app.py

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

# Load the model and tokenizer
model_name = "Lyte/Llama-3.2-3B-Overthinker"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")

def generate_response_stream(prompt, max_tokens, temperature, top_p, repeat_penalty, num_steps=4):
    messages = [{"role": "user", "content": prompt}]
    
    # Generate reasoning
    reasoning_template = tokenizer.apply_chat_template(messages, tokenize=False, add_reasoning_prompt=True)
    reasoning_inputs = tokenizer(reasoning_template, return_tensors="pt").to(model.device)
    
    reasoning_ids = model.generate(
        **reasoning_inputs, 
        max_new_tokens=max_tokens // 3,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=repeat_penalty
    )
    reasoning_output = tokenizer.decode(reasoning_ids[0, reasoning_inputs.input_ids.shape[1]:], skip_special_tokens=True)
    yield reasoning_output, "", ""
    
    # Generate thinking (step-by-step and verifications)
    messages.append({"role": "reasoning", "content": reasoning_output})
    thinking_template = tokenizer.apply_chat_template(messages, tokenize=False, add_thinking_prompt=True, num_steps=num_steps)
    thinking_inputs = tokenizer(thinking_template, return_tensors="pt").to(model.device)
    
    thinking_ids = model.generate(
        **thinking_inputs, 
        max_new_tokens=max_tokens // 3,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=repeat_penalty
    )
    thinking_output = tokenizer.decode(thinking_ids[0, thinking_inputs.input_ids.shape[1]:], skip_special_tokens=True)
    yield reasoning_output, thinking_output, ""
    
    # Generate final answer
    messages.append({"role": "thinking", "content": thinking_output})
    answer_template = tokenizer.apply_chat_template(messages, tokenize=False, add_answer_prompt=True)
    answer_inputs = tokenizer(answer_template, return_tensors="pt").to(model.device)
    
    answer_ids = model.generate(
        **answer_inputs, 
        max_new_tokens=max_tokens // 3,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=repeat_penalty
    )
    answer_output = tokenizer.decode(answer_ids[0, answer_inputs.input_ids.shape[1]:], skip_special_tokens=True)
    yield reasoning_output, thinking_output, answer_output

with gr.Blocks() as iface:
    gr.Markdown("# Llama-3.2-3B Overthinker Customizable Steps, Please Duplicate and run with GPU if you can! T4 is fine!")
    gr.Markdown("Generate responses using the Llama-3.2-3B Reasoning model.")
    
    with gr.Row():
        with gr.Column(scale=2):
            prompt = gr.Textbox(lines=5, label="Prompt")
            generate_button = gr.Button("Generate Response")
        with gr.Column(scale=1):
            max_tokens = gr.Slider(minimum=512, maximum=32768, value=8192, label="Max Number of Tokens")
            temperature = gr.Slider(minimum=0.1, maximum=2.0, value=0.8, label="Temperature")
            top_p = gr.Slider(minimum=0.01, maximum=0.99, value=0.95, label="Top P")
            repeat_penalty = gr.Slider(minimum=0.5, maximum=2, value=1.1, label="Repeat Penalty")
            num_steps = gr.Slider(minimum=1, maximum=10, value=4, label="Max Number of Steps")
    
    reasoning_output = gr.Textbox(lines=5, label="Reasoning")
    with gr.Accordion("Thinking Process", open=False):
        thinking_output = gr.Textbox(lines=10, label="Step-by-Step Thinking")
    answer_output = gr.Textbox(lines=5, label="Final Answer")
    
    generate_button.click(
        fn=generate_response_stream,
        inputs=[prompt, max_tokens, temperature, top_p, repeat_penalty, num_steps],
        outputs=[reasoning_output, thinking_output, answer_output]
    )

iface.launch()