app.py

import gradio as gr
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

# Public 4-bit base (no login) + your LoRA
base_name = "unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit"
lora_name = "Quanttum/crypto-llama-lora-final"

tokenizer = AutoTokenizer.from_pretrained(base_name)
base_model = AutoModelForCausalLM.from_pretrained(
    base_name,
    device_map="auto",           # auto offload = works on T4 and CPU
    torch_dtype=torch.float16,
    load_in_4bit=True,           # 4-bit is already baked into the model → don't override
)

model = PeftModel.from_pretrained(base_model, lora_name)

pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    max_new_tokens=512,
    temperature=0.3,
    repetition_penalty=1.15,
)

def chat(message, history):
    prompt = f"""You are an expert crypto news summarizer and sentiment analyzer.
Summarize the following news/tweet in 2-4 sentences (TL;DR) and then give a sentiment score from -1.0 (extremely negative) to +1.0 (extremely positive).

News/Tweet:
{message}

TL;DR:"""

    output = pipe(prompt)[0]["generated_text"]
    response = output[len(prompt):].strip()
    return response

# Fixed Gradio 6.0 syntax
with gr.Blocks() as demo:
    gr.Markdown("# Crypto News Summarizer + Sentiment Analyzer\nFine-tuned Llama 3.1 8B by @Quanttum")
    
    gr.ChatInterface(
        chat,
        examples=[
            ["Bitcoin ETF inflows hit $1.5B this week!"],
            ["SEC approves spot Ethereum ETF!"],
            ["Mt. Gox starts repaying creditors in BTC."],
            ["China just banned all crypto trading again."],
        ],
        cache_examples=True,
    )

demo.launch()