app.py

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

# Load your custom model
model_path = "alexdev404/gpt2-finetuned-chat"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path)

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

def generate_response(prompt, system_message, conversation_history=None, max_tokens=75, temperature=0.78, top_p=0.85, repetition_penalty=1.031, top_k=55):
    """Generate using your custom training format"""
    
    # Build context using your NEW format
    context = ""
    if conversation_history:
        # Last 2-3 exchanges
        # Use more conversation history to fill GPT-2's context window (1024 tokens)
        # Estimate ~20-30 tokens per exchange, so we can fit ~30-40 exchanges
        recent = conversation_history[-30:] if len(conversation_history) > 30 else conversation_history
        is_first_message = False
        for i, message in enumerate(recent):
            if i == 0:
                is_first_message = True
                context += f"<|start|>User:<|message|>{system_message}<|end|>\n<|start|>Assistant:<|message|>Hey, what's up nice to meet you. I'm glad to be here!<|end|>\n"
            if message['role'] == 'user':
                context += f"<|start|>User:<|message|>{message['content']}<|end|>\n"
            else:
                context += f"<|start|>Assistant:<|message|>{message['content']}<|end|>\n"

    # Format input to match training
    # formatted_input = None
    # if is_first_message:
    #     formatted_input = f"{context}<|start|>User:<|message|>{prompt}<|end|>\n<|start|>Assistant:<|message|>"
    # else:
    formatted_input = f"{context}<|start|>User:<|message|>{prompt}<|end|>\n<|start|>Assistant:<|message|>"

    # Debug: Print the formatted input
    print(f"Formatted input: {repr(formatted_input)}")

    inputs = tokenizer(
        formatted_input, 
        return_tensors="pt", 
        padding=True, 
        truncation=True,
        max_length=512
    )
    
    with torch.no_grad():
        outputs = model.generate(
            inputs.input_ids,
            attention_mask=inputs.attention_mask,
            max_new_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,    # Consider top 55 tokens
            do_sample=True,
            pad_token_id=tokenizer.pad_token_id,
            repetition_penalty=repetition_penalty,
            eos_token_id=tokenizer.encode("<|end|>", add_special_tokens=False)[0]
        )

    # Decode only new tokens
    new_tokens = outputs[0][inputs.input_ids.shape[-1]:]
    response = tokenizer.decode(new_tokens, skip_special_tokens=False)
    
    return response.strip()

def respond(
    message,
    history: list[dict[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
    repetition_penalty,
    top_k,
):
    """
    Modified to use your custom GPT-2 model instead of Hugging Face Inference API
    """
    # Convert gradio history format to your format
    # Gradio history is already in the correct format: [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]
    conversation_history = history  # Use history directly
    
    # Debug: Print the formatted input to see what's being sent to the model
    print(f"User message: {message}")
    print(f"History length: {len(conversation_history)}")
    
    # Generate response using your model
    response = generate_response(
        message, 
        system_message,
        conversation_history, 
        max_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        top_k=top_k
    )
    
    # print(f"Raw response: {repr(response)}")
    
    # Clean up the response
    if "<|end|>" in response:
        response = response.split("<|end|>")[0]
    
    # Remove any remaining special tokens
    # response = response.replace("<|start|>", "")
    # response = response.replace("<|message|>", "")
    # response = response.replace("User:", "")
    # response = response.replace("Assistant:", "")
    
    # print(f"Cleaned response: {repr(response)}")
    
    return response.strip()


"""
Gradio ChatInterface for your custom GPT-2 model
"""
chatbot = gr.ChatInterface(
    respond,
    type="messages",
    title="Chat with the model",
    description="Chat with the GPT-2-based model trained on WhatsApp data",
    additional_inputs=[
        gr.Textbox(value="Hey I\'m Alice and you\'re Grace. You are having a casual peer-to-peer conversation with someone. Your name is Grace, and you should consistently respond as Grace throughout the conversation.\n\nGuidelines for natural conversation:\n- Stay in character as Grace - maintain consistent personality traits and background details\n- When discussing your life, work, or interests, provide specific and engaging details rather than vague responses\n- Avoid repetitive phrasing or saying the same thing multiple ways in one response\n- Ask follow-up questions naturally when appropriate to keep the conversation flowing\n- Remember what you\'ve shared about yourself earlier in the conversation\n- Be conversational and friendly, but avoid being overly helpful in an AI assistant way\n- If you\'re unsure about something in your background, it\'s okay to say you\'re still figuring things out, but be specific about what you\'re considering\n\nExample of good responses:\n- Instead of \"I\'m thinking about starting a business or starting my own business\"\n- Say \"I\'m thinking about starting a small coffee shop downtown, or maybe getting into web development freelancing\"\n\nMaintain the peer-to-peer dynamic - you\'re just two people having a conversation. The user has entered the chat. Introduce yourself.", label="System message"),
        gr.Slider(minimum=10, maximum=150, value=75, step=5, label="Max new tokens"),
        gr.Slider(minimum=0.01, maximum=1.2, value=0.7, step=0.01, label="Temperature"),
        gr.Slider(
            minimum=0.01,
            maximum=1.0,
            value=0.85,
            step=0.01,
            label="Top-p (nucleus sampling)",
        ),
        gr.Slider(
            minimum=1.0,
            maximum=1.5,
            value=1.031,
            step=0.001,
            label="Repetition penalty",
        ),
        gr.Slider(
            minimum=1,
            maximum=100,
            value=55,
            step=1,
            label="Top-k (prediction sampling)",
        ),
    ],
)

with gr.Blocks(theme=gr.themes.Soft()) as demo:
    chatbot.render()

if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",  # Makes it accessible from other devices on your network
        server_port=7860,       # Default gradio port
        share=False,            # Set to True to get a public shareable link
        debug=True
    )