app-kanji-local-llama.py

import argparse
import os
from queue import SimpleQueue
from threading import Thread
from typing import Iterator

import gradio as gr
import spaces
import torch
from gradio import Chatbot
from image_utils import ImageStitcher
from transformers import (AutoModelForCausalLM, AutoTokenizer,
                          TextIteratorStreamer)

from StreamDiffusionIO import LatentConsistencyModelStreamIO

MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))

DESCRIPTION = """\
# Kanji-Streaming Chat

🌍 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).

🔨 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*.

🔎 For more details about Kanji-Streaming, take a look at the [github repository](https://github.com/AgainstEntropy/kanji).
"""

LICENSE = """
<p/>

---
As a derivate work of [Llama-2-7b-chat](https://huggingface.co/meta-llama/Llama-2-7b-chat) by Meta,
this 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).
"""


parser = argparse.ArgumentParser(description="Gradio launcher for Streaming-Kanji.")
parser.add_argument(
    "--llama_model_id_or_path",
    type=str,
    default="meta-llama/Llama-2-7b-chat-hf",
    required=False,
    help="Path to downloaded llama-chat-hf model or model identifier from huggingface.co/models.",
)
parser.add_argument(
    "--sd_model_id_or_path",
    type=str,
    default="runwayml/stable-diffusion-v1-5",
    required=False,
    help="Path to downloaded sd-1-5 model or model identifier from huggingface.co/models.",
)
parser.add_argument(
    "--lora_path",
    type=str,
    default="AgainstEntropy/kanji-lora-sd-v1-5",
    required=False,
    help="Path to downloaded LoRA weight or model identifier from huggingface.co/models.",
)
parser.add_argument(
    "--lcm_lora_path",
    type=str,
    default="AgainstEntropy/kanji-lcm-lora-sd-v1-5",
    required=False,
    help="Path to downloaded LCM-LoRA weight or model identifier from huggingface.co/models.",
)
parser.add_argument(
    "--img_res",
    type=int,
    default=64,
    required=False,
    help="Image resolution for displaying Kanji characters in ChatBot.",
)
parser.add_argument(
    "--img_per_line",
    type=int,
    default=16,
    required=False,
    help="Number of Kanji characters to display in a single line.",
)
parser.add_argument(
    "--tmp_dir",
    type=str,
    default="./tmp",
    required=False,
    help="Path to save temporary images generated by StreamDiffusionIO.",
)

args = parser.parse_args()

if torch.cuda.is_available():
    device = "cuda"
else:
    device = "cpu"
    DESCRIPTION += "\n<p>Running on CPU 🥶 This demo works best on GPU.</p>"

DESCRIPTION += "\n<p>This demo will get the best kanji streaming experience in localhost (or SSH forward), instead of shared link generated by Gradio.</p>"

model = AutoModelForCausalLM.from_pretrained(args.llama_model_id_or_path, torch_dtype=torch.float16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(args.llama_model_id_or_path)
tokenizer.use_default_system_prompt = False
streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)

lcm_stream = LatentConsistencyModelStreamIO(
    model_id_or_path=args.sd_model_id_or_path,
    lcm_lora_path=args.lcm_lora_path,
    lora_dict={args.lora_path: 1},
    resolution=128,
    device=device,
    use_xformers=True,
    verbose=True,
)

tmp_dir_template = f"{args.tmp_dir}/%d"
response_num = 0

stitcher = ImageStitcher(
    tmp_dir=tmp_dir_template % response_num,
    img_res=args.img_res,
    img_per_line=args.img_per_line,
    verbose=True,
)


@spaces.GPU
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    show_original_response: bool,
    seed: int,
    system_prompt: str = '',
    max_new_tokens: int = 1024,
    temperature: float = 0.6,
    top_p: float = 0.9,
    top_k: int = 50,
    repetition_penalty: float = 1.2,
) -> Iterator[str]:
    
    conversation = []
    if system_prompt:
        conversation.append({"role": "system", "content": system_prompt})
    for user, assistant in chat_history:
        if isinstance(assistant, tuple):
            assistant = assistant[1]
        else:
            assistant = str(assistant)
        conversation.extend([
            {"role": "user", "content": user}, 
            {"role": "assistant", "content": assistant},
        ])
    conversation.append({"role": "user", "content": message})
    print(conversation)

    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)

    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,
        pad_token_id=tokenizer.eos_token_id
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    outputs = []
    prompt_queue = SimpleQueue()

    lcm_stream.reset(seed)
    stitcher.reset()

    global response_num
    response_num += 1
    stitcher.update_tmp_dir(tmp_dir_template % response_num)

    def append_to_queue():
        for text in streamer:
            outputs.append(text)
            prompt = text.strip()
            if prompt:
                if prompt.endswith("."): prompt = prompt[:-1]
                prompt_queue.put(prompt)
        prompt_queue.put(None)

    append_thread = Thread(target=append_to_queue)
    append_thread.start()

    def show_image(prompt: str = None):
        image, text = lcm_stream(prompt)
        img_path = None
        if image is not None:
            img_path = stitcher.add(image, text)
        return img_path

    while True:
        prompt = prompt_queue.get()
        if prompt is None:
            break
        img_path = show_image(prompt)
        if img_path is not None:
            yield (img_path, )

    # Continue to display the remaining images
    while True:
        img_path = show_image()
        if img_path is not None:
            yield (img_path, ''.join(outputs))
        if lcm_stream.stop():
            break

    print(outputs)
    if show_original_response:
        yield ''.join(outputs)


chat_interface = gr.ChatInterface(
    fn=generate,
    chatbot=Chatbot(height=400),
    additional_inputs=[
        gr.Checkbox(
            label="Show original response",
            value=False,
        ),
        gr.Number(
            label="Seed",
            info="Random Seed for Kanji Generation (maybe some kind of accent 🤔)",
            step=1,
            value=1026,
        ),
        gr.Textbox(label="System prompt", lines=4),
        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?"],
        ["Can you explain briefly to me what is the Python programming language?"],
        ["Explain the plot of Cinderella in a sentence."],
        ["How many hours does it take a man to eat a Helicopter?"],
        ["Write a 100-word article on 'Benefits of Open-Source in AI research'"],
    ],
)

with gr.Blocks(css="style.css") 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(server_name="0.0.0.0", share=False)