[Bugfix] Preserve object storage URIs during model resolution (#5036)

Signed-off-by: Ziyang Zhang <hafuhafu@qq.com>

tests/entrypoints/test_omni_snapshot_download.py

@@ -10,8 +10,10 @@ ModelScope path.
 
 import sys
 import types
+from pathlib import Path
 
 import pytest
+from pytest_mock import MockerFixture
 from vllm import envs
 
 from vllm_omni.entrypoints import omni_base
@@ -21,6 +23,64 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 MODEL_ID = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
 
 
+@pytest.mark.parametrize(
+    "model_uri",
+    [
+        "s3://bucket/model",
+        "gs://bucket/model",
+        "az://bucket/model",
+    ],
+)
+def test_omni_snapshot_download_preserves_object_storage_uri(
+    model_uri: str,
+    mocker: MockerFixture,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("VLLM_USE_MODELSCOPE", raising=False)
+    hf_download = mocker.patch.object(omni_base, "download_weights_from_hf_specific")
+
+    assert omni_base.omni_snapshot_download(model_uri) == model_uri
+    hf_download.assert_not_called()
+
+
+def test_omni_snapshot_download_preserves_existing_local_path(
+    tmp_path: Path,
+    mocker: MockerFixture,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("VLLM_USE_MODELSCOPE", raising=False)
+    absolute_model_path = tmp_path / "absolute-model"
+    absolute_model_path.mkdir()
+    relative_model_path = "relative-model"
+    (tmp_path / relative_model_path).mkdir()
+    monkeypatch.chdir(tmp_path)
+    hf_download = mocker.patch.object(omni_base, "download_weights_from_hf_specific")
+
+    assert omni_base.omni_snapshot_download(str(absolute_model_path)) == str(absolute_model_path)
+    assert omni_base.omni_snapshot_download(relative_model_path) == relative_model_path
+    hf_download.assert_not_called()
+
+
+def test_omni_snapshot_download_uses_hf_for_relative_repo_id(
+    mocker: MockerFixture,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.delenv("VLLM_USE_MODELSCOPE", raising=False)
+    model_id = "org/model"
+    # Stub the modular-index probe so the test never reaches huggingface.co.
+    hf_file_probe = mocker.patch.object(omni_base, "file_or_path_exists", return_value=False)
+    hf_download = mocker.patch.object(omni_base, "download_weights_from_hf_specific")
+
+    assert omni_base.omni_snapshot_download(model_id) == model_id
+    hf_file_probe.assert_called_once_with(model_id, "modular_model_index.json", revision=None)
+    hf_download.assert_called_once_with(
+        model_name_or_path=model_id,
+        cache_dir=None,
+        allow_patterns=["*"],
+        require_all=True,
+    )
+
+
 @pytest.fixture
 def download_backend(monkeypatch: pytest.MonkeyPatch):
     """Run ``omni_snapshot_download`` and report which backend it selected.

tests/entrypoints/test_utils.py

@@ -22,6 +22,7 @@ from vllm_omni.engine.stage_init_utils import (
 from vllm_omni.entrypoints.utils import (
     _convert_dataclasses_to_dict,
     _filter_dict_like_object,
+    _try_resolve_omni_model_type,
     coerce_param_message_types,
     filter_dataclass_kwargs,
     filter_stages,
@@ -371,6 +372,48 @@ class TestResolveModelConfigPath:
         assert sender["extra"]["connector_get_max_wait_first_chunk"] == 3000
         assert sender["extra"]["connector_get_max_wait"] == 300
 
+    def test_object_storage_uri_resolves_via_materialized_configs(
+        self,
+        mocker: MockerFixture,
+        tmp_path,
+    ):
+        """An object-storage URI must not reach HF helpers; resolution reads the
+        locally materialized config files instead."""
+        hf_config = SimpleNamespace(
+            model_type="qwen3_omni_moe",
+            architectures=["Qwen3OmniMoeForConditionalGeneration"],
+        )
+        uri = "s3://qwen3-tts-models/Qwen3-Omni-30B-A3B-Instruct"
+        materialize = mocker.patch(
+            "vllm_omni.entrypoints.utils._materialize_object_storage_configs",
+            return_value=str(tmp_path),
+        )
+        get_config = mocker.patch(
+            "vllm_omni.entrypoints.utils.get_config",
+            return_value=hf_config,
+        )
+
+        result = resolve_model_config_path(uri)
+
+        materialize.assert_called_once_with(uri)
+        get_config.assert_called_once()
+        assert get_config.call_args[0][0] == str(tmp_path)
+        assert result is not None
+        assert Path(result).as_posix().endswith("qwen3_omni_moe.yaml")
+
+
+class TestTryResolveOmniModelType:
+    """Name matching must only scan the last path component."""
+
+    def test_uri_bucket_name_cannot_hijack_pipeline(self):
+        # "qwen3-tts-models" normalizes to "qwen3ttsmodels", which contains the
+        # registered key "qwen3tts" — only the basename may participate, so a
+        # misleadingly named bucket selects nothing.
+        assert _try_resolve_omni_model_type("s3://qwen3-tts-models/plain-checkpoint") is None
+
+    def test_basename_still_matches_pipeline(self):
+        assert _try_resolve_omni_model_type("gs://any-bucket/Fun-CosyVoice3-0.5B-2512") == "cosyvoice3"
+
 
 class TestLoadAndResolveStageConfigs:
     def test_load_and_resolve_with_kwargs(self, mocker: MockerFixture):