app.py

import os
import gradio as gr
from llama_cpp import Llama

# CPU最適化
os.environ["OMP_NUM_THREADS"] = "8"
os.environ["OPENBLAS_NUM_THREADS"] = "8"
os.environ["MKL_NUM_THREADS"] = "8"

llm = Llama.from_pretrained(
    repo_id="summerMC/ume-GGUF",
    filename="ume-Q4_K_M.gguf",
    n_ctx=768,          # 小さめにすると速い
    n_threads=8,        # SpaceのCPUコア数に合わせて調整
    n_batch=512,        # prompt処理高速化
    n_ubatch=128,
    use_mmap=True,
    use_mlock=False,
    logits_all=False,
    embedding=False,
    verbose=False,
)

def respond(message, history, system_message, max_tokens, temperature, top_p):
    messages = [{"role": "system", "content": system_message}]

    # Gradio ChatInterface の history は [(user, assistant), ...]
    for user_msg, bot_msg in history:
        messages.append({"role": "user", "content": user_msg})
        if bot_msg:
            messages.append({"role": "assistant", "content": bot_msg})

    messages.append({"role": "user", "content": message})

    response = ""
    for chunk in llm.create_chat_completion(
        messages=messages,
        max_tokens=min(int(max_tokens), 256),  # CPUでは長すぎると遅い
        temperature=float(temperature),
        top_p=float(top_p),
        stream=True,
    ):
        delta = chunk["choices"][0].get("delta", {})
        token = delta.get("content", "")
        response += token
        yield response


chatbot = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Textbox(
            value="You are a concise and helpful assistant.",
            label="System message"
        ),
        gr.Slider(1, 256, value=128, step=1, label="Max new tokens"),
        gr.Slider(0.1, 1.5, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p"),
    ],
)

with gr.Blocks() as demo:
    chatbot.render()

if __name__ == "__main__":
    demo.queue(default_concurrency_limit=1).launch()