[Bugfix][Diffusion] Honor HF_HUB_OFFLINE in OmniDiffusionConfig (align with AR stage) (#5403)

Signed-off-by: Weiming Liao <liaowm5@gmail.com>
Signed-off-by: Weiming Liao <fayespica@outlook.com>
Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com>

tests/diffusion/test_omni_diffusion_config_offline_model_path.py

@@ -0,0 +1,60 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Unit tests for OmniDiffusionConfig HF-offline model-path resolution."""
+
+import os
+
+import huggingface_hub
+import pytest
+
+import vllm_omni.diffusion.data as diffusion_data
+from vllm_omni.diffusion.data import OmniDiffusionConfig
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+_LOCAL_SNAPSHOT = "/root/.cache/huggingface/hub/models--foo--bar/snapshots/deadbeef"
+
+
+@pytest.fixture
+def record_get_model_path(monkeypatch: pytest.MonkeyPatch):
+    calls: list[tuple[str, str | None]] = []
+
+    def fake_get_model_path(model, revision=None):
+        calls.append((model, revision))
+        if os.path.exists(model):
+            return model
+        return _LOCAL_SNAPSHOT
+
+    monkeypatch.setattr(diffusion_data, "get_model_path", fake_get_model_path)
+    return calls
+
+
+class TestOmniDiffusionConfigOfflineModelPath:
+    def test_offline_repo_id_replaced_with_local_path(
+        self, monkeypatch: pytest.MonkeyPatch, record_get_model_path
+    ) -> None:
+        monkeypatch.setattr(huggingface_hub.constants, "HF_HUB_OFFLINE", True)
+        config = OmniDiffusionConfig(model="foo/bar")
+        assert config.model == _LOCAL_SNAPSHOT
+        assert record_get_model_path == [("foo/bar", None)]
+
+    def test_online_leaves_model_untouched(self, monkeypatch: pytest.MonkeyPatch, record_get_model_path) -> None:
+        monkeypatch.setattr(huggingface_hub.constants, "HF_HUB_OFFLINE", False)
+        config = OmniDiffusionConfig(model="foo/bar")
+        assert config.model == "foo/bar"
+        assert record_get_model_path == []
+
+    def test_offline_local_path_unchanged(self, monkeypatch: pytest.MonkeyPatch, tmp_path) -> None:
+        monkeypatch.setattr(huggingface_hub.constants, "HF_HUB_OFFLINE", True)
+        local = str(tmp_path)
+        config = OmniDiffusionConfig(model=local)
+        assert config.model == local
+
+    @pytest.mark.parametrize("model", [None, ""])
+    def test_offline_empty_model_unchanged(
+        self, monkeypatch: pytest.MonkeyPatch, record_get_model_path, model: str | None
+    ) -> None:
+        monkeypatch.setattr(huggingface_hub.constants, "HF_HUB_OFFLINE", True)
+        config = OmniDiffusionConfig(model=model)
+        assert config.model == model
+        assert record_get_model_path == []

vllm_omni/diffusion/data.py

@@ -11,6 +11,7 @@ from enum import Enum
 from typing import TYPE_CHECKING, Any
 
 import diffusers
+import huggingface_hub
 import torch
 from PIL import Image
 from pydantic import Field, model_validator
@@ -20,6 +21,7 @@ from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig,
 )
+from vllm.transformers_utils.repo_utils import get_model_path
 
 from vllm_omni.diffusion.model_metadata import get_diffusion_model_metadata
 from vllm_omni.diffusion.utils.network_utils import is_port_available
@@ -1018,6 +1020,18 @@ class OmniDiffusionConfig:
                 "valid together with diffusion_load_format=diffusers"
             )
 
+        # when use hf offline, replace model to local model path
+        # align with ar stage behavior, see vllm/engine/arg_utils.py
+        if huggingface_hub.constants.HF_HUB_OFFLINE and self.model:
+            model_id = self.model
+            self.model = get_model_path(self.model, self.revision)
+            if model_id != self.model:
+                logger.info(
+                    "HF_HUB_OFFLINE is True, replace model_id [%s] to model_path [%s]",
+                    model_id,
+                    self.model,
+                )
+
     def _propagate_quantization_from_tf_config(self, tf_config: "TransformerConfig") -> None:
         if tf_config.quant_config is None:
             return