import gradio as gr
import torch
import numpy as np
from chronos import Chronos2Pipeline

# --- MODEL AYARLARI ---
MODEL_NAME = "amazon/chronos-2" 

print(f"🚀 {MODEL_NAME} Modeli Yükleniyor...")

try:
    pipeline = Chronos2Pipeline.from_pretrained(
        MODEL_NAME,
        device_map="cpu",
        torch_dtype=torch.float32,
    )
    print("✅ Model Başarıyla Yüklendi!")
except Exception as e:
    print(f"❌ Model Yükleme Hatası: {e}")
    pipeline = None

def predict(context_str, prediction_length):
    if pipeline is None:
        return "Error: Model yüklenemedi."

    try:
        # 1. Veriyi Temizle
        clean_s = context_str.strip()
        if not clean_s: return "Error: Veri boş."
            
        data_list = [float(x) for x in clean_s.split(',')]
        
        # 2. Tensor Oluştur (Batch, Dim, Time) -> (1, 1, Zaman)
        # 🔥 DÜZELTME BURADA YAPILDI (Çift unsqueeze) 🔥
        context_tensor = torch.tensor(data_list).unsqueeze(0).unsqueeze(0)
        
        # 3. Tahmin Yap
        prediction_length = int(prediction_length)
        forecast = pipeline.predict(context_tensor, prediction_length)
        
        # 4. İstatistikleri Al
        median_price = forecast[0].quantile(0.5).item()
        low_bound = forecast[0].quantile(0.1).item()
        high_bound = forecast[0].quantile(0.9).item()
        
        # 5. Formatla
        return f"{median_price}|{low_bound}|{high_bound}"

    except Exception as e:
        return f"Error: {str(e)}"

# SSR Mode False yaparak 404 hatalarını engelliyoruz
iface = gr.Interface(fn=predict, inputs=["text", "number"], outputs="text")
iface.launch(server_name="0.0.0.0", server_port=7860, ssr_mode=False)