app.py

#!/usr/bin/env python
import os
import random
from threading import Thread
from typing import Iterator

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

from chat_interface_preference import ChatInterface

MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "8192"))

if torch.cuda.is_available():
    model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"
    model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
    tokenizer = AutoTokenizer.from_pretrained(model_id)


@spaces.GPU
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    max_new_tokens: int = 1024,
    temperature: float = 0.06,
    top_p: float = 0.95,
    top_k: int = 40,
    repetition_penalty: float = 1.2,
) -> Iterator[str]:
    system_message = random.choice(["concise", "explicit", "simple", "complex", "usefull", "helpfull"])
    conversation = [{"role": "system", "content": f"Communicate {system_message}."}]
    for user, assistant in chat_history:
        conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
    conversation.append({"role": "user", "content": message})

    input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
    if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
        input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
        gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
    input_ids = input_ids.to(model.device)

    streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = dict(
        {"input_ids": input_ids},
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        num_beams=1,
        repetition_penalty=repetition_penalty,
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    outputs = []
    for text in streamer:
        outputs.append(text)
        yield "".join(outputs)


chat_interface = ChatInterface(
    fn=generate,
    prefence_techniques="dpo",
    min_turns=1,
    max_turns=10,
    repo_id="llm-human-feedback-collector-chat-interface-dpo",
    chatbot=gr.Chatbot(height=450, label="Meta-Llama-3.1-8B-Instruct", show_share_button=True),
    cache_examples=False,
    additional_inputs=[
        gr.Slider(
            label="Max new tokens",
            minimum=1,
            maximum=MAX_MAX_NEW_TOKENS,
            step=1,
            value=DEFAULT_MAX_NEW_TOKENS,
        ),
        gr.Slider(
            label="Temperature",
            minimum=0.05,
            maximum=1.2,
            step=0.05,
            value=0.2,
        ),
        gr.Slider(
            label="Top-p (nucleus sampling)",
            minimum=0.05,
            maximum=1.0,
            step=0.05,
            value=0.9,
        ),
        gr.Slider(
            label="Top-k",
            minimum=1,
            maximum=1000,
            step=1,
            value=50,
        ),
        gr.Slider(
            label="Repetition penalty",
            minimum=1.0,
            maximum=2.0,
            step=0.05,
            value=1.2,
        ),
    ],
    examples=[
        ["""What word doesn't make sense in this row: "car, airplane, lama, bus"?"""],
        ["Write a news article about the usage of Lama's by the CSI"],
        ["What are great things cook when getting started with Asian cooking?"],
        ["Who was Anthony Bourdain?"],
    ],
    title="💪🏽🦾 Human Feedback Collector | Meta-Llama-3.1-8B-Instruct | (DPO) 🦾💪🏽",
    description="".join(
        [
            "This is an adaptation of the [`gr.ChatInferface`](https://www.gradio.app/docs/gradio/chatinterface) and [`huggingface_hub.CommitScheduler`](https://huggingface.co/docs/huggingface_hub/main/en/package_reference/hf_api#huggingface_hub.CommitScheduler) which allows for human feedback collection. ",
            "Another cool tool for capturing Gradio interactions is the [`gr.HuggingFaceDatasetSaver`](https://www.gradio.app/guides/using-flagging#the-hugging-face-dataset-saver-callback). ",
            "This demo shows how you might capture human feedback directly from applications within Gradio. ",
            "The captured feedback can directly be used for fine-tuning LLMs within framework like [transformers](https://github.com/huggingface/transformers), [TRL](https://github.com/huggingface/trl) or [AutoTrain](https://huggingface.co/autotrain), ",
            "however, it might benefit from additional data curation with something like [Argilla](https://github.com/argilla-io/argilla/) for human feedback and/or [distilabel](https://github.com/argilla-io/distilabel/) for AI feedback. Argilla can even be [deployed for free on Hugging Face Spaces](https://argilla-io.github.io/argilla/latest/getting_started/huggingface-spaces/).",
        ]
    ),
)

with gr.Blocks(css="style.css") as demo:
    chat_interface.render()

if __name__ == "__main__":
    demo.queue(max_size=20).launch()