[Refactor] Remove legacy stage_args YAML loader (#6200)

Signed-off-by: zwhzzz0821 <2831474076@qq.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

tests/config/test_config_factory.py

@@ -1102,6 +1102,13 @@ class TestResolveScheduler:
 
 
 class TestDeployConfigLoading:
+    def test_rejects_legacy_stage_args_schema(self, tmp_path):
+        deploy_path = tmp_path / "legacy.yaml"
+        deploy_path.write_text("stage_args:\n  - stage_id: 0\n", encoding="utf-8")
+
+        with pytest.raises(ValueError, match=r"stage_args.*PipelineConfig.*stages"):
+            load_deploy_config(deploy_path)
+
     def test_load_minicpmo_duplex_deploy_config(self):
         deploy_path = Path(get_deploy_config_path("minicpmo_4_5_duplex.yaml"))
 
@@ -1377,7 +1384,8 @@ stages:
         assert async_stages[1].custom_process_input_func is None
 
     def test_no_bundled_legacy_stage_config_yamls(self):
-        stage_config_dir = Path(__file__).parent.parent / "vllm_omni" / "model_executor" / "stage_configs"
+        repo_root = Path(__file__).resolve().parents[2]
+        stage_config_dir = repo_root / "vllm_omni" / "model_executor" / "stage_configs"
         assert not list(stage_config_dir.glob("*.yaml"))
 
     def test_merge_pipeline_deploy(self):

tests/e2e/offline_inference/test_hunyuanimage3_i2t_expansion.py

@@ -3,19 +3,18 @@
 """Smoke test for HunyuanImage-3.0 Image-to-Text (I2T) pipeline."""
 
 from collections.abc import Generator
-from pathlib import Path
 
 import pytest
 import torch
 from PIL import Image
 
 from tests.helpers.runtime import OmniRunner
+from tests.helpers.stage_config import get_deploy_config_path
 from vllm_omni import Omni
 from vllm_omni.diffusion.models.hunyuan_image3.prompt_utils import build_prompt
 
 MODEL_NAME = "tencent/HunyuanImage-3.0-Instruct"
-REPO_ROOT = Path(__file__).resolve().parents[3]
-STAGE_CONFIG_PATH = REPO_ROOT / "vllm_omni" / "model_executor" / "stage_configs" / "hunyuan_image3_i2t.yaml"
+DEPLOY_CONFIG_PATH = get_deploy_config_path("hunyuan_image3_ar.yaml")
 
 # First 20 generated token IDs from the HF greedy reference on this input.
 # vllm-omni AR output matches this prefix bitwise; the two implementations
@@ -45,14 +44,14 @@ EXPECTED_PREFIX_TOKEN_IDS: list[int] = [
 # Decoded form, kept only for human-readable assertion messages.
 EXPECTED_PREFIX_TEXT = "The image is a solid, uniform green color with no variations, objects, or details present. It"
 
-pytestmark = [pytest.mark.full_model, pytest.mark.diffusion]
+pytestmark = [pytest.mark.full_model, pytest.mark.diffusion, pytest.mark.cuda]
 
 
 @pytest.fixture(scope="module")
 def omni() -> Generator[Omni, None, None]:
     with OmniRunner(
         MODEL_NAME,
-        deploy_config=str(STAGE_CONFIG_PATH),
+        deploy_config=DEPLOY_CONFIG_PATH,
         trust_remote_code=True,
     ) as runner:
         yield runner.omni

tests/entrypoints/test_async_omni_diffusion_config.py

@@ -492,6 +492,24 @@ def test_resolve_stage_configs_injects_additional_config_into_diffusion_stage(mo
     assert stage_configs[1].engine_args.additional_config == {"torchair_graph_config": {"enabled": True}}
 
 
+@pytest.mark.parametrize(
+    ("legacy_arg", "value"),
+    [
+        ("stage_configs_path", "legacy.yaml"),
+        ("stage_configs", [{"stage_id": 0}]),
+    ],
+)
+def test_resolve_stage_configs_rejects_legacy_config_arguments(legacy_arg, value):
+    engine = AsyncOmniEngine.__new__(AsyncOmniEngine)
+
+    with pytest.raises(ValueError, match=rf"`{legacy_arg}`.*`deploy_config`"):
+        engine._resolve_stage_configs(
+            "dummy-model",
+            {legacy_arg: value},
+            trust_remote_code=False,
+        )
+
+
 def test_default_stage_config_includes_quantization_config():
     """Ensure structured quantization_config survives default diffusion-stage creation."""
     quantization_config = {

tests/entrypoints/test_utils.py

@@ -27,7 +27,7 @@ from vllm_omni.entrypoints.utils import (
     filter_dataclass_kwargs,
     filter_stages,
     load_and_resolve_stage_configs,
-    load_stage_configs_from_yaml,
+    load_stage_configs_from_model,
     resolve_model_config_path,
 )
 
@@ -520,71 +520,20 @@ class TestLoadAndResolveStageConfigs:
         assert filtered[0].final_output_type is None
 
 
-class TestLoadStageConfigsFromYaml:
-    """Regression tests for stage-config loading and merging."""
-
-    def test_deep_merges_stage_engine_args(self, mocker: MockerFixture):
-        yaml_config = create_config(
-            {
-                "async_chunk": True,
-                "stage_args": [
-                    {
-                        "stage_id": 0,
-                        "runtime": {"device": 0},
-                        "engine_args": {
-                            "parallel_config": {"tensor_parallel_size": 4},
-                        },
-                    }
-                ],
-            }
-        )
+class TestLoadStageConfigsFromModel:
+    def test_unresolved_model_does_not_fall_back_to_yaml(self, mocker: MockerFixture, caplog):
         mocker.patch(
-            "vllm_omni.entrypoints.utils.load_yaml_config",
-            return_value=yaml_config,
+            "vllm_omni.entrypoints.utils.StageConfigFactory.create_legacy_stage_configs_from_model",
+            return_value=(None, None),
         )
+        resolve_path = mocker.patch("vllm_omni.entrypoints.utils.resolve_model_config_path")
 
-        stages = load_stage_configs_from_yaml(
-            "fake.yaml",
-            base_engine_args={
-                "parallel_config": {
-                    "tensor_parallel_size": 1,
-                    "pipeline_parallel_size": 2,
-                },
-                "model": "base-model",
-            },
-        )
+        result = load_stage_configs_from_model("unregistered-model", trust_remote_code=False)
 
-        merged_engine_args = stages[0]["engine_args"]
-        assert merged_engine_args["parallel_config"]["tensor_parallel_size"] == 4
-        assert merged_engine_args["parallel_config"]["pipeline_parallel_size"] == 2
-        assert merged_engine_args["model"] == "base-model"
-        assert merged_engine_args["async_chunk"] is True
-
-    def test_merges_nested_stage_engine_args(self, mocker: MockerFixture):
-        yaml_config = create_config(
-            {
-                "stage_args": [
-                    {
-                        "stage_id": 0,
-                        "engine_args": {
-                            "nested": {"override": 2},
-                        },
-                    }
-                ],
-            }
-        )
-        mocker.patch(
-            "vllm_omni.entrypoints.utils.load_yaml_config",
-            return_value=yaml_config,
-        )
-
-        stages = load_stage_configs_from_yaml(
-            "fake.yaml",
-            base_engine_args={"nested": {"base": 1}},
-        )
-
-        assert stages[0]["engine_args"]["nested"]["base"] == 1
-        assert stages[0]["engine_args"]["nested"]["override"] == 2
+        assert result == ([], None)
+        resolve_path.assert_not_called()
+        assert "No registered PipelineConfig resolved" in caplog.text
+        assert "deploy_config" in caplog.text
 
 
 class TestCumulativeStreamingCoercion: