app.py

#!/usr/bin/env python3
"""
Ultralekki entrypoint dla HF Spaces (Docker SDK)
✅ Pobiera model → uruchamia llama_cpp.server
✅ OpenAI format | ✅ Brak auth | ✅ Odporny na zerwania
"""
import os
import sys
import signal
import logging
import subprocess
from huggingface_hub import hf_hub_download

logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
logger = logging.getLogger(__name__)

MODEL_REPO = "unsloth/granite-4.1-3b-GGUF"
MODEL_FILE = os.environ.get("MODEL_FILE", "granite-4.1-3b-UD-IQ2_M.gguf")
PORT = os.environ.get("PORT", "7860")
N_CTX = os.environ.get("N_CTX", "2048")
N_THREADS = os.environ.get("N_THREADS", "2")
N_BATCH = os.environ.get("N_BATCH", "512")

def graceful_shutdown(signum, frame):
    logger.info("📡 Otrzymano sygnał zakończenia. Zamykanie...")
    sys.exit(0)

signal.signal(signal.SIGTERM, graceful_shutdown)
signal.signal(signal.SIGINT, graceful_shutdown)

if __name__ == "__main__":
    try:
        logger.info(f"⬇️ Pobieranie/weryfikacja: {MODEL_REPO}/{MODEL_FILE}")
        model_path = hf_hub_download(
            repo_id=MODEL_REPO,
            filename=MODEL_FILE,
            resume_download=True
        )
        logger.info(f"✅ Model gotowy: {model_path}")

        # Komenda startowa llama_cpp.server
        cmd = [
            sys.executable, "-m", "llama_cpp.server",
            "--model", model_path,
            "--host", "0.0.0.0",
            "--port", PORT,
            "--n_ctx", N_CTX,
            "--n_threads", N_THREADS,
            "--n_batch", N_BATCH,
            "--n_gpu_layers", "0",
            "--use_mmap",
            "--no_flash_attn",
            "--chat_format", "chatml" # Domyślny format dla Granite/LLama
        ]

        logger.info(f"🚀 Start serwera: {' '.join(cmd)}")
        
        # Uruchom proces podrzędny i czekaj na jego zakończenie
        process = subprocess.Popen(cmd)
        process.wait()
        
    except Exception as e:
        logger.error(f"❌ Krytyczny błąd: {e}", exc_info=True)
        sys.exit(1)