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 AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from contextlib import asynccontextmanager

# === 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)

# === Global model değişkenleri
tokenizer = None
model = None

# === Lifespan tanımı
@asynccontextmanager
async def lifespan(app: FastAPI):
    global tokenizer, model
    try:
        model_name = "ytu-ce-cosmos/Turkish-Llama-8b-DPO-v0.1"
        log(f"⬇️ Model yükleme başlatılıyor: {model_name}")

        tokenizer = AutoTokenizer.from_pretrained(model_name)
        tokenizer.pad_token = tokenizer.eos_token

        quant_config = BitsAndBytesConfig(
            load_in_8bit=True,  # ✅ 8-bit quantization (modern yöntem)
            llm_int8_threshold=6.0
        )

        model = AutoModelForCausalLM.from_pretrained(
            model_name,
            device_map="auto",
            quantization_config=quant_config
        )

        log("✅ Model ve tokenizer başarıyla hazır (8-bit quantized, BitsAndBytesConfig).")
        yield  # Uygulama burada çalışır

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

# === FastAPI başlat
app = FastAPI(lifespan=lifespan)

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

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

        # === Apply chat template
        t0 = time.time()
        messages = [
            {"role": "system", "content": req.system_prompt},
            {"role": "user", "content": req.user_input}
        ]
        chat_input = tokenizer.apply_chat_template(
            messages,
            add_generation_prompt=True,
            return_tensors="pt"
        ).to(model.device)
        t1 = time.time()
        log(f"⏱️ apply_chat_template süresi: {t1 - t0:.2f} saniye")

        input_len = chat_input.shape[-1]
        total_ctx = model.config.max_position_embeddings if hasattr(model.config, 'max_position_embeddings') else 4096
        max_new_tokens = min(512, max(1, total_ctx - input_len))

        log(f"ℹ️ Input uzunluğu: {input_len}, max_new_tokens ayarlandı: {max_new_tokens}")

        # === Generate
        t2 = time.time()
        terminators = [
            tokenizer.eos_token_id,
            tokenizer.convert_tokens_to_ids("<|eot_id|>") if "<|eot_id|>" in tokenizer.get_vocab() else tokenizer.eos_token_id
        ]
        outputs = model.generate(
            input_ids=chat_input,
            max_new_tokens=max_new_tokens,
            eos_token_id=terminators
        )
        t3 = time.time()
        log(f"⏱️ generate süresi: {t3 - t2:.2f} saniye")

        # === Decode
        t4 = time.time()
        response = outputs[0][input_len:]
        answer = tokenizer.decode(response, skip_special_tokens=True)
        t5 = time.time()
        log(f"⏱️ decode süresi: {t5 - t4:.2f} saniye")

        overall_end = time.time()
        overall_elapsed = overall_end - overall_start
        log(f"✅ Toplam yanıt süresi: {overall_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)