import gradio as gr
from unsloth import FastLanguageModel

print("Loading Gemma...")
model, tokenizer = FastLanguageModel.from_pretrained(
    "./gemma_xauusd", max_seq_length=2048, load_in_4bit=True
)
FastLanguageModel.for_inference(model)
print("Gemma ready!")

def predict(price, sma50, sma200, rsi, atr, returns):
    trend = "uptrend" if price > sma200 else "downtrend"
    momentum = "bullish" if price > sma50 else "bearish"
    
    prompt = f"""Analyze XAU/USD: Price ${price:.2f}, Trend: {trend}, Momentum: {momentum}, RSI: {rsi:.1f}. Direction?"""
    
    inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=50, temperature=0.3)
    result = tokenizer.decode(outputs[0])
    
    if "BULLISH" in result.upper():
        return {"sentiment": "BULLISH", "signal": 1.0, "confidence": 0.8}
    elif "BEARISH" in result.upper():
        return {"sentiment": "BEARISH", "signal": -1.0, "confidence": 0.8}
    else:
        return {"sentiment": "NEUTRAL", "signal": 0.0, "confidence": 0.5}

demo = gr.Interface(
    fn=predict,
    inputs=[
        gr.Number(label="Price", value=2650),
        gr.Number(label="SMA 50", value=2640),
        gr.Number(label="SMA 200", value=2620),
        gr.Number(label="RSI", value=65),
        gr.Number(label="ATR", value=12),
        gr.Number(label="Returns %", value=0.5),
    ],
    outputs=gr.JSON(label="Prediction"),
    title="Gemma XAU/USD Analyzer",
    description="AI-powered market analysis",
    api_name="predict"
)

demo.launch()