AgainstEntropy/Kanji-Streaming

1import argparse2import os3from queue import SimpleQueue4from threading import Thread5from typing import Iterator6 7import gradio as gr8import spaces9import torch10from gradio import Chatbot11from huggingface_hub import InferenceClient12 13from image_utils import ImageStitcher14from StreamDiffusionIO import LatentConsistencyModelStreamIO15 16MAX_MAX_NEW_TOKENS = 204817DEFAULT_MAX_NEW_TOKENS = 102418MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))19 20DESCRIPTION = """\21# Kanji-Streaming Chat22 23🌍 This Space is adapted from [Llama-2-7b-chat](https://huggingface.co/spaces/huggingface-projects/llama-2-7b-chat) space, demonstrating how to "chat" with LLM with [Kanji-Streaming](https://github.com/AgainstEntropy/kanji).24 25🔨 The technique behind Kanji-Streaming is [StreamDiffusionIO](https://github.com/AgainstEntropy/StreamDiffusionIO), which is based on [StreamDiffusion](https://github.com/cumulo-autumn/StreamDiffusion), *but especially allows to render text streams into image streams*.26 27🔎 For more details about Kanji-Streaming, take a look at the [github repository](https://github.com/AgainstEntropy/kanji).28"""29 30LICENSE = """31<p/>32 33---34As a derivate work of [Llama-2-7b-chat](https://huggingface.co/meta-llama/Llama-2-7b-chat) by Meta,35this demo is governed by the original [license](https://huggingface.co/spaces/huggingface-projects/llama-2-7b-chat/blob/main/LICENSE.txt) and [acceptable use policy](https://huggingface.co/spaces/huggingface-projects/llama-2-7b-chat/blob/main/USE_POLICY.md).36"""37 38 39parser = argparse.ArgumentParser(description="Gradio launcher for Streaming-Kanji.")40parser.add_argument(41    "--sd_model_id_or_path",42    type=str,43    default="stable-diffusion-v1-5/stable-diffusion-v1-5",44    required=False,45    help="Path to downloaded sd-1-5 model or model identifier from huggingface.co/models.",46)47parser.add_argument(48    "--lora_path",49    type=str,50    default="AgainstEntropy/kanji-lora-sd-v1-5",51    required=False,52    help="Path to downloaded LoRA weight or model identifier from huggingface.co/models.",53)54parser.add_argument(55    "--lcm_lora_path",56    type=str,57    default="AgainstEntropy/kanji-lcm-lora-sd-v1-5",58    required=False,59    help="Path to downloaded LCM-LoRA weight or model identifier from huggingface.co/models.",60)61parser.add_argument(62    "--img_res",63    type=int,64    default=64,65    required=False,66    help="Image resolution for displaying Kanji characters in ChatBot.",67)68parser.add_argument(69    "--img_per_line",70    type=int,71    default=16,72    required=False,73    help="Number of Kanji characters to display in a single line.",74)75parser.add_argument(76    "--tmp_dir",77    type=str,78    default="./tmp",79    required=False,80    help="Path to save temporary images generated by StreamDiffusionIO.",81)82 83args = parser.parse_args()84 85if torch.cuda.is_available():86    device = "cuda"87else:88    device = "cpu"89    DESCRIPTION += "\n<p>Running on CPU 🥶 This demo works best on GPU.</p>"90 91client = InferenceClient(92    model="mistralai/Mixtral-8x7B-Instruct-v0.1",93)94 95def format_prompt(message, history, system_prompt=''):96    prompt = f"<s> {system_prompt}"97    for user_prompt, bot_response in history:98        prompt += f"[INST] {user_prompt} [/INST]"99        if isinstance(bot_response, tuple):100            bot_response = bot_response[1]101        if not bot_response.endswith("</s>"):102            bot_response += "</s>"103        prompt += f" {bot_response} "104    prompt += f"[INST] {message} [/INST]"105    return prompt106 107lcm_stream = LatentConsistencyModelStreamIO(108    model_id_or_path=args.sd_model_id_or_path,109    lcm_lora_path=args.lcm_lora_path,110    lora_dict={args.lora_path: 1},111    resolution=128,112    device=device,113    use_xformers=True,114    verbose=True,115)116 117tmp_dir_template = f"{args.tmp_dir}/%d"118response_num = 0119 120stitcher = ImageStitcher(121    tmp_dir=tmp_dir_template % response_num,122    img_res=args.img_res,123    img_per_line=args.img_per_line,124    verbose=True,125)126 127 128@spaces.GPU129def generate(130    message: str,131    chat_history: list[tuple[str, str]],132    show_original_response: bool,133    seed: int,134    system_prompt: str = '',135    max_new_tokens: int = 1024,136    temperature: float = 0.6,137    top_p: float = 0.9,138    top_k: int = 50,139    repetition_penalty: float = 1.2,140) -> Iterator[str]:141    142    if temperature < 1e-2:143        temperature = 1e-2144    145    generate_kwargs = dict(146        max_new_tokens=max_new_tokens,147        do_sample=True,148        top_p=top_p,149        top_k=top_k,150        temperature=temperature,151        repetition_penalty=repetition_penalty,152    )153    formatted_prompt = format_prompt(message, chat_history, system_prompt)154    print(formatted_prompt)155    streamer = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)156 157    outputs = []158    prompt_queue = SimpleQueue()159 160    lcm_stream.reset(seed)161    stitcher.reset()162 163    global response_num164    response_num += 1165    stitcher.update_tmp_dir(tmp_dir_template % response_num)166 167    def append_to_queue():168        for response in streamer:169            text = response.token.text170            outputs.append(text)171            prompt = text.strip()172            if prompt and prompt not in ['</s>']:173                if prompt.endswith("."): prompt = prompt[:-1]174                prompt_queue.put(prompt)175        prompt_queue.put(None)176 177    append_thread = Thread(target=append_to_queue)178    append_thread.start()179 180    def show_image(prompt: str = None):181        image, text = lcm_stream(prompt)182        img_path = None183        if image is not None:184            img_path = stitcher.add(image, text)185        return img_path186 187    while True:188        prompt = prompt_queue.get()189        if prompt is None:190            break191        img_path = show_image(prompt)192        if img_path is not None:193            yield (img_path, )194 195    # Continue to display the remaining images196    while True:197        img_path = show_image()198        if img_path is not None:199            yield (img_path, ''.join(outputs))200        if lcm_stream.stop():201            break202 203    print(outputs)204    if show_original_response:205        yield ''.join(outputs)206 207 208chat_interface = gr.ChatInterface(209    fn=generate,210    chatbot=Chatbot(height=400),211    additional_inputs=[212        gr.Checkbox(213            label="Show original response",214            value=False,215        ),216        gr.Number(217            label="Seed",218            info="Random Seed for Kanji Generation (maybe some kind of accent 🤔)",219            step=1,220            value=1026,221        ),222        gr.Textbox(223            label="System prompt", 224            value="",225            lines=4),226        gr.Slider(227            label="Max new tokens",228            minimum=1,229            maximum=MAX_MAX_NEW_TOKENS,230            step=1,231            value=DEFAULT_MAX_NEW_TOKENS,232        ),233        gr.Slider(234            label="Temperature",235            minimum=0.1,236            maximum=4.0,237            step=0.1,238            value=0.6,239        ),240        gr.Slider(241            label="Top-p (nucleus sampling)",242            minimum=0.05,243            maximum=1.0,244            step=0.05,245            value=0.9,246        ),247        gr.Slider(248            label="Top-k",249            minimum=1,250            maximum=1000,251            step=1,252            value=50,253        ),254        gr.Slider(255            label="Repetition penalty",256            minimum=1.0,257            maximum=2.0,258            step=0.05,259            value=1.2,260        ),261    ],262    stop_btn=None,263    examples=[264        ["Hello there! How are you doing?"],265        ["Can you explain briefly to me what is the Python programming language?"],266        ["Explain the plot of Cinderella in a sentence."],267        ["How many hours does it take a man to eat a Helicopter?"],268        ["Write a 100-word article on 'Benefits of Open-Source in AI research'"],269    ],270)271 272with gr.Blocks(css="style.css") as demo:273    gr.Markdown(DESCRIPTION)274    gr.DuplicateButton(value="Duplicate Space for private use", elem_id="duplicate-button")275    chat_interface.render()276    gr.Markdown(LICENSE)277 278if __name__ == "__main__":279    demo.queue(max_size=20).launch(server_name="0.0.0.0", share=False, show_api=False)280