fix(moss): align accelerator detection and routing status

CHANGELOG.md

@@ -20,6 +20,8 @@ the frozen-backend fallback mirror it for their toolchains.
 
 ### Fixed
 
+- MOSS accelerator routing now matches runtime device selection, including native XPU and registered NPU detection (#1830) — thanks @li-lizhe!
+
 
 ## [0.5.2] — 2026-09-02
 

backend/api/routers/settings.py

@@ -150,7 +150,7 @@ def _compute_device_state() -> dict:
     caps = device_caps.detect_host_caps()
     env_pin = (os.environ.get("OMNIVOICE_DEVICE") or "").strip().lower()
     auto_family = next(
-        (f for f in ("cuda", "rocm", "xpu", "mps") if f in caps.available_families),
+        (f for f in device_caps.ACCELERATOR_PRIORITY if f in caps.available_families),
         "cpu",
     )
     value = device_caps.requested_device_override()

backend/core/device_caps.py

@@ -35,7 +35,8 @@ import sys
 from dataclasses import dataclass
 from typing import Literal
 
-DeviceFamily = Literal["cuda", "rocm", "mps", "xpu", "cpu"]
+DeviceFamily = Literal["cuda", "rocm", "mps", "xpu", "npu", "cpu"]
+ACCELERATOR_PRIORITY = ("cuda", "rocm", "xpu", "npu", "mps")
 
 # Stable substring stamped onto notes that represent a real kernel-launch risk
 # (arch/driver mismatch) — as opposed to advisory notes (multi-GPU, VRAM query
@@ -533,9 +534,12 @@ def _probe() -> HostCaps:
         # is the whole truth in that case (CodeRabbit, #1425).
         notes.extend(why_no_gpu(torch))
 
-    # ── Intel XPU via IPEX ───────────────────────────────────────────────
+    # Older builds register XPU through IPEX; modern torch exposes it directly.
     try:
         import intel_extension_for_pytorch  # noqa: F401
+    except Exception:
+        pass
+    try:
         if hasattr(torch, "xpu") and torch.xpu.is_available():
             detected.append("xpu")
             if not device_name:
@@ -546,7 +550,21 @@ def _probe() -> HostCaps:
                     pass
             notes.append("XPU VRAM not queried (unreliable across IPEX versions)")
     except Exception:
-        # IPEX absent or XPU probe failed — no XPU on this host.
+        # XPU probe failed — no usable XPU on this host.
+        pass
+
+    # Vendor extensions may register an NPU with torch. Probe only an already
+    # registered backend; never install or import an optional vendor package.
+    try:
+        if hasattr(torch, "npu") and torch.npu.is_available():
+            detected.append("npu")
+            if not device_name:
+                try:
+                    device_name = torch.npu.get_device_name(0)
+                except Exception:
+                    pass
+            notes.append("NPU VRAM not queried")
+    except Exception:
         pass
 
     # ── Apple Silicon MPS ────────────────────────────────────────────────
@@ -579,7 +597,7 @@ def _probe() -> HostCaps:
 
     # Preferred family by priority; cpu when nothing accelerated was detected.
     family: DeviceFamily = "cpu"
-    for pref in ("cuda", "rocm", "xpu", "mps"):
+    for pref in ACCELERATOR_PRIORITY:
         if pref in detected:
             family = pref  # type: ignore[assignment]
             break

backend/engines/moss_tts_v15/__init__.py

@@ -29,15 +29,11 @@ Do NOT import ``main.py`` from the parent process — it runs under a
 different venv (``transformers==5.0.0``) and importing it in-process would
 re-introduce the exact conflict this isolation exists to avoid.
 
-Hardware honesty (cross-platform rule): MOSS-TTS-v1.5's upstream documents
+Hardware routing follows the sidecar's runtime-available PyTorch accelerator:
-only CUDA and CPU. There is **no documented or tested MPS path** — the
+CUDA/ROCm, XPU, or a registered NPU. MPS remains excluded; CPU is the fallback.
-custom ``trust_remote_code`` modelling code and the separate audio
+XPU/NPU routing is covered with mocked device contracts, not physical-hardware
-tokenizer are unverified on Apple Silicon. We therefore advertise
+synthesis certification; users need a compatible torch/vendor runtime in the
-``gpu_compat = ("cuda", "cpu")`` and the sidecar selects ``cuda`` when
+isolated engine venv.
-present else ``cpu`` — it never silently routes to MPS where it might
-crash. On Apple Silicon the engine honestly resolves to CPU (slow but
-correct), and the engine is opt-in regardless, so it never becomes a
-broken default on any platform.
 """
 from __future__ import annotations
 
@@ -91,7 +87,7 @@ class MossTTSV15Backend(SubprocessBackend):
     _DEFAULT_SAMPLE_RATE = 24000
     # Honest hardware surface: upstream documents CUDA + CPU only. MPS is
     # undocumented / untested, so we do NOT claim it (cross-platform rule).
-    gpu_compat = ("cuda", "cpu")
+    gpu_compat = ("cuda", "rocm", "xpu", "npu", "cpu")
 
     # ── availability ───────────────────────────────────────────────────────
 

backend/services/tts_backend.py

@@ -2371,7 +2371,7 @@ def list_backends() -> list[dict]:
           "one_click_install": bool,                # services.sidecar_install can provision it in-app
           "last_error":     Optional[str],          # cached most-recent failure
           "isolation_mode": "in-process" | "subprocess",
-          "gpu_compat":     list[str],              # subset of {cuda, rocm, mps, xpu, cpu}
+          "gpu_compat":     list[str],              # subset of {cuda, rocm, mps, xpu, npu, cpu}
           "supports_cloning": Optional[bool],       # True/False from the class attr; None when
                                                     #   model-dependent (property, e.g. mlx-audio)
           "effective_device": str,                  # device this engine uses on THIS host

docs/engines/moss-tts-v15.md

@@ -23,10 +23,11 @@ interpreter, so MOSS runs behind
   8 GB GPUs when quantized; the bf16 Transformers path used here is ~16 GB
   of weights, so a 16 GB+ GPU is the realistic CUDA target. It also runs on
   **CPU** (fp32) — correct but slow.
-- **Device:** CUDA when present, else CPU. **There is no MPS path** —
+- **Device:** the sidecar uses a runtime-available PyTorch CUDA/ROCm, XPU,
-  upstream documents only CUDA/CPU and the custom modelling code is
+  or registered NPU backend, otherwise CPU. The isolated engine venv needs the
-  untested on Apple Silicon, so VoiceStudio never routes MOSS to MPS. On a
+  matching torch/vendor integration. MPS still uses CPU. XPU/NPU routing is
-  Mac it runs on CPU.
+  covered by mocked loader tests; physical-device synthesis has not been
+  validated by this change.
 
 ## Install
 

frontend/src/i18n/locales/ar.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "تعذّر تحميل إعداد الجهاز",

frontend/src/i18n/locales/de.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Geräteeinstellung konnte nicht geladen werden",

frontend/src/i18n/locales/en.json

@@ -934,6 +934,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Failed to load device setting",

frontend/src/i18n/locales/es.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "No se pudo cargar el ajuste del dispositivo",

frontend/src/i18n/locales/fr.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Impossible de charger le réglage du périphérique",

frontend/src/i18n/locales/hi.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "डिवाइस सेटिंग लोड नहीं हो सकी",

frontend/src/i18n/locales/id.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Gagal memuat pengaturan perangkat",

frontend/src/i18n/locales/it.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Impossibile caricare l'impostazione del dispositivo",

frontend/src/i18n/locales/ja.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "デバイス設定を読み込めませんでした",

frontend/src/i18n/locales/ko.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "장치 설정을 불러오지 못했습니다",

frontend/src/i18n/locales/nl.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Apparaatinstelling kon niet worden geladen",

frontend/src/i18n/locales/pl.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Nie udało się wczytać ustawienia urządzenia",

frontend/src/i18n/locales/pt.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Falha ao carregar a configuração do dispositivo",

frontend/src/i18n/locales/ru.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Не удалось загрузить настройку устройства",

frontend/src/i18n/locales/sv.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Kunde inte läsa in enhetsinställningen",

frontend/src/i18n/locales/th.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "โหลดการตั้งค่าอุปกรณ์ไม่สำเร็จ",

frontend/src/i18n/locales/tr.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Aygıt ayarı yüklenemedi",

frontend/src/i18n/locales/uk.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Не вдалося завантажити налаштування пристрою",

frontend/src/i18n/locales/vi.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "Không tải được cài đặt thiết bị",

frontend/src/i18n/locales/zh-CN.json

@@ -586,6 +586,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "无法加载设备设置",

frontend/src/i18n/locales/zh-TW.json

@@ -334,6 +334,7 @@
     "device_family_cuda": "NVIDIA GPU (CUDA)",
     "device_family_rocm": "AMD GPU (ROCm)",
     "device_family_xpu": "Intel GPU (XPU)",
+    "device_family_npu": "NPU",
     "device_family_mps": "Apple GPU (MPS)",
     "device_family_cpu": "CPU",
     "device_load_failed": "無法載入裝置設定",

tests/backend/api/test_engines_route_shape.py

@@ -82,7 +82,7 @@ _REQUIRED_KEYS = {
     "effective_device", "routing_status", "routing_reason",
 }
 _TTS_ASR_STATUSES = {"accelerated", "cpu_fallback", "cpu_only", "unavailable"}
-_VALID_FAMILIES = {"cuda", "rocm", "mps", "xpu", "cpu"}
+_VALID_FAMILIES = {"cuda", "rocm", "mps", "xpu", "npu", "cpu"}
 
 
 def test_engines_response_includes_new_fields(fresh_app):

tests/backend/test_compute_device_settings.py

@@ -109,3 +109,14 @@ def test_env_pin_is_reported_and_wins(fresh_app, monkeypatch):
     r = c.put("/api/settings/compute-device", json={"value": "auto"})
     assert r.status_code == 200
     assert r.json()["value"] == "cpu"
+
+
+def test_auto_summary_reports_registered_npu(fresh_app, monkeypatch):
+    from core import device_caps
+    monkeypatch.setattr(device_caps, 'detect_host_caps', lambda: device_caps.HostCaps(
+        family='npu', available_families=('npu', 'cpu'),
+    ))
+    body = _client(fresh_app).get('/api/settings/compute-device').json()
+    assert body['auto_family'] == body['effective_family'] == 'npu'
+    # Detection does not globally opt unrelated engines into NPU execution.
+    assert 'npu' not in body['choices']

tests/test_asr_gpu_compat.py

@@ -34,7 +34,7 @@ _EXPECTED = {
 # is_available — e.g. parakeet-mlx gates on Apple Silicon via mlx_supported()).
 _GPU_ONLY: set[str] = {"parakeet-mlx"}
 
-_VALID = {"cuda", "rocm", "mps", "xpu", "cpu"}
+_VALID = {"cuda", "rocm", "mps", "xpu", "npu", "cpu"}
 
 
 def _cls(engine_id):

tests/test_device_caps.py

@@ -235,3 +235,24 @@ def test_result_is_cached_until_refresh():
 def teardown_module(_module):
     # Drop any cached mock-derived result so other test modules re-probe clean.
     device_caps.detect_host_caps.cache_clear()
+
+
+def test_builtin_xpu_does_not_require_ipex():
+    caps = _probe_with({'torch': _torch_mock(xpu_available=True), 'intel_extension_for_pytorch': None})
+    assert caps.family == 'xpu'
+
+
+def test_registered_npu_is_reported_without_importing_vendor_packages():
+    torch = _torch_mock()
+    torch.npu = types.SimpleNamespace(is_available=lambda: True, get_device_name=lambda i: 'Ascend')
+    caps = _probe_with({'torch': torch, 'intel_extension_for_pytorch': None})
+    assert caps.family == 'npu'
+    assert caps.available_families == ('npu', 'cpu')
+    assert caps.device_name == 'Ascend'
+
+
+def test_unavailable_npu_does_not_claim_acceleration():
+    torch = _torch_mock()
+    torch.npu = types.SimpleNamespace(is_available=lambda: False)
+    caps = _probe_with({'torch': torch, 'intel_extension_for_pytorch': None})
+    assert caps.family == 'cpu'

tests/test_engine_routing.py

@@ -132,7 +132,7 @@ def test_empty_compat_is_defensive_cpu_only():
 
 # ── Contract guarantees ───────────────────────────────────────────────────
 def test_never_emits_n_a():
-    for fam in ("cuda", "rocm", "mps", "xpu", "cpu"):
+    for fam in ("cuda", "rocm", "mps", "xpu", "npu", "cpu"):
         for compat in ((), ("cpu",), ("cuda",), ("cuda", "cpu"), ("mps", "cpu")):
             assert resolve_routing(compat, _caps(fam))["routing_status"] != "n/a"
 

tests/test_moss_tts_v15.py

@@ -106,12 +106,12 @@ def test_audited_custom_remote_code_requires_both_opt_ins(monkeypatch):
 # ── hardware honesty (cross-platform rule) ─────────────────────────────────
 
 
-def test_gpu_compat_cuda_cpu_no_mps():
+def test_gpu_compat_matches_accelerator_paths_without_mps():
     """MPS is undocumented/untested upstream — we must not claim it."""
     from engines.moss_tts_v15 import MossTTSV15Backend
 
-    assert MossTTSV15Backend.gpu_compat == ("cuda", "cpu"), (
+    assert MossTTSV15Backend.gpu_compat == ("cuda", "rocm", "xpu", "npu", "cpu"), (
-        f"expected ('cuda', 'cpu'), got {MossTTSV15Backend.gpu_compat!r}"
+        f"unexpected device targets: {MossTTSV15Backend.gpu_compat!r}"
     )
 
 
@@ -199,3 +199,41 @@ def test_generate_without_ref_audio_omits_reference(monkeypatch):
     MossTTSV15Backend().generate("just text")
     assert "ref_audio" not in captured
     assert "tokens" not in captured
+
+
+@pytest.mark.parametrize('family', [None, 'cuda', 'xpu', 'npu', 'mps'])
+def test_loader_device_matches_routing(monkeypatch, family):
+    """Exercise model + tokenizer placement without importing optional weights."""
+    import io
+    from types import SimpleNamespace
+    from unittest.mock import Mock
+    from engines.moss_tts_v15 import main, MossTTSV15Backend
+    from core.device_caps import HostCaps
+    from services.engine_routing import resolve_routing
+
+    expected = family if family not in (None, 'mps') else 'cpu'
+    accelerator = Mock(return_value=SimpleNamespace(type=family) if family else None)
+    monkeypatch.setitem(sys.modules, 'torch', SimpleNamespace(
+        accelerator=SimpleNamespace(current_accelerator=accelerator),
+        bfloat16='bf16', float32='fp32',
+    ))
+    processor = Mock()
+    processor.model_config.sampling_rate = 24000
+    tokenizer = processor.audio_tokenizer
+    model = Mock()
+    model.to.return_value = model
+    factory = Mock()
+    factory.from_pretrained.return_value = model
+    monkeypatch.setitem(sys.modules, 'transformers', SimpleNamespace(
+        AutoModel=factory,
+        AutoProcessor=SimpleNamespace(from_pretrained=lambda *a, **kw: processor),
+    ))
+    monkeypatch.setattr(main, '_state', None)
+    monkeypatch.setattr(main, '_model_source', lambda: ('local-fixture', 'a' * 40))
+    state = main._load_model(io.BytesIO())
+    accelerator.assert_called_once_with(check_available=True)
+    model.to.assert_called_once_with(expected)
+    tokenizer.to.assert_called_once_with(expected)
+    assert factory.from_pretrained.call_args.kwargs['torch_dtype'] == ('fp32' if expected == 'cpu' else 'bf16')
+    caps = HostCaps(family=family or 'cpu', available_families=(family, 'cpu') if family else ('cpu',))
+    assert resolve_routing(MossTTSV15Backend.gpu_compat, caps)['effective_device'] == state[2]

tests/test_setup_preflight.py

@@ -145,7 +145,7 @@ def test_preflight_device_summary(client):
     assert d["gpu_backend"] in {"cuda", "rocm", "mps", "cpu"}
     assert d["gpu_vendor"] in {"nvidia", "amd", "apple", "intel", "unknown", "none"}
     # #21: canonical-probe family + VRAM joined the device summary.
-    assert d["gpu_family"] in {"cuda", "rocm", "mps", "xpu", "cpu"}
+    assert d["gpu_family"] in {"cuda", "rocm", "mps", "xpu", "npu", "cpu"}
     assert isinstance(d["vram_gb"], (int, float))