fix: tolerate Confucius and DOTS accelerator probe failures

CHANGELOG.md

@@ -20,7 +20,7 @@ the frozen-backend fallback mirror it for their toolchains.
 
 ### Fixed
 
-- Confucius accelerator routing matches device selection, and dots.tts keeps safe CPU precision on non-CUDA accelerators (#1831) — thanks @li-lizhe!
+- Confucius accelerator routing tolerates failed device probes, and dots.tts keeps safe default precision on non-CUDA hosts (#1831) — thanks @li-lizhe!
 
 
 ## [0.5.2] — 2026-09-02

backend/engines/confucius4/main.py

@@ -115,6 +115,7 @@ def _load_model(stdout):
     import torch
     from confuciustts.cli.inference import ConfuciusTTS  # type: ignore[import-not-found]
 
+    try:
         # Existing manually provisioned venvs may predate torch.accelerator.
         current_accelerator = getattr(getattr(torch, "accelerator", None), "current_accelerator", None)
         if current_accelerator is None:
@@ -122,6 +123,8 @@ def _load_model(stdout):
         else:
             device = current_accelerator(check_available=True)
         device = device.type if device is not None else "cpu"  # 'cuda', 'npu', 'mps', 'xpu', 'cpu'
+    except Exception:
+        device = "cpu"  # Broken accelerator drivers must not block CPU loading.
     if device == "mps":
         device = "cpu"  # MPS was slower than CPU in the existing validation run
     _send(stdout, {"op": "progress", "stage": "loading_model", "percent": 50})

backend/engines/dots_tts/main.py

@@ -117,7 +117,10 @@ def _load_runtime(stdout):
     repo = os.environ.get("OMNIVOICE_DOTS_TTS_MODEL", _DEFAULT_REPO)
     # Match DotsTtsRuntime's own CUDA/CPU selection. Its _check_torch_env
     # rejects half precision without CUDA, even when an XPU/NPU is available.
+    try:
         default_precision = "bfloat16" if torch.cuda.is_available() else "float32"
+    except Exception:
+        default_precision = "float32"  # Probe failure must not force half precision.
     precision = os.environ.get("OMNIVOICE_DOTS_TTS_PRECISION", default_precision)
     optimize = os.environ.get("OMNIVOICE_DOTS_TTS_OPTIMIZE", "0") == "1"
 

docs/engines/confucius4-tts.md

@@ -96,3 +96,5 @@ through upstream's device-aware model loading. The engine venv needs a matching
 PyTorch/vendor runtime. XPU/NPU selection is covered by mocked loader and routing
 tests; this change does not certify synthesis on physical XPU/NPU hardware.
 MPS keeps the existing CPU fallback described in the validation record above.
+If modern accelerator detection or the legacy CUDA probe raises, loading falls
+back to CPU instead of aborting before model construction.

docs/engines/dots-tts.md

@@ -119,4 +119,5 @@ disk and how uv keeps the cost down, see
 The upstream runtime selects CUDA or CPU internally. Automatic precision follows
 that selection: bfloat16 on CUDA, float32 otherwise, including XPU/NPU/MPS hosts
 where this runtime executes on CPU. `OMNIVOICE_DOTS_TTS_PRECISION` remains an
-explicit override.
+explicit override. If the CUDA availability probe raises, the automatic precision
+default stays float32; upstream remains responsible for its device selection.

tests/test_confucius4_sidecar.py

@@ -210,3 +210,24 @@ def test_load_model_matches_routing(sc, monkeypatch, family, legacy):
         accelerator.assert_called_once_with(check_available=True)
     caps = HostCaps(family=family or 'cpu', available_families=(family, 'cpu') if family else ('cpu',))
     assert resolve_routing(Confucius4Backend.gpu_compat, caps)['effective_device'] == expected
+
+
+@pytest.mark.parametrize("legacy", [False, True])
+def test_load_model_falls_back_when_accelerator_probe_raises(sc, monkeypatch, legacy):
+    import sys
+    from types import SimpleNamespace
+    from unittest.mock import Mock
+
+    probe = Mock(side_effect=RuntimeError("accelerator driver unavailable"))
+    monkeypatch.setitem(sys.modules, "torch", SimpleNamespace(
+        accelerator=SimpleNamespace() if legacy else SimpleNamespace(current_accelerator=probe),
+        cuda=SimpleNamespace(is_available=probe),
+    ))
+    constructor = Mock()
+    monkeypatch.setitem(sys.modules, "confuciustts.cli.inference", SimpleNamespace(ConfuciusTTS=constructor))
+    monkeypatch.setattr(sc, "_model", None)
+    monkeypatch.setattr(sc, "_config_path", lambda: "fixture.yaml")
+    monkeypatch.setattr(sc, "_ensure_clone_on_sys_path", lambda: None)
+    sc._load_model(io.BytesIO())
+    constructor.assert_called_once_with(config_path="fixture.yaml", device="cpu")
+    assert probe.call_count == 1

tests/test_dots_tts_accelerator_precision.py

@@ -22,3 +22,20 @@ def test_precision_matches_runtime_device(monkeypatch, family):
     monkeypatch.delenv('OMNIVOICE_DOTS_TTS_PRECISION', raising=False)
     main._load_runtime(io.BytesIO())
     assert loader.from_pretrained.call_args.kwargs['precision'] == ('bfloat16' if family == 'cuda' else 'float32')
+
+
+@pytest.mark.parametrize("override", [None, "float16"])
+def test_precision_probe_failure_uses_safe_default_or_explicit_override(monkeypatch, override):
+    from engines.dots_tts import main
+
+    probe = Mock(side_effect=RuntimeError("CUDA driver unavailable"))
+    monkeypatch.setitem(sys.modules, "torch", SimpleNamespace(cuda=SimpleNamespace(is_available=probe)))
+    loader = Mock()
+    monkeypatch.setitem(sys.modules, "dots_tts.runtime", SimpleNamespace(DotsTtsRuntime=loader))
+    monkeypatch.setattr(main, "_runtime", None)
+    if override is None:
+        monkeypatch.delenv("OMNIVOICE_DOTS_TTS_PRECISION", raising=False)
+    else:
+        monkeypatch.setenv("OMNIVOICE_DOTS_TTS_PRECISION", override)
+    main._load_runtime(io.BytesIO())
+    assert loader.from_pretrained.call_args.kwargs["precision"] == (override or "float32")