speech/app/asr/workers/faster_whisper_worker.py

from __future__ import annotations

import argparse
import asyncio
import gc
import tempfile
import time
from pathlib import Path
from typing import Any, Dict, List, Optional, Union

import uvicorn
from fastapi import FastAPI, File, Form, HTTPException, UploadFile
from faster_whisper import WhisperModel

import sys
sys.path.insert(0, "/app")
from asr.config import (
    COMPUTE_TYPE,
    CUSTOM_MODEL_DIR,
    DEFAULT_MODEL,
    DEVICE,
    MODEL_CACHE,
    PYANNOTE_HF_TOKEN,
    ensure_runtime_dirs,
    resolve_custom_model_path,
)

import os
# 컨테이너는 983 유저로 실행되는데 HOME(/app)이 root 소유라, pyannote.audio가 쓰는
# numba/matplotlib/torch가 각자 ~/.cache, ~/.config 아래에 쓰려다 실패한다.
os.environ["HOME"] = "/tmp"  # 컨테이너 기본 HOME=/app은 983 유저가 쓰기 불가 (setdefault로는 덮어쓰기 안 됨)
os.environ.setdefault("NUMBA_CACHE_DIR", "/tmp/numba_cache")

app = FastAPI(title="ASR Faster-Whisper Worker")
_MODEL_CACHE: Dict[str, WhisperModel] = {}
_DIARIZATION_PIPELINE: Any = None

# Idle-unload / model-switch eviction: only one whisper model stays resident
# at a time (switching sizes frees the old one), and everything (including
# the diarization pipeline) is dropped after IDLE_UNLOAD_SECONDS of no
# requests, so this backend doesn't permanently hog GPU memory shared with
# the other ASR workers.
IDLE_UNLOAD_SECONDS = 120
_last_used: float = 0.0
_active_requests: int = 0


def _free_gpu() -> None:
    gc.collect()
    try:
        import torch
        if DEVICE == "cuda" and torch.cuda.is_available():
            torch.cuda.empty_cache()
    except Exception:
        pass


def _unload_whisper_models() -> None:
    if not _MODEL_CACHE:
        return
    print(f"[faster_whisper] unloading {list(_MODEL_CACHE.keys())}", flush=True)
    _MODEL_CACHE.clear()
    _free_gpu()


def _unload_diarization() -> None:
    global _DIARIZATION_PIPELINE
    if _DIARIZATION_PIPELINE is None:
        return
    print("[faster_whisper] unloading diarization pipeline", flush=True)
    _DIARIZATION_PIPELINE = None
    _free_gpu()


async def _idle_unload_loop() -> None:
    while True:
        await asyncio.sleep(30)
        if _active_requests == 0 and _last_used and (time.monotonic() - _last_used) >= IDLE_UNLOAD_SECONDS:
            _unload_whisper_models()
            _unload_diarization()


@app.on_event("startup")
async def startup() -> None:
    ensure_runtime_dirs()
    asyncio.create_task(_idle_unload_loop())


@app.get("/health")
def health() -> Dict[str, Any]:
    return {
        "status": "ok",
        "device": DEVICE,
        "compute_type": COMPUTE_TYPE,
        "diarization_available": bool(PYANNOTE_HF_TOKEN),
    }


@app.post("/transcribe")
async def transcribe(
    file: UploadFile = File(...),
    model: str = Form(DEFAULT_MODEL),
    custom_model_path: Optional[str] = Form(None),
    language: Optional[str] = Form(None),
    task: str = Form("transcribe"),
    beam_size: int = Form(5),
    temperature: float = Form(0.0),
    word_timestamps: bool = Form(False),
    diarize: bool = Form(False),
    num_speakers: Optional[int] = Form(None),
    min_speakers: Optional[int] = Form(None),
    max_speakers: Optional[int] = Form(None),
    no_repeat_ngram_size: int = Form(0),
    repetition_penalty: float = Form(1.0),
    compression_ratio_threshold: float = Form(2.4),
    log_prob_threshold: float = Form(-1.0),
    no_speech_threshold: float = Form(0.6),
    condition_on_previous_text: bool = Form(True),
) -> Dict[str, Any]:
    ensure_runtime_dirs()
    suffix = Path(file.filename or "upload.bin").suffix or ".bin"
    with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
        tmp.write(await file.read())
        tmp_path = Path(tmp.name)

    global _active_requests, _last_used
    _active_requests += 1
    try:
        resolved_model = resolve_custom_model_path(custom_model_path) or model
        result = _transcribe(
            audio_path=tmp_path,
            model_name=resolved_model,
            language=language or None,
            task=task,
            beam_size=beam_size,
            temperature=temperature,
            word_timestamps=word_timestamps,
            no_repeat_ngram_size=no_repeat_ngram_size,
            repetition_penalty=repetition_penalty,
            compression_ratio_threshold=compression_ratio_threshold,
            log_prob_threshold=log_prob_threshold,
            no_speech_threshold=no_speech_threshold,
            condition_on_previous_text=condition_on_previous_text,
        )
        if diarize:
            result = _apply_diarization(
                audio_path=tmp_path,
                result=result,
                num_speakers=num_speakers,
                min_speakers=min_speakers,
                max_speakers=max_speakers,
            )
        return result
    except HTTPException:
        raise
    except Exception as exc:
        raise HTTPException(status_code=500, detail=f"Worker failure: {exc}") from exc
    finally:
        tmp_path.unlink(missing_ok=True)
        _free_gpu()
        _active_requests -= 1
        _last_used = time.monotonic()


def _load_model(model_name: str) -> WhisperModel:
    if model_name not in _MODEL_CACHE:
        if _MODEL_CACHE:
            # Only one whisper model resident at a time — switching sizes frees the old one.
            _unload_whisper_models()
        try:
            _MODEL_CACHE[model_name] = WhisperModel(
                model_name,
                device=DEVICE,
                compute_type=COMPUTE_TYPE,
                download_root=str(MODEL_CACHE),
            )
        except Exception as e:
            # A load interrupted partway (e.g. OOM) can leave CUDA memory
            # fragmented/leaked in ways gc.collect()+empty_cache() don't
            # reliably reclaim, and since _MODEL_CACHE never got populated
            # the idle-unload loop has nothing to clean up either.
            # Restarting the whole process is the only guaranteed way to get
            # that memory back — supervisord's autorestart=true respawns it
            # immediately.
            print(f"[faster_whisper] model load failed, restarting process to reclaim GPU memory: {type(e).__name__}: {e}", flush=True)
            os._exit(1)
    return _MODEL_CACHE[model_name]


def _transcribe(
    audio_path: Path,
    model_name: str,
    language: Optional[str],
    task: str,
    beam_size: int,
    temperature: float,
    word_timestamps: bool,
    no_repeat_ngram_size: int,
    repetition_penalty: float,
    compression_ratio_threshold: float,
    log_prob_threshold: float,
    no_speech_threshold: float,
    condition_on_previous_text: bool,
) -> Dict[str, Any]:
    model = _load_model(model_name)

    temperatures: Union[float, List[float]] = [0.0, 0.2, 0.4, 0.6, 0.8, 1.0] if temperature == 0.0 else temperature

    segments_iter, info = model.transcribe(
        str(audio_path),
        language=language,
        task=task,
        beam_size=beam_size,
        temperature=temperatures,
        word_timestamps=word_timestamps,
        no_repeat_ngram_size=no_repeat_ngram_size,
        repetition_penalty=repetition_penalty,
        compression_ratio_threshold=compression_ratio_threshold,
        log_prob_threshold=log_prob_threshold,
        no_speech_threshold=no_speech_threshold,
        condition_on_previous_text=condition_on_previous_text,
    )

    segment_list = []
    full_text = []
    for seg in segments_iter:
        seg_dict: Dict[str, Any] = {
            "id": seg.id,
            "start": round(float(seg.start), 3),
            "end": round(float(seg.end), 3),
            "text": seg.text,
            "avg_logprob": round(float(seg.avg_logprob), 4) if seg.avg_logprob is not None else None,
            "compression_ratio": round(float(seg.compression_ratio), 4) if seg.compression_ratio is not None else None,
            "no_speech_prob": round(float(seg.no_speech_prob), 4) if seg.no_speech_prob is not None else None,
        }
        if word_timestamps and getattr(seg, "words", None):
            seg_dict["words"] = [
                {
                    "start": round(float(w.start), 3),
                    "end": round(float(w.end), 3),
                    "word": w.word,
                    "probability": round(float(w.probability), 4),
                }
                for w in seg.words
            ]
        segment_list.append(seg_dict)
        full_text.append(seg.text.strip())

    duration = None
    try:
        duration = round(float(info.duration), 3)
    except Exception:
        pass

    return {
        "backend": "faster-whisper",
        "model": model_name,
        "language": getattr(info, "language", language),
        "language_probability": round(float(getattr(info, "language_probability", 0) or 0), 4),
        "duration": duration,
        "text": " ".join(x for x in full_text if x),
        "segments": segment_list,
        "diarized": False,
    }


def _get_diarization_pipeline():
    global _DIARIZATION_PIPELINE
    if _DIARIZATION_PIPELINE is None:
        try:
            from pyannote.audio import Pipeline
            import torch
        except ImportError as exc:
            raise RuntimeError("pyannote.audio가 설치되지 않았습니다.") from exc
        if not PYANNOTE_HF_TOKEN:
            raise RuntimeError("화자 분리를 사용하려면 PYANNOTE_HF_TOKEN 환경변수를 설정하세요.")
        _DIARIZATION_PIPELINE = Pipeline.from_pretrained(
            "pyannote/speaker-diarization-3.1",
            token=PYANNOTE_HF_TOKEN,
        )
        if DEVICE == "cuda":
            import torch
            _DIARIZATION_PIPELINE = _DIARIZATION_PIPELINE.to(torch.device("cuda"))
    return _DIARIZATION_PIPELINE


def _apply_diarization(
    audio_path: Path,
    result: Dict[str, Any],
    num_speakers: Optional[int],
    min_speakers: Optional[int],
    max_speakers: Optional[int],
) -> Dict[str, Any]:
    pipeline = _get_diarization_pipeline()
    kwargs: Dict[str, Any] = {}
    if num_speakers is not None:
        kwargs["num_speakers"] = num_speakers
    else:
        if min_speakers is not None:
            kwargs["min_speakers"] = min_speakers
        if max_speakers is not None:
            kwargs["max_speakers"] = max_speakers

    # pyannote.audio 4.x: pipeline() returns a DiarizeOutput dataclass instead of
    # an Annotation directly; the itertracks-capable Annotation is .speaker_diarization.
    diarization = pipeline(str(audio_path), **kwargs).speaker_diarization
    segments = result.get("segments", [])
    for seg in segments:
        seg_start = float(seg.get("start", 0))
        seg_end = float(seg.get("end", 0))
        speaker_times: Dict[str, float] = {}
        for turn, _, speaker in diarization.itertracks(yield_label=True):
            overlap_start = max(seg_start, turn.start)
            overlap_end = min(seg_end, turn.end)
            if overlap_end > overlap_start:
                speaker_times[speaker] = speaker_times.get(speaker, 0.0) + (overlap_end - overlap_start)
        seg["speaker"] = max(speaker_times, key=speaker_times.get) if speaker_times else "UNKNOWN"

    lines = []
    current_speaker: Optional[str] = None
    current_texts: list = []
    for seg in segments:
        speaker = seg.get("speaker", "UNKNOWN")
        text = seg.get("text", "").strip()
        if not text:
            continue
        if speaker != current_speaker:
            if current_texts and current_speaker is not None:
                lines.append(f"[{current_speaker}]: {' '.join(current_texts)}")
            current_speaker = speaker
            current_texts = [text]
        else:
            current_texts.append(text)
    if current_texts and current_speaker is not None:
        lines.append(f"[{current_speaker}]: {' '.join(current_texts)}")

    result["text"] = "\n".join(lines)
    result["diarized"] = True
    return result


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--host", default="0.0.0.0")
    parser.add_argument("--port", type=int, default=8001)
    args = parser.parse_args()
    ensure_runtime_dirs()
    uvicorn.run(app, host=args.host, port=args.port)