Skip to content

VoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages.

3.0 KB
Raw Blame
"""A warm generate must free idle GPU memory first when the box is tight
(#730/#1190).

The cold LOAD path already evicts via _make_room_before_tts_load (inside
_load_model_with_timeout). The warm path (model already resident, get_model
returns early at the cache check) skipped it, so a generate on a VRAM-tight
MPS box contended with capture-ASR and the clone-prompt side cache until it
exceeded the execution budget and was abandoned (#730/#1190).
make_room_before_generate, called from get_model()'s warm-return path, closes
that gap for every native TTS generate; the policy is the one real
reliability-vs-latency knob.
"""
import services.model_manager as mm
from services.model_manager import make_room_before_generate


def _count(monkeypatch):
    """Patch the three eviction primitives and count how often each runs."""
    calls = {"free_vram": 0, "side_caches": 0, "capture_asr": 0}

    monkeypatch.setattr(mm, "free_vram", lambda: calls.__setitem__("free_vram", calls["free_vram"] + 1))
    monkeypatch.setattr(
        mm, "release_tts_side_caches",
        lambda: calls.__setitem__("side_caches", calls["side_caches"] + 1),
    )
    import services.asr_backend as ab
    monkeypatch.setattr(
        ab, "release_idle_capture_backend",
        lambda _idle_s: calls.__setitem__("capture_asr", calls["capture_asr"] + 1),
    )
    return calls


def _ram(monkeypatch, gb):
    import services.memory_budget as mb
    monkeypatch.setattr(mb, "available_memory", lambda: {"ram_available_gb": gb})


def test_never_mode_frees_nothing(monkeypatch):
    monkeypatch.setenv("OMNIVOICE_FREE_VRAM_BEFORE_GENERATE", "never")
    calls = _count(monkeypatch)
    make_room_before_generate()
    assert calls == {"free_vram": 0, "side_caches": 0, "capture_asr": 0}


def test_always_mode_frees_all(monkeypatch):
    monkeypatch.setenv("OMNIVOICE_FREE_VRAM_BEFORE_GENERATE", "always")
    calls = _count(monkeypatch)
    make_room_before_generate()
    assert calls["free_vram"] == 1 and calls["side_caches"] == 1 and calls["capture_asr"] == 1


def test_auto_skips_on_roomy_machine(monkeypatch):
    monkeypatch.delenv("OMNIVOICE_FREE_VRAM_BEFORE_GENERATE", raising=False)
    _ram(monkeypatch, 999.0)  # ample RAM -> a roomy machine pays nothing
    calls = _count(monkeypatch)
    make_room_before_generate()
    assert calls == {"free_vram": 0, "side_caches": 0, "capture_asr": 0}


def test_auto_frees_on_tight_ram(monkeypatch):
    monkeypatch.delenv("OMNIVOICE_FREE_VRAM_BEFORE_GENERATE", raising=False)
    _ram(monkeypatch, 0.5)  # below the 6.0 GB unified headroom
    calls = _count(monkeypatch)
    make_room_before_generate()
    assert calls["free_vram"] == 1 and calls["side_caches"] == 1


def test_default_mode_is_auto(monkeypatch):
    # No env set at all must behave as auto (tight RAM triggers).
    monkeypatch.delenv("OMNIVOICE_FREE_VRAM_BEFORE_GENERATE", raising=False)
    _ram(monkeypatch, 0.5)
    calls = _count(monkeypatch)
    make_room_before_generate()
    assert calls["free_vram"] == 1