app.py

import os
import sys
import time
import threading
import traceback
from datetime import datetime
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

# === Ortam değişkenleri
os.environ.setdefault("HF_HOME", "/app/.cache")
os.environ.setdefault("HF_HUB_CACHE", "/app/.cache")

# === Zamanlı log fonksiyonu
def log(message):
    timestamp = datetime.now().strftime("%H:%M:%S")
    print(f"[{timestamp}] {message}", flush=True)

# === FastAPI başlat
app = FastAPI()
pipe = None

@app.on_event("startup")
def load_model():
    global pipe
    try:
        model_name = "ytu-ce-cosmos/Turkish-Llama-8b-Instruct-v0.1"
        log(f"⬇️ Model yükleme başlatılıyor: {model_name}")

        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForCausalLM.from_pretrained(
            model_name,
            device_map="auto",
            torch_dtype="auto"
        )

        pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device_map="auto")
        log("✅ Model ve pipeline başarıyla hazır.")

    except Exception as e:
        log(f"❌ Model yükleme hatası: {e}")
        traceback.print_exc()
        raise

class UserInputRequest(BaseModel):
    user_input: str
    system_prompt: str

@app.post("/generate")
def generate(req: UserInputRequest):
    try:
        start_time = time.time()
        log(f"💬 Kullanıcı isteği alındı: {req.user_input}")

        # Kısa ve net system prompt, okunabilir bölünmüş
        concise_system_prompt = (
            f"{req.system_prompt}\n"
            "❗ Cevaplarını sadece aşağıdaki formatta döndür, fazladan açıklama yazma, örnek ekleme:\n"
            "#ANSWER: <cevap>\n"
            "#INTENT: <intent>\n"
            "#PARAMS: {...}\n"
            "#MISSING: [...]\n"
            "#ACTION_JSON: {...}\n"
            "Şimdi sadece kullanıcının sorusunu bekliyorsun ve formatlı cevap veriyorsun."
        )

        # Role separation: System, User, Assistant blokları
        full_prompt = (
            f"### System:\n{concise_system_prompt}\n\n"
            f"### User:\n{req.user_input}\n\n"
            f"### Assistant:"
        )

        result = pipe(
            full_prompt,
            max_new_tokens=200,
            temperature=0.0,
            top_p=1.0,
            repetition_penalty=1.0,
            do_sample=False
        )
        answer = result[0]["generated_text"]

        end_time = time.time()
        elapsed = end_time - start_time
        log(f"✅ Yanıt süresi: {elapsed:.2f} saniye")

        return {"response": answer}

    except Exception as e:
        log(f"❌ /generate hatası: {e}")
        traceback.print_exc()
        raise HTTPException(status_code=500, detail=str(e))

@app.get("/")
def health():
    return {"status": "ok"}

def run_health_server():
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)

threading.Thread(target=run_health_server, daemon=True).start()

log("⏸️ Uygulama bekleme modunda...")
while True:
    time.sleep(60)