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

MAX_MAX_NEW_TOKENS = 4096
DEFAULT_MAX_NEW_TOKENS = 2048
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))

DESCRIPTION = """\
# CantoneseLLM Chat

Please join our [Discord server](https://discord.gg/gG6GPp8XxQ) and give me your feedback
"""


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():
    model_id = "hon9kon9ize/CantoneseLLMChat-preview20240326"
    model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16,)
    model = torch.compile(model)
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    tokenizer.use_default_system_prompt = False


@spaces.GPU
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    system_prompt: str,
    max_new_tokens: int = 2048,
    temperature: float = 0.6,
    top_p: float = 0.9,
    top_k: int = 50,
    repetition_penalty: float = 1.2,
) -> 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})

    print(chat_history)

    input_ids = tokenizer.apply_chat_template(conversation, tokenize=True, 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)
    output_ids = model.generate(
        input_ids,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        repetition_penalty=repetition_penalty
    )

    response = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True)
    return response
        


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=[
        ["Hello there! How are you doing?"],
        ["咩嘢係氣候變化?"],
        ["香港最高嘅山係？"],
        ["邊個係香港特首？"],
        ["香港行政长官是谁？"]
    ],
)

with gr.Blocks() as demo:
    gr.Markdown(DESCRIPTION)
    chat_interface.render()

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