 @@ 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
 
@@ -817,6 +870,17 @@ 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
 
@@ -1011,7 +1075,8 @@ class TestBaseConfigInheritance:
         deploy = load_deploy_config(ci_path)
         assert len(deploy.stages) == 3
         # CI overrides
-        assert deploy.stages[0].engine_extras.get("load_format") == "dummy"
+        assert deploy.stages[0].load_format == "dummy"
+        assert "load_format" not in deploy.stages[0].engine_extras
         assert deploy.stages[0].max_num_seqs == 5
         # Inherited from base
         assert deploy.stages[0].gpu_memory_utilization == 0.9
@@ -1216,7 +1281,7 @@ 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 stages[2].yaml_engine_args.get("max_num_seqs") == 1
+        assert "max_num_seqs" not in stages[2].yaml_engine_args
 
     def test_empty_kwargs_yaml_only(self):
         stages = self._stages({})