app.py

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

access_token = os.getenv('HF_TOKEN')

# Define the repository ID and access token
repo_id = "Mikhil-jivus/Llama-32-3B-FineTuned"

# Load the tokenizer and model from the Hugging Face repository
tokenizer = AutoTokenizer.from_pretrained(repo_id, token=access_token)

model = AutoModelForCausalLM.from_pretrained(
    repo_id, 
    token=access_token, 
    torch_dtype=torch.bfloat16,  # or use torch.bfloat16 if supported
    device_map="auto"  # Automatically use available GPU/CPU efficiently
)

# Define a function to clean up any repeated segments in the generated response
def clean_response(response, history):
    # Check for repetition in the response and remove it
    if len(history) > 0:
        last_user_message, last_bot_response = history[-1]
        if last_bot_response in response:
            response = response.replace(last_bot_response, "").strip()
    
    return response

def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    # Add system prompt only once at the beginning of the conversation
    if len(history) == 0:
        input_text = f"system: {system_message}\nuser: {message}\n"
    else:
        input_text = f"user: {message}\n"

    # Append previous conversation history to the input text
    for user_msg, bot_msg in history:
        input_text += f"user: {user_msg}\nassistant: {bot_msg}\n"

    # Tokenize the input messages
    input_ids = tokenizer.encode(input_text, return_tensors="pt")
    
    # Move input_ids to the GPU
    input_ids = input_ids.to("cuda")
    
    # Create attention mask and move to GPU
    attention_mask = input_ids.ne(tokenizer.pad_token_id).long().to("cuda")
    
    # Generate a response
    chat_history_ids = model.generate(
        input_ids,
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        pad_token_id=tokenizer.eos_token_id,
        do_sample=True,
        attention_mask=attention_mask,
    )
    
    # Decode the response
    response = tokenizer.decode(chat_history_ids[:, input_ids.shape[-1]:][0], skip_special_tokens=True)

    # Clean the response to remove any repeated or unnecessary text
    response = clean_response(response, history)
    
    # Update history with the new user message and bot response
    history.append((message, response))

    return response

# Set up the Gradio app interface
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="You are a helpful and friendly assistant.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.95,
            step=0.05,
            label="Top-p (nucleus sampling)",
        ),
    ],
)

if __name__ == "__main__":
    demo.launch(share=True)