feat: ship Wave 5-6 multi-surface hygiene, optimization, and bug fixes

This commit is contained in:
Yun Chan 2026-11-14 11:51:41 +09:00
parent 01ef3293cf
commit 66b6b45ba9
399 changed files with 17828 additions and 11330 deletions

View file

@ -29,9 +29,10 @@ import logging
import os
import signal
import sys
import tempfile
import threading
import time
from contextlib import asynccontextmanager
from contextlib import asynccontextmanager, suppress
from pathlib import Path
from typing import AsyncGenerator
@ -599,28 +600,40 @@ async def load_model(body: dict) -> JSONResponse: # noqa: ANN001
@app.post("/unload")
async def unload_model(request: Request) -> JSONResponse:
"""보조 자리 모델을 내려 VRAM/RAM 을 돌려준다.
"""보조 자리 또는 기본(primary) 자리 모델을 내려 VRAM/RAM 을 돌려준다.
자막·회의가 끝난 뒤에도 보조 모델(예: large-v3-turbo)이 받아쓰기 모델과 함께 상주하면
GPU 여유가 줄어 Ollama 모델 로드가 느려지거나 OOM 이 난다. 기본(primary) 모델은 받아쓰기
지연을 막기 위해 내리지 않는다.
사이드카 HTTP 프로세스 자체는 UIA 브리지(/uia/focus)가 계속 쓰므로 살려 둔 채,
클라우드 STT 전환·장시간 유휴·자막 종료 시 Whisper CTranslate2 모델 메모리만 회수한다.
Request body: { "slot": "aux" }
Request body: { "slot": "aux" | "primary" | "all" }
"""
global _model, _model_id
body = await request.json()
slot: str = body.get("slot", "aux")
if slot != "aux":
if slot not in ("aux", "primary", "all"):
return JSONResponse(
status_code=400,
content={"status": "error", "message": "only the aux slot can be unloaded"},
content={"status": "error", "message": "slot must be aux, primary, or all"},
)
released: list[str] = []
async with _inference_lock:
released = list(_aux_models)
_aux_models.clear()
if slot in ("primary", "all") and _model is not None:
if _model_id:
released.append(_model_id)
_model_devices.pop(_model_id, None)
_model = None
_model_id = None
if slot in ("aux", "all") and _aux_models:
for aux_id in list(_aux_models):
released.append(aux_id)
_model_devices.pop(aux_id, None)
_aux_models.clear()
if released:
gc.collect()
logger.info("보조 모델 언로드: %s", ", ".join(released))
logger.info("모델 언로드(%s): %s", slot, ", ".join(released))
return JSONResponse(content={"status": "unloaded", "models": released})
@ -708,14 +721,21 @@ async def transcribe(
start_time = time.monotonic()
try:
pcm_bytes = await audio.read()
try:
pcm_bytes = await audio.read()
finally:
await audio.close()
if len(pcm_bytes) == 0:
if len(pcm_bytes) < 2:
return JSONResponse(
status_code=400,
content={"status": "error", "message": "오디오 데이터가 비어있습니다"},
)
if len(pcm_bytes) % 2 != 0:
logger.warning("홀수 바이트 PCM 입력(%dB) — 마지막 1바이트를 버립니다", len(pcm_bytes))
pcm_bytes = pcm_bytes[:-1]
audio_array = (
np.frombuffer(pcm_bytes, dtype=np.int16).astype(np.float32) / 32768.0
)
@ -756,6 +776,8 @@ async def transcribe(
finally:
cancel.set()
watcher.cancel()
with suppress(asyncio.CancelledError, Exception):
await watcher
processing_time = int((time.monotonic() - start_time) * 1000)
@ -872,11 +894,134 @@ async def uia_focus(timeoutMs: int = 1500) -> JSONResponse:
return JSONResponse(content=snapshot)
def _run_pyannote_diarization(
audio_bytes: bytes,
hf_token: str,
num_speakers: int | None,
) -> dict[str, object]:
"""선택 설치된 pyannote.audio 파이프라인으로 화자 구간을 추출한다."""
try:
from pyannote.audio import Pipeline # type: ignore[import-not-found]
except Exception as exc:
raise ModuleNotFoundError(str(exc)) from exc
fd, temp_path_str = tempfile.mkstemp(suffix=".wav", prefix="d3ro-diarize-")
temp_path = Path(temp_path_str)
try:
with os.fdopen(fd, "wb") as handle:
handle.write(audio_bytes)
try:
pipeline = Pipeline.from_pretrained(
"pyannote/speaker-diarization-3.1",
token=hf_token,
)
except TypeError:
pipeline = Pipeline.from_pretrained(
"pyannote/speaker-diarization-3.1",
use_auth_token=hf_token,
)
call_kwargs: dict[str, object] = {}
if num_speakers is not None and num_speakers > 0:
call_kwargs["num_speakers"] = num_speakers
annotation = pipeline(str(temp_path), **call_kwargs)
segments: list[dict[str, object]] = []
speakers: set[str] = set()
for turn, _track, speaker in annotation.itertracks(yield_label=True):
speaker_label = str(speaker)
speakers.add(speaker_label)
segments.append(
{
"speaker": speaker_label,
"start": round(float(turn.start), 3),
"end": round(float(turn.end), 3),
}
)
return {
"status": "ok",
"segments": segments,
"num_speakers": len(speakers),
}
finally:
try:
temp_path.unlink(missing_ok=True)
except OSError:
pass
@app.post("/diarize")
async def diarize(
audio: UploadFile = File(...),
hf_token: str = Form(""),
num_speakers: int | None = Form(None),
) -> JSONResponse:
"""pyannote.audio 기반 화자 분리 엔드포인트 (미설치 시 503 반환으로 LLM 폴백 유도)."""
token = hf_token.strip()
if not token:
return JSONResponse(
status_code=400,
content={
"status": "error",
"code": "missing_hf_token",
"message": "Hugging Face 토큰이 비어 있습니다",
},
)
try:
audio_bytes = await audio.read()
finally:
await audio.close()
if not audio_bytes:
return JSONResponse(
status_code=400,
content={
"status": "error",
"code": "empty_audio",
"message": "오디오 파일이 비어 있습니다",
},
)
async with _inference_lock:
try:
result = await asyncio.to_thread(
_run_pyannote_diarization,
audio_bytes,
token,
num_speakers,
)
return JSONResponse(content=result)
except ModuleNotFoundError:
return JSONResponse(
status_code=503,
content={
"status": "unavailable",
"code": "pyannote_not_installed",
"message": "pyannote.audio 패키지가 설치되어 있지 않습니다",
},
)
except Exception as exc:
logger.warning("pyannote 화자 분리 실패: %s", exc)
return JSONResponse(
status_code=500,
content={
"status": "error",
"code": "diarization_failed",
"message": f"화자 분리 실행 실패: {exc}",
},
)
@app.post("/shutdown")
async def shutdown() -> JSONResponse:
"""서버를 graceful하게 종료한다."""
logger.info("종료 요청 수신")
if _download_thread is not None and _download_thread.is_alive():
_download_cancel.set()
await asyncio.to_thread(_download_thread.join, 2.0)
try:
import uia_bridge