app.py

import re
import threading

import gradio as gr
import spaces
import transformers
from transformers import pipeline

# loading model and tokenizer
model_name = "Qwen/Qwen2-1.5B-Instruct"
if gr.NO_RELOAD:
    pipe = pipeline(
        "text-generation",
        model=model_name,
        device_map="auto",
        torch_dtype="auto",
    )

# the answer marker to detect final answer
ANSWER_MARKER = "**ANSWER**"

# the sentences starting the reasoning step by step
rethink_prepends = [
    "OK, I need to figure out ",
    "I think ",
    "Wait, I think ",
    "Let me check if ",
    "I should also remember that ",
    "Another thing to note is that ",
    "I also recall that ",
    "I think I have a good grasp ",
    "Now, using all the above information, I can answer the question using the original language used for the question:"
    "\n{question}\n"
    f"\n{ANSWER_MARKER}\n",
]


# to fix some problems with math display
latex_delimiters = [
    {"left": "$$", "right": "$$", "display": True},
    {"left": "$", "right": "$", "display": False},
]


def reformat_math(text):
    """Fix MathJax delimiters to use the Gradio syntax (Katex).

    This is a workaround to display math formulas in Gradio. For now, I havn't found a way to
    make it work as expected using others latex_delimiters...
    """
    text = re.sub(r"\\\[\s*(.*?)\s*\\\]", r"$$\1$$", text, flags=re.DOTALL)
    text = re.sub(r"\\\(\s*(.*?)\s*\\\)", r"$\1$", text, flags=re.DOTALL)
    return text


def user_input(message, history: list):
    """Append the user input in the history and clean the input textbox"""
    return "", history + [
        gr.ChatMessage(role="user", content=message.replace(ANSWER_MARKER, ""))
    ]


def rebuild_messages(history: list):
    """Rebuid the messages from the history to be used by the model without the intermediate thoughs"""
    messages = []
    for h in history:
        if isinstance(h, dict) and not h.get("metadata", {}).get("title", False):
            messages.append(h)
        elif (
            isinstance(h, gr.ChatMessage)
            and h.metadata.get("title")
            and isinstance(h.content, str)
        ):
            messages.append({"role": h.role, "content": h.content})
    return messages


@spaces.GPU
def bot(
    history: list,
    max_num_tokens: int,
    final_num_tokens: int,
    do_sample: bool,
    temperature: float,
):
    """Make the model answering the question"""

    # to get token as a stream, later in a thread
    streamer = transformers.TextIteratorStreamer(
        pipe.tokenizer,  # pyright: ignore
        skip_special_tokens=True,
        skip_prompt=True,
    )

    # to reinsert the question in the reasoning if needed
    question = history[-1]["content"]

    # prepare the assistant message
    history.append(
        gr.ChatMessage(
            role="assistant",
            content=str(""),
            metadata={"title": "🧠 Thinking...", "status": "pending"},
        )
    )

    # for the moment, make the reasoning to be displayed in the chat
    messages = rebuild_messages(history)
    for i, prepend in enumerate(rethink_prepends):
        if i > 0:
            messages[-1]["content"] += "\n\n"
        messages[-1]["content"] += prepend.format(question=question)

        num_tokens = int(
            max_num_tokens if ANSWER_MARKER not in prepend else final_num_tokens
        )
        t = threading.Thread(
            target=pipe,
            args=(messages,),
            kwargs=dict(
                max_new_tokens=num_tokens,
                streamer=streamer,
                do_sample=do_sample,
                temperature=temperature,
            ),
        )
        t.start()

        # rebuild the history with the new content
        history[-1].content += prepend.format(question=question)
        if ANSWER_MARKER in prepend:
            history[-1].metadata = {"title": "💭 Thoughs", "status": "done"}
            # stop thinking, this is the answer now (no metadata for intermediate steps)
            history.append(gr.ChatMessage(role="assistant", content=""))
        for token in streamer:
            history[-1].content += token
            history[-1].content = reformat_math(history[-1].content)
            yield history
        t.join()

    yield history


with gr.Blocks(fill_height=True, title="Making any LLM model reasoning") as demo:
    with gr.Row(scale=1):
        with gr.Column(scale=5):
            gr.Markdown(f"""
            # Force reasoning for any LLM
            
            This is a simple proof-of-concept to get any LLM (Large language Model) to reason ahead of its response.
            This interface uses *{model_name}* model **which is not a reasoning model**. The used method
            is only to force some "reasoning" steps with prefixes to help the model to enhance the answer.

            See my related article here: [Make any model reasoning](https://huggingface.co/blog/Metal3d/making-any-model-reasoning)
            """)
            chatbot = gr.Chatbot(
                scale=1,
                type="messages",
                latex_delimiters=latex_delimiters,
            )
            msg = gr.Textbox(
                submit_btn=True,
                label="",
                show_label=False,
                placeholder="Type your question here.",
                autofocus=True,
            )
        with gr.Column(scale=1):
            gr.Markdown("""## Tweaking""")
            num_tokens = gr.Slider(
                50,
                1024,
                100,
                step=1,
                label="Max tokens per reasoning step",
                interactive=True,
            )
            final_num_tokens = gr.Slider(
                50,
                1024,
                512,
                step=1,
                label="Max token for the final answer",
                interactive=True,
            )
            do_sample = gr.Checkbox(True, label="Do sample")
            temperature = gr.Slider(0.1, 1.0, 0.7, step=0.1, label="Temperature")
            gr.Markdown("""
            Using smaller number of tokens in the reasoning steps will make the model
            faster to answer, but it may not be able to go deep enough in its reasoning.
            A good value is 100 to 512.

            Using smaller number of tokens for the final answer will make the model
            to be less verbose, but it may not be able to give a complete answer.
            A good value is 512 to 1024.

            **Do sample** uses another strategie to select the next token to complete the
            answer. It's commonly better to leave it checked.

            **Temperature** indicates how much the model could be "creative". 0.7 is a common value.
            If you set a too high value (like 1.0) the model could be incoherent. With a low value 
            (like 0.3), the model will produce very predictives answers.
            """)
            gr.Markdown("""
            This interface can work on personal computer with 6Go VRAM (e.g. NVidia 3050/3060 on laptop). 
            Feel free to fork the application and try others instruct models.
            """)

    # when the user submit a message, the bot will answer
    msg.submit(
        user_input,
        [msg, chatbot],  # inputs
        [msg, chatbot],  # outputs
    ).then(
        bot,
        [
            chatbot,
            num_tokens,
            final_num_tokens,
            do_sample,
            temperature,
        ],  # actually, the "history" input
        chatbot,  # to store the new history from the output
    )

if __name__ == "__main__":
    demo.queue().launch()