app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer, pipeline
import torch
from threading import Thread

MODEL_ID = "HODACHI/Llama-3.1-8B-EZO-1.1-it"
DTYPE = torch.bfloat16

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=DTYPE,
    device_map="auto",
    low_cpu_mem_usage=True,
)

pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device_map="auto",
)

def generate_text(prompt, max_new_tokens, temperature, top_p):
    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
    generation_kwargs = dict(
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        top_p=top_p,
        do_sample=True,
        streamer=streamer,
    )
    
    thread = Thread(target=pipe, kwargs=dict(text_inputs=prompt, **generation_kwargs))
    thread.start()
    
    return streamer

def respond(message, history, max_tokens, temperature, top_p):
    chat = []
    chat.append({"role": "system", "content": "あなたは誠実で優秀な日本人のアシスタントです。特に指示が無い場合は、原則日本語で回答してください。"})
    for user, assistant in history:
        chat.append({"role": "user", "content": user})
        chat.append({"role": "assistant", "content": assistant})
    chat.append({"role": "user", "content": message})
    prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
    
    streamer = generate_text(prompt, max_tokens, temperature, top_p)
    
    response = ""
    for new_text in streamer:
        response += new_text
        yield response

demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Slider(minimum=1, maximum=2048, value=150, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"),
    ],
)

if __name__ == "__main__":
    demo.launch()