import gradio as gr
import torch
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor, BitsAndBytesConfig
from peft import PeftModel
from qwen_vl_utils import process_vision_info
from PIL import Image
import os

# --- AYARLAR ---
CLASS_DESCRIPTIONS = {
    "AnnualCrop": "🌾 Yıllık Tarım Alanı",
    "Forest": "🌲 Orman",
    "HerbaceousVegetation": "🌿 Otsu Bitki Örtüsü",
    "Highway": "🛣️ Otoyol",
    "Industrial": "🏭 Endüstriyel Alan",
    "Pasture": "🐄 Mera/Otlak",
    "PermanentCrop": "🍇 Kalıcı Tarım (Bağ/Bahçe)",
    "Residential": "🏘️ Yerleşim Alanı",
    "River": "🌊 Nehir",
    "SeaLake": "🏞️ Deniz/Göl"
}

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"🚀 Kullanılan cihaz: {DEVICE}")

# --- MODEL YÜKLEME ---
def load_model():
    print("⏳ Model yükleniyor...")
    try:
        model_id = "Qwen/Qwen2-VL-2B-Instruct"
        adapter_id = "tugrulkaya/GeoQwen-VL-2B-EuroSAT"
        
        if DEVICE == "cuda":
            # GPU Varsa: 4-bit quantization
            bnb_config = BitsAndBytesConfig(
                load_in_4bit=True,
                bnb_4bit_quant_type="nf4",
                bnb_4bit_compute_dtype=torch.float16
            )
            base_model = Qwen2VLForConditionalGeneration.from_pretrained(
                model_id,
                quantization_config=bnb_config,
                device_map="auto",
                trust_remote_code=True,
                _attn_implementation="flash_attention_2"
            )
        else:
            # CPU (Hugging Face Spaces Free Tier)
            # Offload ve device_map="auto" KAL