debpalashandClaude Fable 5
63fd497caf
Only the TTS model (~2.4 GB) is required on first run; ASR models are per-platform curated picks (curated_on in models.yaml) installed on demand. Every transcription surface returns a typed asr_model_missing error with a one-click download CTA instead of silently pulling multi-GB Whisper weights. Settings -> Models is a grouped, platform-aware catalog. New guided permissions UX (wizard System Check + Settings -> Permissions + mic pre-flight) with native mic-state checks and OS settings deep-links. New parakeet-mlx engine brings Parakeet TDT v3 to Apple Silicon (language-gated capture preference so multilingual dictation never regresses). Docs: expressive-speech page, Flush/Unload + CPU-fallback triage, clone-length FAQ. Hardening: preflight fails open for custom model pins, ROCm curation no longer inherits NVIDIA picks, Windows mic probe reads the NonPackaged consent key, CaptureWidget setup race fixed, offline-cache CI simulation fixes so empty-cache runners stay green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
90 lines
3.4 KiB
Python
90 lines
3.4 KiB
Python
"""ASR backends now carry an explicit ``gpu_compat`` (mirroring TTSBackend) so
|
||
engine_routing can surface the effective device per host. Verifies the ABC
|
||
default and every subclass's declared tuple, plus the IndexTTS2 fix.
|
||
|
||
Backend classes are resolved at RUNTIME (inside each test, via the registry)
|
||
rather than imported at module scope — other suites purge ``services.*`` from
|
||
``sys.modules`` for DB isolation, which would otherwise leave this module
|
||
holding stale class objects depending on collection/run order.
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import pytest
|
||
|
||
# id → declared gpu_compat. nemo-parakeet was CUDA-gated until 2026-07-02,
|
||
# when parakeet-tdt-0.6b-v3 was measured at RTF 0.08–0.23 on an M2 CPU —
|
||
# every ASR engine now has a cpu path.
|
||
_EXPECTED = {
|
||
"whisperx": ("cuda", "cpu"),
|
||
"faster-whisper": ("cuda", "cpu"),
|
||
"mlx-whisper": ("mps", "cpu"),
|
||
"pytorch-whisper": ("cuda", "mps", "cpu"),
|
||
"nemo-parakeet": ("cuda", "cpu"),
|
||
# MLX runs on Apple Silicon's unified-memory GPU only; is_available
|
||
# hard-gates on mlx_supported(), so claiming cpu would be false.
|
||
"parakeet-mlx": ("mps",),
|
||
"moonshine": ("cpu",),
|
||
"funasr": ("cuda", "cpu"),
|
||
# Crash-isolated sidecar wraps the same CTranslate2 engine as
|
||
# faster-whisper — same device support (#730 residual B).
|
||
"faster-whisper-isolated": ("cuda", "cpu"),
|
||
}
|
||
|
||
# Engines that legitimately have NO cpu path (hard platform/GPU gate in
|
||
# 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"}
|
||
|
||
|
||
def _cls(engine_id):
|
||
from services.asr_backend import _REGISTRY
|
||
return _REGISTRY[engine_id]
|
||
|
||
|
||
def test_abc_default_is_cpu_only():
|
||
from services.asr_backend import ASRBackend
|
||
assert ASRBackend.gpu_compat == ("cpu",)
|
||
|
||
|
||
@pytest.mark.parametrize("engine_id,expected", list(_EXPECTED.items()))
|
||
def test_subclass_gpu_compat(engine_id, expected):
|
||
assert _cls(engine_id).gpu_compat == expected
|
||
|
||
|
||
@pytest.mark.parametrize("engine_id", list(_EXPECTED))
|
||
def test_compat_values_are_valid(engine_id):
|
||
compat = _cls(engine_id).gpu_compat
|
||
assert compat, "gpu_compat must be non-empty"
|
||
assert set(compat) <= _VALID
|
||
# Every engine has a cpu path EXCEPT the known hard-GPU-gated ones.
|
||
if engine_id not in _GPU_ONLY:
|
||
assert "cpu" in compat
|
||
else:
|
||
assert "cpu" not in compat # would be a false claim — is_available gates on CUDA
|
||
|
||
|
||
def test_no_asr_engine_falsely_claims_rocm():
|
||
# ROCm is intentionally unclaimed until verified per engine (see ABC note).
|
||
for engine_id in _EXPECTED:
|
||
assert "rocm" not in _cls(engine_id).gpu_compat
|
||
|
||
|
||
def test_nemo_parakeet_has_no_cuda_gate(monkeypatch):
|
||
"""Regression (CPU un-gating, 2026-07-02): on a CUDA-less host,
|
||
is_available() must never claim a GPU is required — availability is a
|
||
pure nemo_toolkit dependency check now."""
|
||
import torch
|
||
monkeypatch.setattr(torch.cuda, "is_available", lambda: False)
|
||
ok, reason = _cls("nemo-parakeet").is_available()
|
||
assert "NVIDIA GPU" not in reason
|
||
if not ok: # env without nemo_toolkit — the only legitimate blocker
|
||
assert "nemo_toolkit" in reason
|
||
|
||
|
||
def test_indextts2_overrides_cpu_only_default():
|
||
from engines.indextts import IndexTTS2Backend
|
||
assert IndexTTS2Backend.gpu_compat == ("cuda", "cpu")
|
||
# must NOT be the inherited TTSBackend default
|
||
assert IndexTTS2Backend.gpu_compat != ("cpu",)
|