Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""Folders audio.cpp v0.9.2 runs that Unsloth Studio now offers: their name hints and the language
each transcription request carries."""
import sys
from pathlib import Path
import pytest
from core.inference import audio_cpp_models as acm
from core.inference import stt_audiocpp_sidecar as stt
from core.inference.audio_cpp_models import AUDIO_CPP_REPO
sys.path.insert(0, str(Path(__file__).resolve().parent))
from test_audio_cpp_models import _gguf_bytes, _put, _snapshot, hub # noqa: E402, F401
from test_stt_audiocpp_details import ABSENT, fake, side, source # noqa: E402, F401
def _names(folder, gguf = None):
return (f"{AUDIO_CPP_REPO}/{folder}", folder, *([gguf] if gguf else []))
@pytest.mark.parametrize(
("names", "family"),
[
# Hub names on audio-cpp/audio.cpp-gguf; Voxtral TTS speaks, it is not the realtime ASR.
(_names("Voxtral-4B-TTS-2603-GGUF"), None),
(_names("Voxtral-4B-TTS-2603-GGUF", "voxtral-4b-tts-2603-q8_0.gguf"), None),
(_names("Voxtral-4B-TTS-2603-GGUF", "voxtral-4b-tts-2603-bf16.gguf"), None),
(_names("Voxtral-Mini-4B-Realtime-2602-GGUF"), "voxtral_realtime"),
(
_names("Voxtral-Mini-4B-Realtime-2602-GGUF", "voxtral-mini-4b-realtime-2602-q4_k.gguf"),
"voxtral_realtime",
),
(_names("KittenTTS-GGUF"), "kitten_tts"),
(("dignome/kitten_tts2", "kitten-tts2", "kitten-tts2-native-q8-multilingual.gguf"), None),
# A version dot is dropped before matching: IndexTTS2.5 is still index_tts2.
(_names("IndexTTS2.5-GGUF"), "index_tts2"),
(_names("IndexTTS2.5-GGUF", "index-tts2_5-q8_0.gguf"), "index_tts2"),
],
)
def test_name_hints(names, family):
assert acm.family_from_names(names) == family
OWSM = f"{AUDIO_CPP_REPO}/OWSM-GGUF"
HVISKE_V6 = f"{AUDIO_CPP_REPO}/Hviske-v6-GGUF"
def _asr_spec(*options):
return {
"schema_version": 1,
"tasks": ["asr"],
"options": {"request": [{"name": n, "type": "string"} for n in options]},
}
@pytest.fixture
def v092_asr(fake, hub, monkeypatch):
monkeypatch.setattr(acm, "runtime_knows_family", lambda family: True)
snap = _snapshot(hub)
_put(
snap,
"OWSM-GGUF/owsm-v4-base-102m-q8_0.gguf",
_gguf_bytes(family = "owsm", spec = _asr_spec("language", "max_tokens")),
)
_put(
snap,
"OWSM-CTC-GGUF/owsm-ctc-v4-1b-q8_0.gguf",
_gguf_bytes(family = "owsm_ctc", spec = _asr_spec("language", "max_tokens")),
)
_put(
snap,
"Hviske-v6-GGUF/hviske-v6-q8_0.gguf",
_gguf_bytes(family = "hviske_asr_v6", spec = _asr_spec("max_tokens", "num_beams")),
)
acm.forget()
with acm._resolve_lock:
acm._downloaded_cache.clear()
return fake
@pytest.mark.parametrize("language", ["fr", "da-DK", None])
def test_owsm_detects_the_language_rather_than_a_two_letter_code(v092_asr, side, source, language):
result = side.transcribe_path(source, OWSM, language)
(body,) = v092_asr.bodies
assert body["language"] == "auto"
# The language the model reports hearing (the fake server answers English), not the ignored choice.
assert result["language"] == "English"
@pytest.mark.parametrize("language, reported", [("fr", "fr"), (None, None)])
def test_owsm_ctc_reports_the_choice_rather_than_an_echoed_auto(
v092_asr, side, source, language, reported
):
result = side.transcribe_path(source, f"{AUDIO_CPP_REPO}/OWSM-CTC-GGUF", language)
(body,) = v092_asr.bodies
assert body["language"] == "auto"
# OWSM-CTC detects nothing and echoes "auto"; that is never reported as a language.
assert result["language"] == reported
@pytest.mark.parametrize("family", ["owsm", "owsm_ctc"])
def test_owsm_families_always_read_auto(family):
assert stt._runtime_language(family, "fr") == "auto"
assert stt._runtime_language(family, None) == "auto"
def test_a_spec_without_a_language_option_gets_none(v092_asr, side, source):
assert acm.resolve(HVISKE_V6, network = False).request_keys == {"max_tokens", "num_beams"}
side.transcribe_path(source, HVISKE_V6, "da")
(body,) = v092_asr.bodies
assert body.get("language", ABSENT) is ABSENT