The open-source AI voice studio. Clone, dictate, create.
Fix infinite HF retry storm when loading a cached model offline
Roman Dolgov0517c0687ade
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 0b03f5e4..f598cc41 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -7,6 +7,14 @@## [Unreleased]+### Reliability++- **Cached models no longer retry HuggingFace when offline.** Loading a fully-downloaded+ model now forces offline mode for the duration of the load, so it skips the network HEAD+ request (and its 5-retry backoff) for every config file — `config.json`,+ `generation_config.json`, and the rest — instead of retrying each one in sequence before the+ app becomes ready.+### Linux- **ROCm setup works on Linux AMD systems.** Docker ROCm builds now keep PyTorchdiff --git a/backend/backends/base.py b/backend/backends/base.pyindex 70ec11ef..672022e8 100644--- a/backend/backends/base.py+++ b/backend/backends/base.py@@ -14,6 +14,7 @@ from typing import Callable, List, Optional, Tupleimport numpy as npfrom ..utils.audio import normalize_audio, load_audio+from ..utils.hf_offline_patch import force_offline_if_cachedfrom ..utils.progress import get_progress_managerfrom ..utils.hf_progress import HFProgressTracker, create_hf_progress_callbackfrom ..utils.tasks import get_task_manager@@ -280,7 +281,8 @@ def model_load_progress()try:- yield tracker_context+ with force_offline_if_cached(is_cached, model_name):+ yield tracker_contextexcept Exception as e:# Report error to both managersprogress_manager.mark_error(model_name, str(e))diff --git a/backend/tests/test_model_load_progress_offline.py b/backend/tests/test_model_load_progress_offline.pynew file mode 100644index 00000000..4befe58b--- /dev/null+++ b/backend/tests/test_model_load_progress_offline.py@@ -0,0 +1,55 @@+"""+Regression test for voicebox #434 (infinite HF retry loop on a cached model).++``model_load_progress`` is the single context manager every backend's+``_load_model_sync`` uses around its ``from_pretrained()`` call. It already+receives ``is_cached`` but never forwarded it to ``force_offline_if_cached``,+so a fully-cached model still resolved config files against huggingface.co+and ate the default 5-retry backoff per file offline. This asserts the guard+is actually active for the duration of the ``with`` block when ``is_cached``+is ``True``, and restored on exit.++NOTE: mutates process-global state in ``huggingface_hub.constants`` and+``transformers.utils.hub`` (via ``force_offline_if_cached``); run serially,+same caveat as ``test_offline_guard.py``.+"""++import sys+from pathlib import Path++sys.path.insert(0, str(Path(__file__).parent.parent.parent))++from backend.backends.base import model_load_progress+++def _hf_const():+ import huggingface_hub.constants as hf_const++ return hf_const+++def _tf_hub():+ import transformers.utils.hub as tf_hub++ return tf_hub+++def test_cached_model_load_forces_offline_mode():+ original_hf = _hf_const().HF_HUB_OFFLINE+ original_tf = _tf_hub()._is_offline_mode++ with model_load_progress("test-model", is_cached=True):+ assert _hf_const().HF_HUB_OFFLINE is True+ assert _tf_hub()._is_offline_mode is True++ assert original_hf == _hf_const().HF_HUB_OFFLINE+ assert original_tf == _tf_hub()._is_offline_mode+++def test_uncached_model_load_does_not_force_offline_mode():+ original_hf = _hf_const().HF_HUB_OFFLINE++ with model_load_progress("test-model", is_cached=False):+ assert original_hf == _hf_const().HF_HUB_OFFLINE++ assert original_hf == _hf_const().HF_HUB_OFFLINE