Effi/effi-stt/main.py

# Effi-STT - Speech-to-Text pour Effi
# API OpenAI-compatible avec faster-whisper
# Inclut déchargement automatique après inactivité

from fastapi import FastAPI, File, UploadFile, Form, HTTPException
from fastapi.responses import JSONResponse
import uvicorn
from faster_whisper import WhisperModel
import tempfile
import os
from typing import Optional
import logging
import threading
import time

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

app = FastAPI(
    title="Effi-STT",
    description="Speech-to-Text API compatible OpenAI pour Effi",
    version="1.1.0"
)

MODEL_SIZE = os.getenv("MODEL_SIZE", "distil-large-v3")
DEVICE = os.getenv("DEVICE", "cpu")
COMPUTE_TYPE = os.getenv("COMPUTE_TYPE", "int8")
LANGUAGE = os.getenv("LANGUAGE", "fr")
INACTIVITY_TIMEOUT = int(os.getenv("INACTIVITY_TIMEOUT", "300"))
PORT = int(os.getenv("PORT", "8200"))
HOST = os.getenv("HOST", "0.0.0.0")

model = None
model_loaded = False
model_lock = threading.Lock()
last_request_time = time.time()
unload_timer = None

def load_model():
    global model, model_loaded
    with model_lock:
        if not model_loaded:
            logger.info(f"Chargement du modèle Whisper: {MODEL_SIZE}")
            logger.info(f"   Device: {DEVICE}, Compute type: {COMPUTE_TYPE}")
            model = WhisperModel(
                MODEL_SIZE,
                device=DEVICE,
                compute_type=COMPUTE_TYPE,
                cpu_threads=int(os.getenv("CPU_THREADS", "8"))
            )
            model_loaded = True
            logger.info("✅ Modèle chargé avec succès!")

def unload_model():
    global model, model_loaded
    with model_lock:
        if model_loaded:
            logger.info("🕐 Inactivité détectée - Déchargement du modèle Whisper...")
            del model
            model = None
            model_loaded = False
            import gc
            gc.collect()
            logger.info("✅ Modèle déchargé, mémoire libérée")

def reset_unload_timer():
    global last_request_time, unload_timer
    last_request_time = time.time()
    if unload_timer is not None:
        unload_timer.cancel()
    if model_loaded:
        timer = threading.Timer(INACTIVITY_TIMEOUT, unload_model)
        timer.daemon = True
        timer.start()
        unload_timer = timer
        logger.debug(f"⏱️ Timer déchargement programmé dans {INACTIVITY_TIMEOUT}s")

def check_and_load_model():
    global model_loaded
    if not model_loaded:
        logger.info("📥 Chargement du modèle Whisper à la demande...")
        load_model()
        reset_unload_timer()

@app.on_event("startup")
async def startup():
    logger.info(f"🚀 Démarrage effi-stt v1.1.0 sur http://{HOST}:{PORT}")
    logger.info(f"   Timeout d'inactivité: {INACTIVITY_TIMEOUT}s")
    load_model()
    reset_unload_timer()

@app.on_event("shutdown")
async def shutdown():
    if unload_timer:
        unload_timer.cancel()
    unload_model()

@app.get("/health")
async def health():
    check_and_load_model()
    return {
        "status": "healthy",
        "model": MODEL_SIZE,
        "model_loaded": model_loaded,
        "inactivity_timeout": INACTIVITY_TIMEOUT
    }

@app.post("/v1/audio/transcriptions")
async def transcribe(
    file: UploadFile = File(..., description="Fichier audio à transcrire"),
    model_param: Optional[str] = Form(None, alias="model"),
    language: Optional[str] = Form(LANGUAGE, description="Code langue (fr, en, auto)"),
    prompt: Optional[str] = Form(None, description="Prompt optionnel"),
    response_format: Optional[str] = Form("json", description="Format de réponse")
):
    try:
        check_and_load_model()
        reset_unload_timer()
        logger.info(f"📝 Transcription: lang={language}, file={file.filename}")
        
        suffix = os.path.splitext(file.filename)[1] or ".mp3"
        with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
            content = await file.read()
            tmp.write(content)
            tmp_path = tmp.name
            logger.info(f"   Fichier: {tmp_path} ({len(content)} bytes)")
        
        try:
            with model_lock:
                segments, info = model.transcribe(
                    tmp_path,
                    language="fr",
                    task="transcribe",
                    beam_size=5,
                    best_of=5,
                    condition_on_previous_text=True,
                    initial_prompt="Transcription en français. L'utilisateur parle français."
                )
            
            text_parts = []
            for segment in segments:
                text_parts.append(segment.text)
            
            full_text = " ".join(text_parts).strip()
            logger.info(f"✅ Transcription: '{full_text[:50]}...'")
            logger.info(f"   Langue: {info.language} ({info.language_probability:.2f})")
            
            if response_format == "text":
                return full_text
            else:
                return JSONResponse(content={"text": full_text})
                
        finally:
            os.unlink(tmp_path)
            
    except Exception as e:
        logger.error(f"❌ Erreur: {str(e)}")
        raise HTTPException(status_code=500, detail=str(e))

@app.post("/unload")
async def unload():
    unload_model()
    return {"status": "unloaded"}

@app.post("/reload")
async def reload():
    load_model()
    reset_unload_timer()
    return {"status": "loaded"}

@app.get("/v1/models")
async def list_models():
    return {
        "data": [{"id": "whisper-1", "object": "model", "created": 1677610602, "owned_by": "openai", "permission": [], "root": "whisper-1", "parent": None}],
        "object": "list"
    }

@app.get("/stats")
async def stats():
    return {
        "model_loaded": model_loaded,
        "last_request_age_seconds": int(time.time() - last_request_time),
        "inactivity_timeout": INACTIVITY_TIMEOUT
    }

if __name__ == "__main__":
    logger.info(f"🚀 Serveur sur http://{HOST}:{PORT}")
    logger.info(f"   Endpoints:")
    logger.info(f"   - GET  /health")
    logger.info(f"   - POST /v1/audio/transcriptions")
    logger.info(f"   - GET  /v1/models")
    logger.info(f"   - POST /unload (déchargement manuel)")
    logger.info(f"   - POST /reload (rechargement manuel)")
    logger.info(f"   - GET  /stats")
    uvicorn.run(app, host=HOST, port=PORT, log_level="info")