diff --git a/tests/benchmarks/test_accuracy_bench_utils.py b/tests/benchmarks/test_accuracy_bench_utils.py
index ab539234616..8d78da2b2da 100644
--- a/tests/benchmarks/test_accuracy_bench_utils.py
+++ b/tests/benchmarks/test_accuracy_bench_utils.py
@@ -1,9 +1,11 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 # ruff: noqa: E402, I001
 import argparse
 import math
 import os
 import sys
-import types
 from pathlib import Path
 
 import pytest
@@ -106,11 +108,9 @@ def fake_snapshot_download(*, repo_id, repo_type, allow_patterns):
         captured["allow_patterns"] = allow_patterns
         return str(downloaded_root)
 
-    monkeypatch.setitem(
-        sys.modules,
-        "huggingface_hub",
-        types.SimpleNamespace(snapshot_download=fake_snapshot_download),
-    )
+    from vllm_omni.transformers_utils import repo_utils
+
+    monkeypatch.setattr(repo_utils.hf_api(), "snapshot_download", fake_snapshot_download)
 
     resolved = resolve_seed_tts_root(
         "zhaochenyang20/seed-tts-eval",
diff --git a/tests/dfx/perf/scripts/run_diffusion_benchmark.py b/tests/dfx/perf/scripts/run_diffusion_benchmark.py
index bb69391fdf1..ffcfc45b721 100644
--- a/tests/dfx/perf/scripts/run_diffusion_benchmark.py
+++ b/tests/dfx/perf/scripts/run_diffusion_benchmark.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """
 Performance benchmark CI runner for diffusion models.
 
@@ -356,6 +359,8 @@ def _resolve_offline_model(model: str) -> str:
     """
     import huggingface_hub
 
+    from vllm_omni.transformers_utils.repo_utils import hf_api
+
     if not model or os.path.isdir(model):
         return model
 
@@ -380,9 +385,7 @@ def _resolve_offline_model(model: str) -> str:
     if len(parts) >= 3:
         repo_id = "/".join(parts[:2])
         subfolder = "/".join(parts[2:])
-        from huggingface_hub import snapshot_download
-
-        snapshot_root = snapshot_download(
+        snapshot_root = hf_api().snapshot_download(
             repo_id,
             allow_patterns=[f"{subfolder}/**"],
             local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
@@ -391,9 +394,7 @@ def _resolve_offline_model(model: str) -> str:
 
     if not huggingface_hub.constants.HF_HUB_OFFLINE:
         return model
-    from huggingface_hub import snapshot_download
-
-    return snapshot_download(model, local_files_only=True)
+    return hf_api().snapshot_download(model, local_files_only=True)
 
 
 class DiffusionServer:
diff --git a/tests/diffusion/model_loader/test_diffusers_loader.py b/tests/diffusion/model_loader/test_diffusers_loader.py
index 555cb19bb4d..565949ff616 100644
--- a/tests/diffusion/model_loader/test_diffusers_loader.py
+++ b/tests/diffusion/model_loader/test_diffusers_loader.py
@@ -13,7 +13,6 @@
 import pytest
 import torch
 import torch.nn as nn
-from huggingface_hub import snapshot_download
 from safetensors.torch import save_file
 from vllm.config.load import LoadConfig
 
@@ -32,6 +31,7 @@
 from vllm_omni.diffusion.models.host_weight_contract import FinalLayoutModelContract
 from vllm_omni.diffusion.registry import initialize_model
 from vllm_omni.quantization.component_config import ComponentQuantizationConfig
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
 
@@ -41,7 +41,7 @@
 @pytest.fixture(scope="module")
 def prefetch_helios_model():
     """Downloads the tiny helios model prior to running a test."""
-    snapshot_download(model_path)
+    hf_api().snapshot_download(model_path)
 
 
 @pytest.fixture(scope="function")
diff --git a/tests/diffusion/model_loader/test_hub_prefetch.py b/tests/diffusion/model_loader/test_hub_prefetch.py
index 39eb2b27711..94170edcc3f 100644
--- a/tests/diffusion/model_loader/test_hub_prefetch.py
+++ b/tests/diffusion/model_loader/test_hub_prefetch.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 import contextlib
 
@@ -14,10 +14,10 @@
 def test_prefetch_subfolders_propagates_revision(monkeypatch):
     calls = []
 
-    def fake_snapshot_download(**kwargs):
+    def fake_snapshot_download(self, **kwargs):
         calls.append(kwargs)
 
-    monkeypatch.setattr(huggingface_hub, "snapshot_download", fake_snapshot_download)
+    monkeypatch.setattr(huggingface_hub.HfApi, "snapshot_download", fake_snapshot_download)
     monkeypatch.setattr(hub_prefetch, "_repo_prefetch_lock", lambda _model: contextlib.nullcontext())
 
     hub_prefetch.prefetch_subfolders(
diff --git a/tests/diffusion/models/cosmos3/test_cosmos3_sound_tokenizer.py b/tests/diffusion/models/cosmos3/test_cosmos3_sound_tokenizer.py
index 7ab04cc212f..1aa653d4a37 100644
--- a/tests/diffusion/models/cosmos3/test_cosmos3_sound_tokenizer.py
+++ b/tests/diffusion/models/cosmos3/test_cosmos3_sound_tokenizer.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -81,9 +81,8 @@ def test_from_config_loads_local_diffusers_component(tmp_path, monkeypatch: pyte
 
 
 def test_from_config_downloads_component_from_hf_repo(tmp_path, monkeypatch: pytest.MonkeyPatch) -> None:
-    import huggingface_hub
-
     from vllm_omni.diffusion.models.cosmos3 import sound_tokenizer
+    from vllm_omni.transformers_utils import repo_utils
 
     cache_dir = tmp_path / "hf"
     _write_component(cache_dir, checkpoint_name=DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME)
@@ -95,7 +94,7 @@ def fake_snapshot_download(repo_id: str, *, revision: str | None, allow_patterns
         calls.append((repo_id, revision, allow_patterns))
         return str(cache_dir)
 
-    monkeypatch.setattr(huggingface_hub, "snapshot_download", fake_snapshot_download)
+    monkeypatch.setattr(repo_utils.hf_api(), "snapshot_download", fake_snapshot_download)
 
     sound_tokenizer.Cosmos3SoundTokenizer.from_config(
         SimpleNamespace(
diff --git a/tests/diffusion/models/ltx2/test_ltx2_pipeline.py b/tests/diffusion/models/ltx2/test_ltx2_pipeline.py
index 3e630d8ddd6..a7f6b95dfb8 100644
--- a/tests/diffusion/models/ltx2/test_ltx2_pipeline.py
+++ b/tests/diffusion/models/ltx2/test_ltx2_pipeline.py
@@ -10,6 +10,7 @@
 from types import SimpleNamespace
 from typing import Any
 
+import huggingface_hub
 import numpy as np
 import pytest
 import torch
@@ -302,11 +303,11 @@ def test_ltx_artifact_uses_source_revision_and_hub_fallback(
     filename = "ltx-sidecar.safetensors"
     calls = []
 
-    def fake_download(**kwargs):
+    def fake_download(self, **kwargs):
         calls.append(kwargs)
         return "/cache/ltx-sidecar.safetensors"
 
-    monkeypatch.setattr(ltx2_components, "hf_hub_download", fake_download)
+    monkeypatch.setattr(huggingface_hub.HfApi, "hf_hub_download", fake_download)
 
     assert (
         resolve_ltx_artifact(
@@ -331,7 +332,9 @@ def test_ltx_artifact_prefers_model_root(tmp_path, monkeypatch):
     filename = "ltx-sidecar.safetensors"
     expected = tmp_path / filename
     expected.write_bytes(b"sidecar")
-    monkeypatch.setattr(ltx2_components, "hf_hub_download", lambda **_kwargs: pytest.fail("unexpected Hub lookup"))
+    monkeypatch.setattr(
+        huggingface_hub.HfApi, "hf_hub_download", lambda *_args, **_kwargs: pytest.fail("unexpected Hub lookup")
+    )
 
     assert resolve_ltx_artifact(
         str(tmp_path),
@@ -345,11 +348,11 @@ def test_ltx_artifact_prefers_model_root(tmp_path, monkeypatch):
 def test_ltx_artifact_local_model_missing_sidecar_falls_back_to_hub(tmp_path, monkeypatch):
     calls = []
 
-    def fake_download(**kwargs):
+    def fake_download(self, **kwargs):
         calls.append(kwargs)
         return "/cache/ltx-sidecar.safetensors"
 
-    monkeypatch.setattr(ltx2_components, "hf_hub_download", fake_download)
+    monkeypatch.setattr(huggingface_hub.HfApi, "hf_hub_download", fake_download)
 
     assert (
         resolve_ltx_artifact(
diff --git a/tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py b/tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py
index 0fb043a1ec5..a45631f7dce 100644
--- a/tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py
+++ b/tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py
@@ -42,10 +42,10 @@
 
 
 def _resolve_fl2va_model_ref() -> str:
-    from huggingface_hub import snapshot_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
 
     repo_root = Path(
-        snapshot_download(
+        hf_api().snapshot_download(
             repo_id=_MINIMAX_H3_REPO,
             revision=_MINIMAX_H3_REVISION,
             allow_patterns=["FL2VA/**"],
@@ -253,13 +253,13 @@ def test_resolve_fl2va_model_ref(tmp_path, monkeypatch):
     fl2va_root.mkdir()
     (fl2va_root / "model_index.json").write_text("{}", encoding="utf-8")
 
-    def fake_snapshot_download(*, repo_id, revision, allow_patterns):
+    def fake_snapshot_download(self, *, repo_id, revision, allow_patterns):
         assert repo_id == _MINIMAX_H3_REPO
         assert revision == _MINIMAX_H3_REVISION
         assert allow_patterns == ["FL2VA/**"]
         return str(tmp_path)
 
-    monkeypatch.setattr("huggingface_hub.snapshot_download", fake_snapshot_download)
+    monkeypatch.setattr("huggingface_hub.HfApi.snapshot_download", fake_snapshot_download)
     assert _resolve_fl2va_model_ref() == str(fl2va_root)
 
 
diff --git a/tests/e2e/accuracy/ltx/test_ltx_official_similarity.py b/tests/e2e/accuracy/ltx/test_ltx_official_similarity.py
index 63ce7392371..c1b083495e9 100644
--- a/tests/e2e/accuracy/ltx/test_ltx_official_similarity.py
+++ b/tests/e2e/accuracy/ltx/test_ltx_official_similarity.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """E2E accuracy guard against a pinned Lightricks LTX pipeline revision.
 
@@ -22,11 +22,11 @@
 import numpy as np
 import pytest
 import torch
-from huggingface_hub import hf_hub_download, snapshot_download
 from torchmetrics.image import PeakSignalNoiseRatio, StructuralSimilarityIndexMeasure
 
 from tests.e2e.accuracy.helpers import reset_artifact_dir
 from tests.helpers.mark import hardware_test
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 OFFICIAL_REPOSITORY = "https://github.com/Lightricks/LTX-2.git"
 OFFICIAL_REVISION = "9377758131b1ffde4b7f766804590a6617bf2ab9"
@@ -326,7 +326,7 @@ def _resolve_model(case: LTXAccuracyCase) -> Path:
     if revision is None and model_id == case.model_id:
         revision = case.model_revision
     return Path(
-        snapshot_download(
+        hf_api().snapshot_download(
             repo_id=model_id,
             revision=revision,
             allow_patterns=[
@@ -359,7 +359,7 @@ def _resolve_gemma_root(case: LTXAccuracyCase, model: Path) -> Path:
         if configured_model and Path(configured_model).is_dir():
             return Path(configured_model)
         return Path(
-            snapshot_download(
+            hf_api().snapshot_download(
                 repo_id=case.gemma_model_id,
                 revision=case.gemma_model_revision,
                 allow_patterns=[
@@ -383,7 +383,7 @@ def _resolve_artifact(artifact: LTXArtifact, model: Path | None = None) -> Path:
         if model_path.is_file():
             return model_path
     return Path(
-        hf_hub_download(
+        hf_api().hf_hub_download(
             repo_id=artifact.repo_id,
             repo_type=None if artifact.repo_type == "model" else artifact.repo_type,
             filename=artifact.filename,
diff --git a/tests/e2e/features/rlhf_test/test_verl_omni_e2e.py b/tests/e2e/features/rlhf_test/test_verl_omni_e2e.py
index dab5feef1f8..8d8062092a1 100644
--- a/tests/e2e/features/rlhf_test/test_verl_omni_e2e.py
+++ b/tests/e2e/features/rlhf_test/test_verl_omni_e2e.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """E2E test that follows the EXACT same flow as
 ``verl-omni/tests/workers/rollout/rollout_vllm/test_vllm_omni_generate.py``
@@ -35,7 +35,6 @@
 import pytest
 import ray
 import torch
-from huggingface_hub import snapshot_download
 from omegaconf import OmegaConf
 from transformers import AutoTokenizer
 
@@ -46,6 +45,7 @@
     vLLMOmniHttpServerLocal,
 )
 from tests.helpers.mark import hardware_test
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 MODEL = "tiny-random/Qwen-Image"
 TOKENIZER_MODEL = "Qwen/Qwen2-1.5B-Instruct"
@@ -55,7 +55,7 @@
 def _resolve_model_path(repo_id: str) -> str:
     if os.path.isdir(repo_id):
         return repo_id
-    return snapshot_download(repo_id=repo_id)
+    return hf_api().snapshot_download(repo_id=repo_id)
 
 
 @lru_cache(maxsize=1)
diff --git a/tests/e2e/offline_inference/test_cosyvoice3_expansion.py b/tests/e2e/offline_inference/test_cosyvoice3_expansion.py
index 7e57433ef5c..dd7c8ec8509 100644
--- a/tests/e2e/offline_inference/test_cosyvoice3_expansion.py
+++ b/tests/e2e/offline_inference/test_cosyvoice3_expansion.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """
 Offline E2E smoke test for CosyVoice3 zero-shot reference inference.
 
@@ -20,7 +20,6 @@
 import numpy as np
 import pytest
 import soundfile as sf
-from huggingface_hub import snapshot_download
 from vllm.sampling_params import SamplingParams
 
 from tests.helpers.mark import hardware_test
@@ -29,6 +28,7 @@
 from tests.helpers.stage_config import get_deploy_config_path
 from vllm_omni.model_executor.models.cosyvoice3.tokenizer import get_qwen_tokenizer
 from vllm_omni.transformers_utils.configs.cosyvoice3 import CosyVoice3Config
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 MODEL = "FunAudioLLM/Fun-CosyVoice3-0.5B-2512"
 MODEL_DIR_ENV = "VLLM_OMNI_COSYVOICE3_MODEL_DIR"
@@ -58,7 +58,7 @@ def _resolve_model_dir() -> Path:
     override = os.environ.get(MODEL_DIR_ENV)
     if override:
         return Path(override).expanduser().resolve()
-    return Path(snapshot_download(MODEL, allow_patterns=["*"]))
+    return Path(hf_api().snapshot_download(MODEL, allow_patterns=["*"]))
 
 
 def _reference_zero_shot_stage0_sampling(*, text: str) -> SamplingParams:
diff --git a/tests/e2e/offline_inference/test_mammoth_moda2_expansion.py b/tests/e2e/offline_inference/test_mammoth_moda2_expansion.py
index f570a6ea717..4f21dd1c485 100644
--- a/tests/e2e/offline_inference/test_mammoth_moda2_expansion.py
+++ b/tests/e2e/offline_inference/test_mammoth_moda2_expansion.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """
 End-to-end test for MammothModa2 text-to-image generation.
 
@@ -19,12 +22,12 @@
 
 import pytest
 import torch
-from huggingface_hub import snapshot_download
 from vllm.sampling_params import SamplingParams
 
 from tests.helpers.mark import hardware_test
 from tests.helpers.runtime import OmniRunner
 from tests.helpers.stage_config import get_deploy_config_path
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 # ---------------------------------------------------------------------------
 # Constants
@@ -64,7 +67,7 @@
 # Helpers
 # ---------------------------------------------------------------------------
 def _load_t2i_gen_config(repo_id: str) -> dict:
-    weights_dir = Path(snapshot_download(repo_id))
+    weights_dir = Path(hf_api().snapshot_download(repo_id))
     cfg_path = weights_dir / "t2i_generation_config.json"
     if not cfg_path.exists():
         pytest.skip(f"t2i_generation_config.json not found at {cfg_path}")
diff --git a/tests/e2e/offline_inference/test_moss_tts_realtime.py b/tests/e2e/offline_inference/test_moss_tts_realtime.py
index 34d36118d99..81d97cd743c 100644
--- a/tests/e2e/offline_inference/test_moss_tts_realtime.py
+++ b/tests/e2e/offline_inference/test_moss_tts_realtime.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """E2E offline inference tests for MOSS-TTS-Realtime (MossTTSRealtime, 1.7B).
 
 Uses the standard omni_runner + pytestmark pattern (one module-scoped engine
@@ -160,12 +160,13 @@ def _build_request(ref_audio_path: str, text: str) -> dict:
     Frees the codec model before returning to avoid competing with the
     running vllm engine for GPU/CPU memory.
     """
-    from huggingface_hub import snapshot_download
     from transformers import AutoModel, AutoTokenizer
 
+    from vllm_omni.transformers_utils.repo_utils import hf_api
+
     # Step 1: locate the realtime processor module in the snapshot.
     try:
-        snap_dir = Path(snapshot_download(repo_id=MODEL))
+        snap_dir = Path(hf_api().snapshot_download(repo_id=MODEL))
     except Exception as exc:
         msg = f"Cannot locate snapshot for {MODEL}: {exc}"
         if os.environ.get("MOSS_TTS_SKIP_ON_NET_FAIL"):
diff --git a/tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py b/tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
index 432124e76ff..66f8be616f6 100644
--- a/tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
+++ b/tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
@@ -12,7 +12,6 @@
 from types import SimpleNamespace
 
 import pytest
-from huggingface_hub import snapshot_download
 
 from tests.helpers.runtime import OmniServerParams, get_model_prefix
 from tests.helpers.stage_config import (
@@ -20,6 +19,7 @@
     get_deploy_duplex_max_sessions,
     modify_stage_config,
 )
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 MODEL = "openbmb/MiniCPM-o-4_5"
 DEPLOY_CONFIG_REL = "minicpmo_4_5.yaml"
@@ -111,7 +111,7 @@ def resolve_ref_audio() -> Path:
     model_prefix = get_model_prefix()
     model_root = Path(model_prefix) / MODEL if model_prefix else Path(MODEL)
     if not model_root.is_dir():
-        model_root = Path(snapshot_download(MODEL, local_files_only=True))
+        model_root = Path(hf_api().snapshot_download(MODEL, local_files_only=True))
     ref_audio = model_root / REF_AUDIO_RELATIVE_PATH
     if not ref_audio.is_file():
         raise FileNotFoundError(f"MiniCPM-o checkpoint ref audio is missing: {ref_audio}")
diff --git a/tests/engine/test_arg_utils.py b/tests/engine/test_arg_utils.py
index fb8231a75c6..0e88a00ef88 100644
--- a/tests/engine/test_arg_utils.py
+++ b/tests/engine/test_arg_utils.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """
 Tests for Omni config utils. For stability, these tests should largely be
 invariant to the specific attributes of vLLM config except in cases where we
@@ -63,7 +66,7 @@ class ConnectorRunner(OmniConnectorModelRunnerMixin):
     class ConnectorWorker:
         model_runner_cls = ConnectorRunner
 
-    worker_module.ConnectorWorker = ConnectorWorker
+    setattr(worker_module, "ConnectorWorker", ConnectorWorker)
     monkeypatch.setitem(sys.modules, worker_module.__name__, worker_module)
 
     monkeypatch.setattr(OmniEngineArgs, "_ensure_omni_models_registered", lambda _self: None)
@@ -95,7 +98,7 @@ def test_stage_without_connector_configuration_accepts_plain_runner(monkeypatch)
     class WorkerWithoutConnector:
         model_runner_cls = object
 
-    worker_module.WorkerWithoutConnector = WorkerWithoutConnector
+    setattr(worker_module, "WorkerWithoutConnector", WorkerWithoutConnector)
     monkeypatch.setitem(sys.modules, worker_module.__name__, worker_module)
 
     args = OmniEngineArgs(
@@ -113,7 +116,7 @@ def test_full_payload_capability_requires_selected_worker_connector(monkeypatch)
     class WorkerWithoutConnector:
         model_runner_cls = object
 
-    worker_module.WorkerWithoutConnector = WorkerWithoutConnector
+    setattr(worker_module, "WorkerWithoutConnector", WorkerWithoutConnector)
     monkeypatch.setitem(sys.modules, worker_module.__name__, worker_module)
 
     with pytest.raises(ValueError, match="does not provide an Omni connector model runner"):
@@ -130,7 +133,7 @@ def test_full_payload_capability_validates_platform_selected_worker(monkeypatch)
     class WorkerWithoutConnector:
         model_runner_cls = object
 
-    worker_module.WorkerWithoutConnector = WorkerWithoutConnector
+    setattr(worker_module, "WorkerWithoutConnector", WorkerWithoutConnector)
     monkeypatch.setitem(sys.modules, worker_module.__name__, worker_module)
     monkeypatch.setattr(
         current_omni_platform,
@@ -293,7 +296,7 @@ def test_remote_tokenizer_subfolder_download_does_not_report_failure(tmp_path, m
     baseline_config = Mock()
     warning = mocker.patch("vllm_omni.engine.arg_utils.logger.warning")
 
-    monkeypatch.setattr("huggingface_hub.snapshot_download", lambda *args, **kwargs: str(tmp_path))
+    monkeypatch.setattr("huggingface_hub.HfApi.snapshot_download", lambda *args, **kwargs: str(tmp_path))
     monkeypatch.setattr(OmniEngineArgs, "_patch_empty_hf_config", lambda *args, **kwargs: None)
     monkeypatch.setattr(EngineArgs, "create_model_config", lambda _self: baseline_config)
     monkeypatch.setattr(
@@ -339,6 +342,7 @@ def test_patch_missing_local_hf_config(tmp_path):
 def test_non_missing_local_hf_config_error_reaches_parent_loader(tmp_path, monkeypatch, config_entry):
     """Non-missing config errors must reach vLLM's normal loader."""
     config_path = tmp_path / "config.json"
+    loader_error: Exception
     if config_entry == "malformed":
         config_path.write_text("{not valid json", encoding="utf-8")
         loader_error = json.JSONDecodeError("invalid config", "{not valid json", 1)
diff --git a/tests/helpers/__init__.py b/tests/helpers/__init__.py
index a4e641a571a..9120e183f8d 100644
--- a/tests/helpers/__init__.py
+++ b/tests/helpers/__init__.py
@@ -29,12 +29,13 @@ def skip_if_gated_repo_inaccessible(
     and we skip cleanly.
     """
     try:
-        from huggingface_hub import hf_hub_download
         from huggingface_hub.errors import GatedRepoError, RepositoryNotFoundError
+
+        from vllm_omni.transformers_utils.repo_utils import hf_api
     except Exception:
         return
     try:
-        hf_hub_download(repo_id=repo_id, filename=filename, revision=revision)
+        hf_api().hf_hub_download(repo_id=repo_id, filename=filename, revision=revision)
     except GatedRepoError as exc:
         import pytest
 
diff --git a/tests/model_executor/models/minicpmo_4_5/test_code2wav_model_dir.py b/tests/model_executor/models/minicpmo_4_5/test_code2wav_model_dir.py
index 5aa78262ae8..8d5aeade657 100644
--- a/tests/model_executor/models/minicpmo_4_5/test_code2wav_model_dir.py
+++ b/tests/model_executor/models/minicpmo_4_5/test_code2wav_model_dir.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Unit tests for MiniCPM-o 4.5 code2wav model-dir resolution (#5442).
 
 In hub/CI deployments ``model_config.model`` is a repo id rather than a local
@@ -27,7 +27,7 @@ def _no_hub(monkeypatch):
     def _fail(*args, **kwargs):  # pragma: no cover - must not be reached
         raise AssertionError("snapshot_download must not be called here")
 
-    monkeypatch.setattr(huggingface_hub, "snapshot_download", _fail)
+    monkeypatch.setattr(huggingface_hub.HfApi, "snapshot_download", _fail)
 
 
 def test_local_directory_is_returned_unchanged(tmp_path, monkeypatch):
@@ -38,13 +38,13 @@ def test_local_directory_is_returned_unchanged(tmp_path, monkeypatch):
 def test_repo_id_resolves_via_snapshot_download(tmp_path, monkeypatch):
     calls = {}
 
-    def _fake_snapshot_download(model_ref, revision=None, allow_patterns=None):
+    def _fake_snapshot_download(self, model_ref, revision=None, allow_patterns=None):
         calls["model_ref"] = model_ref
         calls["revision"] = revision
         calls["allow_patterns"] = allow_patterns
         return str(tmp_path / "snapshot")
 
-    monkeypatch.setattr(huggingface_hub, "snapshot_download", _fake_snapshot_download)
+    monkeypatch.setattr(huggingface_hub.HfApi, "snapshot_download", _fake_snapshot_download)
     resolved = _resolve_model_dir("openbmb/MiniCPM-o-4_5", revision="abc123")
     assert resolved == str(tmp_path / "snapshot")
     assert calls["model_ref"] == "openbmb/MiniCPM-o-4_5"
@@ -56,7 +56,7 @@ def test_snapshot_download_failure_propagates(monkeypatch):
     def _raise(*args, **kwargs):
         raise FileNotFoundError("offline and not cached")
 
-    monkeypatch.setattr(huggingface_hub, "snapshot_download", _raise)
+    monkeypatch.setattr(huggingface_hub.HfApi, "snapshot_download", _raise)
     with pytest.raises(FileNotFoundError):
         _resolve_model_dir("openbmb/MiniCPM-o-4_5")
 
diff --git a/tests/model_executor/models/omnivoice/test_nonverbal_tags_tokenize.py b/tests/model_executor/models/omnivoice/test_nonverbal_tags_tokenize.py
index fbf5ac0b010..95524fdced3 100644
--- a/tests/model_executor/models/omnivoice/test_nonverbal_tags_tokenize.py
+++ b/tests/model_executor/models/omnivoice/test_nonverbal_tags_tokenize.py
@@ -1,12 +1,15 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 from __future__ import annotations
 
 import pytest
-from huggingface_hub import hf_hub_download
 from tokenizers import Tokenizer as HFTokenizer
 
 from vllm_omni.diffusion.models.omnivoice.pipeline_omnivoice import (
     _tokenize_with_nonverbal_tags,
 )
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
@@ -18,7 +21,7 @@ def _ids(tok: HFTokenizer, s: str) -> list[int]:
 class TestNonVerbalTags:
     @classmethod
     def setup_class(cls):
-        tokenizer_path = hf_hub_download(repo_id="k2-fsa/OmniVoice", filename="tokenizer.json")
+        tokenizer_path = hf_api().hf_hub_download(repo_id="k2-fsa/OmniVoice", filename="tokenizer.json")
         cls.tokenizer = HFTokenizer.from_file(tokenizer_path)
 
     def test_plain_text_fallback(self):
diff --git a/tools/pre_commit/check_forbidden_imports.py b/tools/pre_commit/check_forbidden_imports.py
index 22ba693d4f2..c5a96148ea2 100644
--- a/tools/pre_commit/check_forbidden_imports.py
+++ b/tools/pre_commit/check_forbidden_imports.py
@@ -17,7 +17,7 @@
 _HF_NAMES = (
     r"HfApi|HfFileSystem|hf_hub_download|snapshot_download"
     r"|list_repo_files|file_exists|try_to_load_from_cache"
-    r"|list_repo_refs|repo_exists"
+    r"|list_repo_refs|repo_exists|get_safetensors_metadata"
 )
 
 # Non-library trees: examples, tests, and tooling may keep stdlib imports.
@@ -169,56 +169,20 @@ class ForbiddenImport:
             r"^\s*from\s+huggingface_hub\s+import\b[^\n]*\b(?:" + _HF_NAMES + r")\b"
         ),
         tip=(
-            "Use the shared, vLLM-tagged helpers from "
-            "vllm.transformers_utils.repo_utils (e.g. hf_api(), hf_fs(), "
-            "list_repo_files, file_exists) instead of calling "
-            "huggingface_hub directly."
+            "Use 'hf_api()' from 'vllm_omni.transformers_utils.repo_utils' (or "
+            "add a tagged helper there) instead, so requests are tagged with "
+            "vLLM-Omni's library info."
         ),
         allowed_dirs=_NON_LIBRARY_DIRS,
         allowed_files={
-            "vllm_omni/benchmarks/data_modules/daily_omni_dataset.py",
-            "vllm_omni/benchmarks/data_modules/seed_tts_dataset.py",
-            "vllm_omni/benchmarks/data_modules/seed_tts_eval.py",
-            "vllm_omni/diffusion/lora/loader.py",
-            "vllm_omni/diffusion/model_loader/hub_prefetch.py",
-            "vllm_omni/diffusion/models/cosmos3/sound_tokenizer.py",
-            "vllm_omni/diffusion/models/dreamzero/pipeline_dreamzero.py",
-            "vllm_omni/diffusion/models/helios/pipeline_helios.py",
-            "vllm_omni/diffusion/models/longcat_video/pipeline_longcat_video_avatar.py",
-            "vllm_omni/diffusion/models/ltx2/ltx2_components.py",
-            "vllm_omni/diffusion/models/magi_human/pipeline_magi_human.py",
-            "vllm_omni/diffusion/models/omnivoice/pipeline_omnivoice.py",
-            "vllm_omni/diffusion/models/pi0/pipeline_pi0.py",
-            "vllm_omni/diffusion/models/sensenova_u1/pipeline_sensenova_u1.py",
-            "vllm_omni/diffusion/models/soulx_singer/utils.py",
-            "vllm_omni/diffusion/models/utils.py",
-            "vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py",
-            "vllm_omni/engine/arg_utils.py",
-            "vllm_omni/entrypoints/openai/api_server.py",
-            "vllm_omni/entrypoints/openai/serving_speech.py",
-            "vllm_omni/entrypoints/openai/tts_adapters/cosyvoice3.py",
-            "vllm_omni/model_executor/model_loader/weight_utils.py",
-            "vllm_omni/model_executor/models/audex/checkpoint.py",
-            "vllm_omni/model_executor/models/bagel/bagel.py",
-            "vllm_omni/model_executor/models/cosyvoice3/cosyvoice3.py",
-            "vllm_omni/model_executor/models/covo_audio/covo_audio_code2wav.py",
-            "vllm_omni/model_executor/models/dots_tts/dots_tts_talker.py",
-            "vllm_omni/model_executor/models/dynin_omni/dynin_omni_common.py",
-            "vllm_omni/model_executor/models/dynin_omni/dynin_omni_token2audio.py",
-            "vllm_omni/model_executor/models/glm_tts/glm_tts.py",
+            # ``try_to_load_from_cache`` is a local cache lookup that issues no
+            # Hub request, so there is nothing for ``hf_api()`` to tag.
             "vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_code2wav.py",
             "vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_tokenizer.py",
             "vllm_omni/model_executor/models/higgs_audio_v3/higgs_audio_v3_code2wav.py",
-            "vllm_omni/model_executor/models/indextts2/preprocess_utils.py",
-            "vllm_omni/model_executor/models/indextts2/s2mel/modules/bigvgan.py",
-            "vllm_omni/model_executor/models/indextts2/tokenizer_v2_5.py",
-            "vllm_omni/model_executor/models/ming_tts/speaker_extractor.py",
-            "vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_code2wav.py",
-            "vllm_omni/model_executor/models/omnivoice/omnivoice.py",
-            "vllm_omni/model_executor/models/personaplex/personaplex_mimi.py",
-            "vllm_omni/model_executor/models/step_audio2/step_audio2_token2wav.py",
-            "vllm_omni/model_executor/models/voxtral_tts/voxtral_tts.py",
             "vllm_omni/transformers_utils/configs/higgs_audio_v3.py",
+            # Defines the tagged helper itself.
+            "vllm_omni/transformers_utils/repo_utils.py",
         },
     ),
 }
@@ -330,6 +294,7 @@ def matches(rule: str, content: str) -> bool:
         ("from huggingface_hub import HfFileSystem", True),
         ("from huggingface_hub import list_repo_files", True),
         ("from huggingface_hub import try_to_load_from_cache", True),
+        ("from huggingface_hub import get_safetensors_metadata", True),
         ("    from huggingface_hub import snapshot_download", True),
         ("from huggingface_hub import PyTorchModelHubMixin, hf_hub_download", True),
         ("from huggingface_hub import (snapshot_download)", True),
@@ -346,6 +311,7 @@ def matches(rule: str, content: str) -> bool:
         ("from huggingface_hub.constants import HF_HUB_CACHE", False),
         ("from huggingface_hub.utils import EntryNotFoundError", False),
         ("from vllm.transformers_utils.repo_utils import hf_api", False),
+        ("from vllm_omni.transformers_utils.repo_utils import hf_api", False),
         ("from huggingface_hub import (\n    PyTorchModelHubMixin,\n)", False),
         ("# from huggingface_hub import snapshot_download", False),
     ]
diff --git a/vllm_omni/benchmarks/data_modules/daily_omni_dataset.py b/vllm_omni/benchmarks/data_modules/daily_omni_dataset.py
index abb02b1d500..3bf852e8c54 100644
--- a/vllm_omni/benchmarks/data_modules/daily_omni_dataset.py
+++ b/vllm_omni/benchmarks/data_modules/daily_omni_dataset.py
@@ -170,7 +170,7 @@ def ensure_daily_omni_hub_videos_dir(repo_id: str) -> Path:
         raise ValueError("repo_id is required to download Daily-Omni Videos.tar")
 
     try:
-        from huggingface_hub import hf_hub_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
     except ImportError as e:
         raise ImportError(
             "Daily-Omni Hub media download requires huggingface_hub. "
@@ -180,7 +180,7 @@ def ensure_daily_omni_hub_videos_dir(repo_id: str) -> Path:
     tar_path: Path | None = None
     for fname in ("Videos.tar", "videos.tar"):
         try:
-            tar_path = Path(hf_hub_download(repo_id=rid, filename=fname, repo_type="dataset"))
+            tar_path = Path(hf_api().hf_hub_download(repo_id=rid, filename=fname, repo_type="dataset"))
             break
         except Exception:
             continue
@@ -237,14 +237,14 @@ def ensure_daily_omni_hub_root(repo_id: str) -> Path:
         raise ValueError("repo_id is required to download Daily-Omni from Hugging Face")
 
     try:
-        from huggingface_hub import snapshot_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
     except ImportError as e:
         raise ImportError(
             "Install huggingface_hub to load Daily-Omni from the Hub, or pass a local "
             "--daily-omni-qa-json / --daily-omni-video-dir mirror."
         ) from e
 
-    cache = snapshot_download(
+    cache = hf_api().snapshot_download(
         repo_id=rid,
         repo_type="dataset",
         allow_patterns=_DAILY_OMNI_QA_ALLOW_PATTERNS,
diff --git a/vllm_omni/benchmarks/data_modules/seed_tts_dataset.py b/vllm_omni/benchmarks/data_modules/seed_tts_dataset.py
index 23e8b8ade63..996f4f24865 100644
--- a/vllm_omni/benchmarks/data_modules/seed_tts_dataset.py
+++ b/vllm_omni/benchmarks/data_modules/seed_tts_dataset.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """Seed-TTS zero-shot evaluation-style prompts for ``vllm bench serve``.
 
 Loads rows from the `meta.lst` format used in `BytedanceSpeech/seed-tts-eval`_ (or any
@@ -121,7 +124,7 @@ def resolve_seed_tts_root(dataset_path: str | None, *, explicit_root: str | None
 
     repo_id = dataset_path.strip()
     try:
-        from huggingface_hub import snapshot_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
     except ImportError as e:
         raise ImportError(
             "Install huggingface_hub to download Seed-TTS from the Hub, or clone the dataset "
@@ -130,7 +133,7 @@ def resolve_seed_tts_root(dataset_path: str | None, *, explicit_root: str | None
     # Download only the requested locale subtree instead of the whole dataset
     # repo. This avoids large, flaky nightly downloads when we only need e.g.
     # ``en/meta.lst`` + ``en/prompt-wavs/**``.
-    cache = snapshot_download(
+    cache = hf_api().snapshot_download(
         repo_id=repo_id,
         repo_type="dataset",
         allow_patterns=[f"{locale}/**"],
diff --git a/vllm_omni/benchmarks/data_modules/seed_tts_eval.py b/vllm_omni/benchmarks/data_modules/seed_tts_eval.py
index ddfbbd63058..9dc99548aeb 100644
--- a/vllm_omni/benchmarks/data_modules/seed_tts_eval.py
+++ b/vllm_omni/benchmarks/data_modules/seed_tts_eval.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """Seed-TTS WER aligned with Bytedance ``seed-tts-eval`` / ``run_wer.py``.
 
 Matches the published protocol (see Hugging Face dataset card and
@@ -329,7 +332,8 @@ def _ensure_utmos_jit_model() -> Any | None:
             return _utmos_jit_model
         try:
             import torch
-            from huggingface_hub import hf_hub_download
+
+            from vllm_omni.transformers_utils.repo_utils import hf_api
 
             repo = os.environ.get("SEED_TTS_UTMOS_HF_REPO", "balacoon/utmos").strip() or "balacoon/utmos"
             fname = os.environ.get("SEED_TTS_UTMOS_JIT_FILE", "utmos.jit").strip() or "utmos.jit"
@@ -338,7 +342,7 @@ def _ensure_utmos_jit_model() -> Any | None:
                 repo,
                 fname,
             )
-            path = hf_hub_download(repo_id=repo, filename=fname, repo_type="model")
+            path = hf_api().hf_hub_download(repo_id=repo, filename=fname, repo_type="model")
 
             # TODO The model weights in UTMOS must be loaded in cuda:0; otherwise, the model execution will fail.
             want = "cuda:0"
diff --git a/vllm_omni/diffusion/lora/loader.py b/vllm_omni/diffusion/lora/loader.py
index 3f74b4853a2..a13363c7a77 100644
--- a/vllm_omni/diffusion/lora/loader.py
+++ b/vllm_omni/diffusion/lora/loader.py
@@ -10,11 +10,11 @@
     _convert_non_diffusers_qwen_lora_to_diffusers,
     _convert_non_diffusers_wan_lora_to_diffusers,
 )
-from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file
 from vllm.logger import init_logger
 
 from vllm_omni.diffusion.utils.tf_utils import get_transformer_from_pipeline
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 logger = init_logger(__name__)
 
@@ -131,7 +131,7 @@ def _load_lora_state_dict(
 
     # finally, we try to load it from the internet
     try:
-        model_file = hf_hub_download(
+        model_file = hf_api().hf_hub_download(
             pretrained_model_name_or_path,
             filename=weights_name,
             subfolder=subfolder,
diff --git a/vllm_omni/diffusion/model_loader/hub_prefetch.py b/vllm_omni/diffusion/model_loader/hub_prefetch.py
index 3594aa4e1c0..a1e893b3be6 100644
--- a/vllm_omni/diffusion/model_loader/hub_prefetch.py
+++ b/vllm_omni/diffusion/model_loader/hub_prefetch.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """Best-effort HuggingFace Hub prefetch for multi-subfolder pipelines.
 
@@ -297,7 +297,7 @@ def prefetch_subfolders(
     logger.info("Prefetching %s subfolders: %s", model, subfolders)
 
     try:
-        from huggingface_hub import snapshot_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
     except ImportError:  # pragma: no cover - huggingface_hub is a hard dep
         logger.debug("huggingface_hub unavailable; skipping prefetch of %s", model)
         return
@@ -334,7 +334,7 @@ def prefetch_subfolders(
     for attempt in range(1, _PREFETCH_MAX_ATTEMPTS + 1):
         try:
             with _repo_prefetch_lock(model):
-                snapshot_download(
+                hf_api().snapshot_download(
                     repo_id=model,
                     revision=revision,
                     allow_patterns=allow_patterns,
diff --git a/vllm_omni/diffusion/models/cosmos3/sound_tokenizer.py b/vllm_omni/diffusion/models/cosmos3/sound_tokenizer.py
index 5d6ff8dd968..8c1337ec90d 100644
--- a/vllm_omni/diffusion/models/cosmos3/sound_tokenizer.py
+++ b/vllm_omni/diffusion/models/cosmos3/sound_tokenizer.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Cosmos3 sound tokenizer integration.
 
 The tokenizer decodes model-generated sound latents for video+audio output. It
@@ -338,9 +338,9 @@ def from_config(cls, od_config: OmniDiffusionConfig) -> Cosmos3SoundTokenizer:
 
         model_root = str(model_path) if model_path and os.path.isdir(model_path) else None
         if model_root is None and model_path and not explicit_avae_path:
-            from huggingface_hub import snapshot_download
+            from vllm_omni.transformers_utils.repo_utils import hf_api
 
-            model_root = snapshot_download(
+            model_root = hf_api().snapshot_download(
                 repo_id=str(model_path),
                 revision=getattr(od_config, "revision", None),
                 allow_patterns=[
diff --git a/vllm_omni/diffusion/models/dreamzero/pipeline_dreamzero.py b/vllm_omni/diffusion/models/dreamzero/pipeline_dreamzero.py
index bf19f0a0fc0..09bd536cd7f 100644
--- a/vllm_omni/diffusion/models/dreamzero/pipeline_dreamzero.py
+++ b/vllm_omni/diffusion/models/dreamzero/pipeline_dreamzero.py
@@ -21,7 +21,6 @@
 import numpy as np
 import torch
 import torch.nn as nn
-from huggingface_hub import hf_hub_download
 from transformers import AutoTokenizer, UMT5Config, UMT5EncoderModel
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
@@ -67,6 +66,7 @@
     resolve_session_state_config,
 )
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 logger = logging.getLogger(__name__)
 MAX_DREAMZERO_SESSIONS = 64
@@ -581,7 +581,7 @@ def _load_repo_json(model_path: str, relative_path: str, local_files_only: bool)
                 return json.load(f)
 
         try:
-            json_path = hf_hub_download(model_path, relative_path)
+            json_path = hf_api().hf_hub_download(model_path, relative_path)
             with open(json_path) as f:
                 return json.load(f)
         except Exception:
diff --git a/vllm_omni/diffusion/models/helios/pipeline_helios.py b/vllm_omni/diffusion/models/helios/pipeline_helios.py
index 820f9a1d98a..610fb2b8fbe 100644
--- a/vllm_omni/diffusion/models/helios/pipeline_helios.py
+++ b/vllm_omni/diffusion/models/helios/pipeline_helios.py
@@ -76,9 +76,9 @@ def load_json_config(model_path: str, subfolder: str, filename: str, local_files
                 return json.load(f)
     else:
         try:
-            from huggingface_hub import hf_hub_download
+            from vllm_omni.transformers_utils.repo_utils import hf_api
 
-            config_path = hf_hub_download(
+            config_path = hf_api().hf_hub_download(
                 repo_id=model_path,
                 filename=f"{subfolder}/{filename}",
             )
diff --git a/vllm_omni/diffusion/models/longcat_video/pipeline_longcat_video_avatar.py b/vllm_omni/diffusion/models/longcat_video/pipeline_longcat_video_avatar.py
index 33b0e2a458a..6cb2dd7fb50 100644
--- a/vllm_omni/diffusion/models/longcat_video/pipeline_longcat_video_avatar.py
+++ b/vllm_omni/diffusion/models/longcat_video/pipeline_longcat_video_avatar.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -376,9 +376,9 @@ def _ensure_local_dir(model: str | os.PathLike[str], allow_patterns: list[str] |
     model_path = Path(model)
     if model_path.exists():
         return model_path
-    from huggingface_hub import snapshot_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
 
-    return Path(snapshot_download(str(model), allow_patterns=allow_patterns))
+    return Path(hf_api().snapshot_download(str(model), allow_patterns=allow_patterns))
 
 
 def _ensure_avatar_model_index(model_dir: Path, use_int8: bool) -> None:
diff --git a/vllm_omni/diffusion/models/ltx2/ltx2_components.py b/vllm_omni/diffusion/models/ltx2/ltx2_components.py
index c00fdd8355a..72e63a22f3b 100644
--- a/vllm_omni/diffusion/models/ltx2/ltx2_components.py
+++ b/vllm_omni/diffusion/models/ltx2/ltx2_components.py
@@ -20,7 +20,6 @@
 from diffusers.pipelines.ltx2.latent_upsampler import LTX2LatentUpsamplerModel
 from diffusers.pipelines.ltx2.vocoder import LTX2Vocoder
 from diffusers.video_processor import VideoProcessor
-from huggingface_hub import hf_hub_download
 from safetensors import safe_open
 from safetensors.torch import load_file
 from transformers import AutoModelForImageTextToText, AutoTokenizer, Gemma3ForConditionalGeneration
@@ -32,6 +31,7 @@
 from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
 from vllm_omni.diffusion.model_loader.hub_prefetch import from_pretrained_with_prefetch, prefetch_subfolders
 from vllm_omni.diffusion.offloader.module_collector import ModuleDiscovery
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 if TYPE_CHECKING:
     from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
@@ -279,7 +279,7 @@ def resolve_ltx_artifact(
     # independently pinned artifact revision.
     revision = model_revision if model == repo_id else artifact_revision
     try:
-        return hf_hub_download(
+        return hf_api().hf_hub_download(
             repo_id=repo_id,
             filename=filename,
             revision=revision,
@@ -351,7 +351,7 @@ def _load_ltx_metadata_json(model: str, filename: str, revision: str | None = No
             return {}
     else:
         try:
-            path = hf_hub_download(repo_id=model, filename=filename, revision=revision)
+            path = hf_api().hf_hub_download(repo_id=model, filename=filename, revision=revision)
         except Exception:
             return {}
     try:
@@ -529,7 +529,7 @@ def _detect_vocoder_output_sample_rate(model: str, revision: str | None = None)
     vocoder_config_path = os.path.join(model, "vocoder", "config.json")
     if not os.path.exists(vocoder_config_path):
         try:
-            vocoder_config_path = hf_hub_download(model, "vocoder/config.json", revision=revision)
+            vocoder_config_path = hf_api().hf_hub_download(model, "vocoder/config.json", revision=revision)
         except Exception:
             return None
     try:
@@ -839,7 +839,7 @@ def load_transformer_config(
         if not os.path.exists(config_path):
             raise FileNotFoundError(f"LTX transformer config not found: {config_path}")
     else:
-        config_path = hf_hub_download(
+        config_path = hf_api().hf_hub_download(
             repo_id=model_path,
             filename=f"{subfolder}/config.json",
             revision=revision,
diff --git a/vllm_omni/diffusion/models/omnivoice/pipeline_omnivoice.py b/vllm_omni/diffusion/models/omnivoice/pipeline_omnivoice.py
index 62ac027ca8b..cfc40fcb6bc 100644
--- a/vllm_omni/diffusion/models/omnivoice/pipeline_omnivoice.py
+++ b/vllm_omni/diffusion/models/omnivoice/pipeline_omnivoice.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """
 OmniVoice TTS Pipeline for vLLM-Omni diffusion engine.
 
@@ -141,9 +141,9 @@ def __init__(self, *, od_config: OmniDiffusionConfig, prefix: str = ""):
 
         # Resolve model path (HF hub ID → local cache)
         if not os.path.isdir(self.model_path):
-            from huggingface_hub import snapshot_download
+            from vllm_omni.transformers_utils.repo_utils import hf_api
 
-            self.model_path = snapshot_download(self.model_path)
+            self.model_path = hf_api().snapshot_download(self.model_path)
 
         # Load OmniVoice config
         config_path = os.path.join(self.model_path, "config.json")
diff --git a/vllm_omni/diffusion/models/pi0/pipeline_pi0.py b/vllm_omni/diffusion/models/pi0/pipeline_pi0.py
index 1eeca29692c..a03801c8d4e 100644
--- a/vllm_omni/diffusion/models/pi0/pipeline_pi0.py
+++ b/vllm_omni/diffusion/models/pi0/pipeline_pi0.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """π0 (Pi-Zero) VLA pipeline for vllm-omni.
 
 Entry point for ``DiffusionEngine.step() → pipeline.forward(req)``. Mirrors the
@@ -94,9 +94,9 @@ def _resolve_model_dir(model: str | None) -> str | None:
             return None
         if os.path.isdir(model):
             return model
-        from huggingface_hub import snapshot_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
 
-        return snapshot_download(
+        return hf_api().snapshot_download(
             repo_id=model,
             allow_patterns=["*.json", "*.safetensors", "*.model", "tokenizer*"],
         )
diff --git a/vllm_omni/diffusion/models/sensenova_u1/pipeline_sensenova_u1.py b/vllm_omni/diffusion/models/sensenova_u1/pipeline_sensenova_u1.py
index de636fb6dca..2be1a9c7a39 100644
--- a/vllm_omni/diffusion/models/sensenova_u1/pipeline_sensenova_u1.py
+++ b/vllm_omni/diffusion/models/sensenova_u1/pipeline_sensenova_u1.py
@@ -27,7 +27,6 @@
 import torch
 import torch.nn as nn
 import torchvision.transforms as T
-from huggingface_hub import snapshot_download
 from PIL import Image
 from transformers import AutoTokenizer
 from vllm.logger import init_logger
@@ -109,7 +108,9 @@ def _resolve_model_path(model_path: str) -> str:
     """Resolve a HuggingFace model ID or local path to a local directory."""
     if os.path.isdir(model_path):
         return model_path
-    return snapshot_download(model_path)
+    from vllm_omni.transformers_utils.repo_utils import hf_api
+
+    return hf_api().snapshot_download(model_path)
 
 
 def _round_by_factor(number: float, factor: int) -> int:
diff --git a/vllm_omni/diffusion/models/utils.py b/vllm_omni/diffusion/models/utils.py
index 2d783d90712..429c34b9d5c 100644
--- a/vllm_omni/diffusion/models/utils.py
+++ b/vllm_omni/diffusion/models/utils.py
@@ -168,8 +168,8 @@ def _load_json(model_path: str, filename: str, local_files_only: bool = True) ->
         with open(path) as f:
             return json.load(f)
     else:
-        from huggingface_hub import hf_hub_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
 
-        cached = hf_hub_download(repo_id=model_path, filename=filename)
+        cached = hf_api().hf_hub_download(repo_id=model_path, filename=filename)
         with open(cached) as f:
             return json.load(f)
diff --git a/vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py b/vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py
index 2383fc6fd09..c9ef325c4ee 100644
--- a/vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py
+++ b/vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py
@@ -156,9 +156,9 @@ def load_transformer_config(model_path: str, subfolder: str = "transformer", loc
     else:
         # Try to download config from HF Hub
         try:
-            from huggingface_hub import hf_hub_download
+            from vllm_omni.transformers_utils.repo_utils import hf_api
 
-            config_path = hf_hub_download(
+            config_path = hf_api().hf_hub_download(
                 repo_id=model_path,
                 filename=f"{subfolder}/config.json",
             )
@@ -364,9 +364,9 @@ def __init__(
         else:
             # For remote models, download and read model_index.json
             try:
-                from huggingface_hub import hf_hub_download
+                from vllm_omni.transformers_utils.repo_utils import hf_api
 
-                model_index_path = hf_hub_download(repo_id=model, filename="model_index.json")
+                model_index_path = hf_api().hf_hub_download(repo_id=model, filename="model_index.json")
                 with open(model_index_path) as f:
                     model_index = json.load(f)
                     self.expand_timesteps = model_index.get("expand_timesteps", False)
diff --git a/vllm_omni/engine/arg_utils.py b/vllm_omni/engine/arg_utils.py
index f9ee40cbbff..bd6a7156215 100644
--- a/vllm_omni/engine/arg_utils.py
+++ b/vllm_omni/engine/arg_utils.py
@@ -381,9 +381,9 @@ def create_model_config(self) -> OmniModelConfig:
                 if tokenizer_subfolder:
                     # Download just the tokenizer files from the subfolder
                     try:
-                        from huggingface_hub import snapshot_download
+                        from vllm_omni.transformers_utils.repo_utils import hf_api
 
-                        local_dir = snapshot_download(
+                        local_dir = hf_api().snapshot_download(
                             model_path,
                             allow_patterns=[
                                 f"{tokenizer_subfolder}/tokenizer*",
diff --git a/vllm_omni/engine/stage_init_utils.py b/vllm_omni/engine/stage_init_utils.py
index eed686837a6..5645d2a366c 100644
--- a/vllm_omni/engine/stage_init_utils.py
+++ b/vllm_omni/engine/stage_init_utils.py
@@ -31,7 +31,6 @@
 from vllm.renderers import BaseRenderer
 from vllm.sampling_params import SamplingParams
 from vllm.tokenizers import cached_tokenizer_from_config
-from vllm.transformers_utils.repo_utils import hf_api
 from vllm.transformers_utils.runai_utils import is_runai_obj_uri
 from vllm.usage.usage_lib import UsageContext
 from vllm.v1.engine.input_processor import InputProcessor
@@ -59,6 +58,7 @@
 from vllm_omni.outputs.output_processor import MultimodalOutputProcessor
 from vllm_omni.platforms import current_omni_platform
 from vllm_omni.quantization.inc_config import OmniINCConfig
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 logger = init_logger(__name__)
 
diff --git a/vllm_omni/entrypoints/openai/api_server.py b/vllm_omni/entrypoints/openai/api_server.py
index d49d95175b5..1d53c387094 100644
--- a/vllm_omni/entrypoints/openai/api_server.py
+++ b/vllm_omni/entrypoints/openai/api_server.py
@@ -201,9 +201,9 @@ def _load_model_chat_template_json(model: str) -> str | None:
 
     if template_path is None:
         try:
-            from huggingface_hub import hf_hub_download
+            from vllm_omni.transformers_utils.repo_utils import hf_api
 
-            template_path = hf_hub_download(
+            template_path = hf_api().hf_hub_download(
                 repo_id=model,
                 filename="chat_template.json",
                 local_files_only=True,
diff --git a/vllm_omni/entrypoints/openai/tts_adapters/cosyvoice3.py b/vllm_omni/entrypoints/openai/tts_adapters/cosyvoice3.py
index 25993ed6e3e..1196c5c709d 100644
--- a/vllm_omni/entrypoints/openai/tts_adapters/cosyvoice3.py
+++ b/vllm_omni/entrypoints/openai/tts_adapters/cosyvoice3.py
@@ -116,9 +116,9 @@ def apply_sampling_overrides(
         if tokenizer is None:
             model_path = server.engine_client.model_config.model
             if not os.path.isdir(model_path):
-                from huggingface_hub import snapshot_download
+                from vllm_omni.transformers_utils.repo_utils import hf_api
 
-                model_path = snapshot_download(model_path)
+                model_path = hf_api().snapshot_download(model_path)
             tokenizer = get_qwen_tokenizer(
                 token_path=os.path.join(model_path, hf_cfg.qwen_pretrain_path),
                 skip_special_tokens=hf_cfg.skip_special_tokens,
diff --git a/vllm_omni/model_executor/model_loader/weight_utils.py b/vllm_omni/model_executor/model_loader/weight_utils.py
index 59c2cd86022..559a90b5062 100644
--- a/vllm_omni/model_executor/model_loader/weight_utils.py
+++ b/vllm_omni/model_executor/model_loader/weight_utils.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 import os
 import time
 from pathlib import Path
@@ -10,7 +13,11 @@
 if envs.VLLM_USE_MODELSCOPE:
     from modelscope.hub.snapshot_download import snapshot_download
 else:
-    from huggingface_hub import snapshot_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
+
+    def snapshot_download(*args, **kwargs):
+        return hf_api().snapshot_download(*args, **kwargs)
+
 
 logger = init_logger(__name__)
 
diff --git a/vllm_omni/model_executor/models/audex/checkpoint.py b/vllm_omni/model_executor/models/audex/checkpoint.py
index 38b0531e660..ae8230e05d6 100644
--- a/vllm_omni/model_executor/models/audex/checkpoint.py
+++ b/vllm_omni/model_executor/models/audex/checkpoint.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Audex checkpoint layout preparation.
 
 The nvidia/Nemotron-Labs-Audex-2B repo deduplicates weights: the
@@ -57,12 +57,12 @@ def _download_audiogen_index(model: str) -> str:
     metadata round trip per stage-engine boot; a cache miss falls back to
     the network.
     """
-    from huggingface_hub import hf_hub_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
 
     try:
-        return hf_hub_download(model, _AUDIOGEN_INDEX_FILE, local_files_only=True)
+        return hf_api().hf_hub_download(model, _AUDIOGEN_INDEX_FILE, local_files_only=True)
     except Exception:
-        return hf_hub_download(model, _AUDIOGEN_INDEX_FILE)
+        return hf_api().hf_hub_download(model, _AUDIOGEN_INDEX_FILE)
 
 
 def _dedup_shard_patterns(model: str) -> list[str]:
@@ -111,13 +111,13 @@ def ensure_audex_snapshot(model: str, profile: str = "tts") -> str:
     if profile in ("tts", "tta"):
         patterns += _dedup_shard_patterns(model)
 
-    from huggingface_hub import snapshot_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
 
     try:
-        return snapshot_download(model, allow_patterns=patterns)
+        return hf_api().snapshot_download(model, allow_patterns=patterns)
     except Exception as download_exc:
         try:
-            return snapshot_download(model, allow_patterns=patterns, local_files_only=True)
+            return hf_api().snapshot_download(model, allow_patterns=patterns, local_files_only=True)
         except Exception:
             raise RuntimeError(
                 f"Could not resolve the Audex repo {model!r} (profile {profile!r}): the download "
@@ -135,13 +135,13 @@ def ensure_xcodec1_snapshot(model: str | None) -> str:
     if os.path.isdir(model):
         return model
 
-    from huggingface_hub import snapshot_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
 
     try:
-        return snapshot_download(model)
+        return hf_api().snapshot_download(model)
     except Exception as download_exc:
         try:
-            return snapshot_download(model, local_files_only=True)
+            return hf_api().snapshot_download(model, local_files_only=True)
         except Exception:
             raise RuntimeError(
                 f"Could not resolve the XCodec1 checkpoint {model!r}: the download failed and "
diff --git a/vllm_omni/model_executor/models/bagel/bagel.py b/vllm_omni/model_executor/models/bagel/bagel.py
index a3dfd7bf151..495e0037f29 100644
--- a/vllm_omni/model_executor/models/bagel/bagel.py
+++ b/vllm_omni/model_executor/models/bagel/bagel.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 from collections.abc import Iterable, Mapping, Sequence
 from math import isqrt
 from typing import Any
@@ -122,9 +125,9 @@ def _patch_max_latent_size(self, config):
             if p.is_dir():
                 index_path = p / "model.safetensors.index.json"
             else:
-                from huggingface_hub import hf_hub_download
+                from vllm_omni.transformers_utils.repo_utils import hf_api
 
-                index_path = Path(hf_hub_download(model_name, "model.safetensors.index.json"))
+                index_path = Path(hf_api().hf_hub_download(model_name, "model.safetensors.index.json"))
 
             if not index_path.exists():
                 return
diff --git a/vllm_omni/model_executor/models/cosyvoice3/cosyvoice3.py b/vllm_omni/model_executor/models/cosyvoice3/cosyvoice3.py
index 81bae6d6c2b..84652cadabf 100644
--- a/vllm_omni/model_executor/models/cosyvoice3/cosyvoice3.py
+++ b/vllm_omni/model_executor/models/cosyvoice3/cosyvoice3.py
@@ -10,7 +10,6 @@
 import onnxruntime
 import torch
 import torch.nn as nn
-from huggingface_hub import snapshot_download
 from scipy.signal import resample_poly
 from transformers import Qwen2Config
 from transformers.feature_extraction_utils import BatchFeature
@@ -52,6 +51,7 @@
 from vllm_omni.model_executor.models.output_templates import OmniOutput
 from vllm_omni.platforms import current_omni_platform
 from vllm_omni.transformers_utils.configs.cosyvoice3 import CosyVoice3Config
+from vllm_omni.transformers_utils.repo_utils import hf_api
 from vllm_omni.utils.speaker_cache import get_speaker_cache
 
 logger = init_logger(__name__)
@@ -137,7 +137,7 @@ def _build_runtime_components(self, model_dir: str, config: CosyVoice3Config) ->
         """Build the per-model runtime components once (cached process-wide)."""
         # If model_dir is an HF repo ID (not a local path), resolve to cache.
         if not os.path.isdir(model_dir):
-            model_dir = snapshot_download(model_dir)
+            model_dir = hf_api().snapshot_download(model_dir)
 
         tokenizer = get_qwen_tokenizer(
             token_path=os.path.join(model_dir, config.qwen_pretrain_path),
@@ -463,7 +463,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.model_stage = vllm_config.model_config.model_stage
         model_dir = vllm_config.model_config.model
         if not os.path.isdir(model_dir):
-            model_dir = snapshot_download(model_dir)
+            model_dir = hf_api().snapshot_download(model_dir)
         self.model_dir = model_dir
         self.model = None
         if self.model_stage == "cosyvoice3_talker":
@@ -918,7 +918,7 @@ def _resolve_flow_estimator_onnx(self) -> str | None:
         repo = getattr(self.config, "flow_estimator_onnx_repo", None)
         if repo:
             try:
-                fetched_dir = snapshot_download(repo, allow_patterns=[fp16_name])
+                fetched_dir = hf_api().snapshot_download(repo, allow_patterns=[fp16_name])
                 fetched = os.path.join(fetched_dir, fp16_name)
                 if os.path.exists(fetched):
                     return fetched
diff --git a/vllm_omni/model_executor/models/covo_audio/covo_audio_code2wav.py b/vllm_omni/model_executor/models/covo_audio/covo_audio_code2wav.py
index 4b0b2fe4bea..15cd0875e64 100644
--- a/vllm_omni/model_executor/models/covo_audio/covo_audio_code2wav.py
+++ b/vllm_omni/model_executor/models/covo_audio/covo_audio_code2wav.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 # Copyright 2026 Tencent.
 import os
 from collections.abc import Iterable
@@ -25,9 +28,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         if os.path.isdir(model_name):
             model_path = model_name
         else:
-            from huggingface_hub import snapshot_download
+            from vllm_omni.transformers_utils.repo_utils import hf_api
 
-            model_path = snapshot_download(model_name)
+            model_path = hf_api().snapshot_download(model_name)
         token2wav_path = os.path.join(model_path, "token2wav")
 
         code2wav_config = CovoAudioCode2WavConfig()
diff --git a/vllm_omni/model_executor/models/dots_tts/dots_tts_talker.py b/vllm_omni/model_executor/models/dots_tts/dots_tts_talker.py
index ff2f460d637..862af8e17cf 100644
--- a/vllm_omni/model_executor/models/dots_tts/dots_tts_talker.py
+++ b/vllm_omni/model_executor/models/dots_tts/dots_tts_talker.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """dots.tts talker — vLLM-native AR base LM + audio side path.
 
 Mirrors upstream rednote-hilab/dots.tts (pinned @ a393d2e):
@@ -516,9 +516,9 @@ def _resolve_latent_stats_path(model_arg: str) -> str | None:
         if os.path.exists(local):
             return local
         try:
-            from huggingface_hub import hf_hub_download
+            from vllm_omni.transformers_utils.repo_utils import hf_api
 
-            return hf_hub_download(
+            return hf_api().hf_hub_download(
                 repo_id=model_arg,
                 filename="latent_stats.pt",
                 local_files_only=True,
diff --git a/vllm_omni/model_executor/models/dynin_omni/dynin_omni_common.py b/vllm_omni/model_executor/models/dynin_omni/dynin_omni_common.py
index 5c67decaa56..125822c1ef4 100644
--- a/vllm_omni/model_executor/models/dynin_omni/dynin_omni_common.py
+++ b/vllm_omni/model_executor/models/dynin_omni/dynin_omni_common.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 from __future__ import annotations
 
 import hashlib
@@ -20,9 +23,9 @@
 logger = init_logger(__name__)
 
 try:
-    from huggingface_hub import snapshot_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
 except Exception:  # pragma: no cover
-    snapshot_download = None
+    hf_api = None
 
 
 class DetokTarget(IntEnum):
@@ -397,13 +400,13 @@ def _find_dynin_config_under_root(root: Path) -> Path | None:
 
 @lru_cache(maxsize=16)
 def _resolve_dynin_config_from_hf_repo(repo_id: str) -> str | None:
-    if not _looks_like_hf_repo_id(repo_id) or snapshot_download is None:
+    if not _looks_like_hf_repo_id(repo_id) or hf_api is None:
         return None
 
     try:
         snapshot_dir = (
             Path(
-                snapshot_download(
+                hf_api().snapshot_download(
                     repo_id=repo_id,
                     repo_type="model",
                     allow_patterns=list(_DYNIN_CONFIG_CANDIDATE_RELPATHS),
@@ -719,7 +722,7 @@ def _resolve_remote_snapshot_dir(
     if source_path.is_dir():
         return str(source_path.resolve())
 
-    if snapshot_download is None:
+    if hf_api is None:
         raise RuntimeError("huggingface_hub is required to load remote code.")
 
     kwargs: dict[str, Any] = {
@@ -732,10 +735,10 @@ def _resolve_remote_snapshot_dir(
         kwargs["revision"] = revision
 
     try:
-        return str(snapshot_download(**kwargs))
+        return str(hf_api().snapshot_download(**kwargs))
     except TypeError:
         kwargs.pop("local_files_only", None)
-        return str(snapshot_download(**kwargs))
+        return str(hf_api().snapshot_download(**kwargs))
 
 
 def _ensure_remote_package(snapshot_dir: str) -> str:
diff --git a/vllm_omni/model_executor/models/dynin_omni/dynin_omni_token2audio.py b/vllm_omni/model_executor/models/dynin_omni/dynin_omni_token2audio.py
index 80bc0a78b47..dbbb635245f 100644
--- a/vllm_omni/model_executor/models/dynin_omni/dynin_omni_token2audio.py
+++ b/vllm_omni/model_executor/models/dynin_omni/dynin_omni_token2audio.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 from __future__ import annotations
 
 import os
@@ -80,7 +83,7 @@ def _ensure_remote_s2u_vendor_root(
             return str(existing_path)
 
     try:
-        from huggingface_hub import snapshot_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
     except Exception as e:
         logger.warning("huggingface_hub unavailable; cannot fetch s2u_vendor from %s: %s", repo_id, e)
         return None
@@ -91,7 +94,7 @@ def _ensure_remote_s2u_vendor_root(
 
     for revision in revisions:
         try:
-            snapshot_dir = snapshot_download(
+            snapshot_dir = hf_api().snapshot_download(
                 repo_id=repo_id,
                 revision=revision,
                 allow_patterns=["s2u_vendor/**"],
@@ -99,7 +102,7 @@ def _ensure_remote_s2u_vendor_root(
             )
         except TypeError:
             try:
-                snapshot_dir = snapshot_download(
+                snapshot_dir = hf_api().snapshot_download(
                     repo_id=repo_id,
                     revision=revision,
                     allow_patterns=["s2u_vendor/**"],
diff --git a/vllm_omni/model_executor/models/glm_tts/glm_tts.py b/vllm_omni/model_executor/models/glm_tts/glm_tts.py
index 664a794c4c0..fb3517036d0 100644
--- a/vllm_omni/model_executor/models/glm_tts/glm_tts.py
+++ b/vllm_omni/model_executor/models/glm_tts/glm_tts.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """GLM-TTS AR Model (Stage 0): Text → Speech Tokens.
 
 Based on Llama architecture, generates speech token sequences from input text.
@@ -113,9 +113,9 @@ def resolve_glm_tts_tokenizer_path(model_name_or_path: Any) -> str:
                 return candidate
         return model_path
 
-    from huggingface_hub import snapshot_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
 
-    local_dir = snapshot_download(
+    local_dir = hf_api().snapshot_download(
         model_path,
         allow_patterns=[
             f"{_GLM_TTS_TOKENIZER_SUBDIR}/tokenizer*",
@@ -169,9 +169,9 @@ def _has_files(root: str) -> bool:
         except Exception:
             pass
 
-    from huggingface_hub import snapshot_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
 
-    return snapshot_download(model_name_or_path)
+    return hf_api().snapshot_download(model_name_or_path)
 
 
 def _first_glm_tts_value(value: Any) -> Any:
diff --git a/vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_code2wav.py b/vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_code2wav.py
index 91609e1a4ee..c6a54a39350 100644
--- a/vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_code2wav.py
+++ b/vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_code2wav.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Stage 1 (codec decoder) for higgs-audio v2.
 
 Two surfaces exposed:
@@ -231,9 +231,9 @@ def _resolve_hf_id_to_local(repo_id: str) -> str:
                 if os.path.isfile(os.path.join(candidate, "config.json")):
                     return candidate
 
-        from huggingface_hub import snapshot_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
 
-        return snapshot_download(repo_id=repo_id)
+        return hf_api().snapshot_download(repo_id=repo_id)
 
     # ------------------------------------------------------ direct decode API
     @torch.inference_mode()
diff --git a/vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_tokenizer.py b/vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_tokenizer.py
index c3f5024b3bc..587bcda6718 100755
--- a/vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_tokenizer.py
+++ b/vllm_omni/model_executor/models/higgs_audio_v2/higgs_audio_v2_tokenizer.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """TTS prompt builder + scope validators for higgs-audio v2.
 
 vllm-omni's higgs path supports two request shapes:
@@ -248,9 +248,9 @@ def _load_audio_tokenizer():
 
     audio_tokenizer_dir = _resolve_audio_tokenizer_dir()
     if audio_tokenizer_dir is None:
-        from huggingface_hub import snapshot_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
 
-        repo_path = snapshot_download(
+        repo_path = hf_api().snapshot_download(
             _K2_OMNIVOICE_REPO,
             allow_patterns=[f"{_K2_OMNIVOICE_SUBDIR}/*"],
         )
diff --git a/vllm_omni/model_executor/models/indextts2/preprocess_utils.py b/vllm_omni/model_executor/models/indextts2/preprocess_utils.py
index 88e9a063952..82729160270 100644
--- a/vllm_omni/model_executor/models/indextts2/preprocess_utils.py
+++ b/vllm_omni/model_executor/models/indextts2/preprocess_utils.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """External model loading, audio I/O, and emotion conditioning for IndexTTS2."""
 
 from __future__ import annotations
@@ -171,9 +171,10 @@ def load_semantic_codec(
         codec.load_state_dict(state, strict=False)
     else:
         import safetensors.torch
-        from huggingface_hub import hf_hub_download
 
-        ckpt_path = hf_hub_download("amphion/MaskGCT", filename="semantic_codec/model.safetensors")
+        from vllm_omni.transformers_utils.repo_utils import hf_api
+
+        ckpt_path = hf_api().hf_hub_download("amphion/MaskGCT", filename="semantic_codec/model.safetensors")
         safetensors.torch.load_model(codec, ckpt_path)
     _freeze(codec.to(device=device, dtype=torch.float32))
     _semantic_codec_cache[cache_key] = codec
@@ -198,9 +199,11 @@ def load_campplus(model_path: str, device: torch.device):
         if ckpt_path is not None:
             break
     if ckpt_path is None:
-        from huggingface_hub import hf_hub_download
+        if os.path.isdir(model_path):
+            raise FileNotFoundError(f"IndexTTS CAMPPlus checkpoint is missing from local bundle {model_path!r}")
+        from vllm_omni.transformers_utils.repo_utils import hf_api
 
-        ckpt_path = hf_hub_download("funasr/campplus", filename="campplus_cn_common.bin")
+        ckpt_path = hf_api().hf_hub_download("funasr/campplus", filename="campplus_cn_common.bin")
     state = torch.load(ckpt_path, map_location="cpu", weights_only=True)
     campplus.load_state_dict(state, strict=False)
     _freeze(campplus.to(device=device, dtype=torch.float32))
diff --git a/vllm_omni/model_executor/models/indextts2/s2mel/modules/bigvgan.py b/vllm_omni/model_executor/models/indextts2/s2mel/modules/bigvgan.py
index 5876bb14742..048d1de2dcf 100644
--- a/vllm_omni/model_executor/models/indextts2/s2mel/modules/bigvgan.py
+++ b/vllm_omni/model_executor/models/indextts2/s2mel/modules/bigvgan.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 # Copyright (c) 2024 NVIDIA CORPORATION.
 #   Licensed under the MIT license.
 
@@ -9,13 +12,14 @@
 
 import torch
 import torch.nn as nn
-from huggingface_hub import PyTorchModelHubMixin, hf_hub_download
+from huggingface_hub import PyTorchModelHubMixin
 from torch.nn import Conv1d, ConvTranspose1d
 from torch.nn.utils import remove_weight_norm, weight_norm
 
 from vllm_omni.model_executor.models.common.alias_free_activation import AliasFreeActivation1d
 from vllm_omni.model_executor.models.common.snake_activation import Snake, SnakeBeta
 from vllm_omni.model_executor.models.indextts2.s2mel.modules.commons import AttrDict
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 # ---------------------------------------------------------------------------
 # Helpers (inlined from env.py / utils.py)
@@ -328,7 +332,7 @@ def _from_pretrained(cls, *, model_id: str, map_location: str = "cpu", **kwargs)
         def _resolve(filename):
             if os.path.isdir(model_id):
                 return os.path.join(model_id, filename)
-            return hf_hub_download(
+            return hf_api().hf_hub_download(
                 repo_id=model_id,
                 filename=filename,
                 **{
diff --git a/vllm_omni/model_executor/models/indextts2/tokenizer_v2_5.py b/vllm_omni/model_executor/models/indextts2/tokenizer_v2_5.py
index 6eae97703a7..ba27cad6243 100644
--- a/vllm_omni/model_executor/models/indextts2/tokenizer_v2_5.py
+++ b/vllm_omni/model_executor/models/indextts2/tokenizer_v2_5.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """IndexTTS 2.5 multilingual tiktoken vocabulary.
 
 The vocabulary layout mirrors the official IndexTTS 2.5 tokenizer: 58,836
@@ -234,9 +234,9 @@ def resolve_indextts25_tokenizer_file(
         return str(model_path)
 
     try:
-        from huggingface_hub import hf_hub_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
 
-        return hf_hub_download(repo_id=model_dir, filename=tokenizer_file)
+        return hf_api().hf_hub_download(repo_id=model_dir, filename=tokenizer_file)
     except Exception as exc:
         raise FileNotFoundError(
             f"Could not resolve IndexTTS 2.5 tokenizer {tokenizer_file!r} from {model_dir!r}"
diff --git a/vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_code2wav.py b/vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_code2wav.py
index 8b3926cc0ca..f701e8b2699 100644
--- a/vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_code2wav.py
+++ b/vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_code2wav.py
@@ -39,9 +39,9 @@ def _resolve_model_dir(model_ref: str, revision: str | None = None) -> str:
     """
     if Path(model_ref).is_dir():
         return model_ref
-    from huggingface_hub import snapshot_download
+    from vllm_omni.transformers_utils.repo_utils import hf_api
 
-    return snapshot_download(model_ref, revision=revision, allow_patterns=["assets/*"])
+    return hf_api().snapshot_download(model_ref, revision=revision, allow_patterns=["assets/*"])
 
 
 def _batch_error(reason: str, **details: Any) -> RuntimeError:
diff --git a/vllm_omni/model_executor/models/omnivoice/omnivoice.py b/vllm_omni/model_executor/models/omnivoice/omnivoice.py
index c167aa1c0a1..26a6f6b5935 100644
--- a/vllm_omni/model_executor/models/omnivoice/omnivoice.py
+++ b/vllm_omni/model_executor/models/omnivoice/omnivoice.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """
 OmniVoice model for vLLM-Omni two-stage TTS pipeline.
 
@@ -487,9 +487,9 @@ def _resolve_model_dir(self) -> str:
         if os.path.isdir(model_dir):
             return model_dir
         # HF hub model ID — resolve to local cache
-        from huggingface_hub import snapshot_download
+        from vllm_omni.transformers_utils.repo_utils import hf_api
 
-        return snapshot_download(model_dir)
+        return hf_api().snapshot_download(model_dir)
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         try:
diff --git a/vllm_omni/model_executor/models/personaplex/personaplex_mimi.py b/vllm_omni/model_executor/models/personaplex/personaplex_mimi.py
index 48df20fab1e..3b8813faf70 100644
--- a/vllm_omni/model_executor/models/personaplex/personaplex_mimi.py
+++ b/vllm_omni/model_executor/models/personaplex/personaplex_mimi.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Streaming Mimi codec for PersonaPlex (moshi-free).
 
 Frame-clocked duplex needs a codec that encodes/decodes exactly one 80 ms frame
@@ -288,10 +288,11 @@ class PersonaPlexMimiCodec(nn.Module):
 
     def __init__(self, hf_repo: str = DEFAULT_HF_REPO, checkpoint: str | None = None, device: str = "cuda") -> None:
         super().__init__()
-        from huggingface_hub import hf_hub_download
         from safetensors.torch import load_file
         from transformers import MimiConfig, MimiModel
 
+        from vllm_omni.transformers_utils.repo_utils import hf_api
+
         self.device = torch.device(device)
 
         # The PersonaPlex repo ships the reference mimi checkpoint in the moshi
@@ -300,7 +301,7 @@ def __init__(self, hf_repo: str = DEFAULT_HF_REPO, checkpoint: str | None = None
         # transformer stacks remain absent because the streaming implementations
         # below replace them with the checkpoint's fused QKV layout.
         if checkpoint is None:
-            checkpoint = hf_hub_download(
+            checkpoint = hf_api().hf_hub_download(
                 "nvidia/personaplex-7b-v1",
                 "tokenizer-e351c8d8-checkpoint125.safetensors",
             )
diff --git a/vllm_omni/model_executor/models/step_audio2/step_audio2_token2wav.py b/vllm_omni/model_executor/models/step_audio2/step_audio2_token2wav.py
index e9ca78b21eb..8e3c56b327d 100644
--- a/vllm_omni/model_executor/models/step_audio2/step_audio2_token2wav.py
+++ b/vllm_omni/model_executor/models/step_audio2/step_audio2_token2wav.py
@@ -461,9 +461,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         if model_path is None:
             # Resolve HF repo names to local cache path
             if not os.path.isdir(model_name_or_path):
-                from huggingface_hub import snapshot_download
+                from vllm_omni.transformers_utils.repo_utils import hf_api
 
-                model_name_or_path = snapshot_download(model_name_or_path)
+                model_name_or_path = hf_api().snapshot_download(model_name_or_path)
             model_path = f"{model_name_or_path}/token2wav"
 
         float16 = getattr(self.config, "token2wav_float16", False)
diff --git a/vllm_omni/model_executor/models/voxtral_tts/voxtral_tts.py b/vllm_omni/model_executor/models/voxtral_tts/voxtral_tts.py
index 69cfa6d2667..c61aeaaadbb 100644
--- a/vllm_omni/model_executor/models/voxtral_tts/voxtral_tts.py
+++ b/vllm_omni/model_executor/models/voxtral_tts/voxtral_tts.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 from collections.abc import Iterable, Mapping
 from dataclasses import replace
 from functools import cached_property
@@ -6,7 +9,6 @@
 import regex as re
 import torch
 import torch.nn as nn
-from huggingface_hub import hf_hub_download
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
 from vllm.model_executor.models.interfaces import SupportsMultiModal
@@ -28,6 +30,7 @@
     VoxtralTTSMultiModalProcessor,
     VoxtralTTSProcessingInfo,
 )
+from vllm_omni.transformers_utils.repo_utils import hf_api
 
 logger = init_logger(__name__)
 
@@ -133,7 +136,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 self.voice_to_embedding = {}
                 for sid in speaker_id:
                     if self.is_hf_model:
-                        path = hf_hub_download(repo_id=self.repo_id, filename=f"voice_embedding/{sid}.pt")
+                        path = hf_api().hf_hub_download(repo_id=self.repo_id, filename=f"voice_embedding/{sid}.pt")
                     else:
                         path = Path(self.repo_id) / "voice_embedding" / f"{sid}.pt"
                     if Path(path).exists():
diff --git a/vllm_omni/transformers_utils/repo_utils.py b/vllm_omni/transformers_utils/repo_utils.py
new file mode 100644
index 00000000000..e780c4b53fc
--- /dev/null
+++ b/vllm_omni/transformers_utils/repo_utils.py
@@ -0,0 +1,20 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Utilities for model repo interaction."""
+
+from huggingface_hub import HfApi
+
+from vllm_omni.version import __version__ as VLLM_OMNI_VERSION
+
+_hf_api: HfApi | None = None
+
+
+def hf_api() -> HfApi:
+    """Return a shared HfApi instance tagged with vLLM-Omni's library info."""
+    global _hf_api
+    if _hf_api is None:
+        _hf_api = HfApi(
+            library_name="vllm-omni",
+            library_version=VLLM_OMNI_VERSION,
+        )
+    return _hf_api