import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import spaces

# Model config
MODEL_ID = "WeiboAI/VibeThinker-1.5B"
SYSTEM_PROMPT = "You are a concise solver. Respond briefly."

# Load model
print("Loading model...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto",
)
print("Model loaded!")

@spaces.GPU
def chat_with_stream(message, history, progress=gr.Progress()):
    """Chat with streaming output"""
    
    # Handle inputs safely
    if message is None:
        message = "Hello"
    if history is None:
        history = []
    
    # Convert to string
    message = str(message)
    
    progress(0.1, desc="Building conversation...")
    
    # Build messages
    messages = [{"role": "system", "content": SYSTEM_PROMPT}]
    
    # Add history
    for user_msg, assistant_msg in history:
        if user_msg is not None:
            messages.append({"role": "user", "content": str(user_msg)})
        if assistant_msg is not None:
            messages.append({"role": "assistant", "content": str(assistant_msg)})
    
    progress(0.3, desc="Adding your message...")
    messages.append({"role": "user", "content": message})
    
    progress(0.5, desc="Formatting input...")
    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    
    progress(0.6, desc="Tokenizing...")
    inputs = tokenizer([prompt], return_tensors="pt").to(model.device)
    
    progress(0.7, desc="Starting generation...")
    
    # Generate with streaming
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=1000,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
            pad_token_id=tokenizer.eos_token_id,
            return_dict_in_generate=True,
            output_scores=False,
        )
    
    progress(0.9, desc="Decoding response...")
    
    # Decode
    full_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # Extract assistant response
    if "assistant" in full_text:
        response = full_text.split("assistant")[-1].strip()
    else:
        response = full_text
    
    progress(1.0, desc="Complete!")
    return response

def create_demo():
    """Create simple demo"""
    demo = gr.ChatInterface(
        fn=chat_with_stream,
        title="VibeThinker Chat",
        description="Simple chat with VibeThinker-1.5B",
        examples=["2+2", "What is AI?", "Write a poem"]
    )
    return demo

if __name__ == "__main__":
    print("Starting...")
    demo = create_demo()
    demo.launch(share=False)