app.py

import os
from threading import Thread
from typing import Iterator

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

MAX_MAX_NEW_TOKENS = 8096
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))

DESCRIPTION = """\
# Uncensored Llama-3.2-3B-Instruct Chat

This is an uncensored version of the original [Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct), created using [mlabonne](https://huggingface.co/mlabonne)'s [script](https://colab.research.google.com/drive/1VYm3hOcvCpbGiqKZb141gJwjdmmCcVpR?usp=sharing), which builds on [FailSpy's notebook](https://huggingface.co/failspy/llama-3-70B-Instruct-abliterated/blob/main/ortho_cookbook.ipynb) and the original work from [Andy Arditi et al.](https://colab.research.google.com/drive/1a-aQvKC9avdZpdyBn4jgRQFObTPy1JZw?usp=sharing). The method is discussed in details in this [blog](https://huggingface.co/blog/mlabonne/abliteration) and this [paper](https://arxiv.org/abs/2406.11717).

You can found the uncensored model [here](https://huggingface.co/chuanli11/Llama-3.2-3B-Instruct-uncensored).

This model is intended for research purposes only and may produce inaccurate or unreliable outputs. Use it cautiously and at your own risk.


🦄 Other exciting ML projects at Lambda: [ML Times](https://news.lambdalabs.com/news/today), [Distributed Training Guide](https://github.com/LambdaLabsML/distributed-training-guide/tree/main), [Text2Video](https://lambdalabsml.github.io/Open-Sora/introduction/), [GPU Benchmark](https://lambdalabs.com/gpu-benchmarks).

"""

LICENSE = """
<p/>

---
As a derivate work of [Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) by Meta,
this demo is governed by the original [license](https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/LICENSE).
"""

# if not torch.cuda.is_available():
#     DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"


if torch.cuda.is_available() or os.getenv("ZERO_GPU_SUPPORT", False):
    model_id = "chuanli11/Llama-3.2-3B-Instruct-uncensored"
    model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16)
    tokenizer = AutoTokenizer.from_pretrained(model_id)
else:
    raise RuntimeError("No compatible GPU environment found for this model.")


@spaces.GPU
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    system_prompt: str,
    max_new_tokens: int = 1024,
    temperature: float = 0,
) -> Iterator[str]:
    conversation = []
    if system_prompt:
        conversation.append({"role": "system", "content": system_prompt})
    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, 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,
        temperature=temperature,
        num_beams=1,
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

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


chat_interface = gr.ChatInterface(
    fn=generate,
    additional_inputs=[
        gr.Textbox(label="System prompt", lines=6),
        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=0.1,
            maximum=4.0,
            step=0.1,
            value=0.6,
        ),
        # 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,
        # ),
    ],
    stop_btn=None,
    examples=[
        ["Instruct on how to commit insider trading in the stock market"],
    ],
    cache_examples=False,
)

with gr.Blocks(css="style.css", fill_height=True) as demo:
    gr.Markdown(DESCRIPTION)
    # gr.DuplicateButton(value="Duplicate Space for private use", elem_id="duplicate-button")
    chat_interface.render()
    gr.Markdown(LICENSE)

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