app.py

import spaces
import json
import subprocess
from llama_cpp import Llama
from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType
from llama_cpp_agent.providers import LlamaCppPythonProvider
from llama_cpp_agent.chat_history import BasicChatHistory
from llama_cpp_agent.chat_history.messages import Roles
import gradio as gr
import random
from datasets import load_dataset
from huggingface_hub import hf_hub_download

# モデルのダウンロード
hf_hub_download(
    repo_id="team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GGUF",
    filename="Tanuki-8x8B-dpo-v1.0-IQ4_NL.gguf",
    local_dir="./models"
)

hf_hub_download(
    repo_id="team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GGUF",
    filename="Tanuki-8x8B-dpo-v1.0-Q6_K.gguf",
    local_dir="./models"
)

hf_hub_download(
    repo_id="team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GGUF",
    filename="Tanuki-8x8B-dpo-v1.0-IQ3_M.gguf",
    local_dir="./models"
)



llm = None
llm_model = None

# データセットをロードしてスプリットを確認
dataset = load_dataset("elyza/ELYZA-tasks-100")
print(dataset)

# 使用するスプリット名を確認
split_name = "train" if "train" in dataset else "test"  # デフォルトをtrainにし、なければtestにフォールバック

# 適切なスプリットから10個の例を取得
examples_list = list(dataset[split_name])  # スプリットをリストに変換
examples = random.sample(examples_list, 10)  # リストからランダムに10個選択
example_inputs = [[example['input']] for example in examples]  # ネストされたリストに変換


@spaces.GPU(duration=120)
def respond(
    message,
    history: list[tuple[str, str]],
    model,
    template,
    system_message,
    max_tokens,
    temperature,
    top_p,
    top_k,
    repeat_penalty,
):
    chat_template = MessagesFormatterType[template]

    global llm
    global llm_model
    
    if llm is None or llm_model != model:
        llm = Llama(
            model_path=f"models/{model}",
            flash_attn=True,
            n_gpu_layers=81,
            n_batch=1024,
            n_ctx=8192,
        )
        llm_model = model

    provider = LlamaCppPythonProvider(llm)

    agent = LlamaCppAgent(
        provider,
        system_prompt=f"{system_message}",
        predefined_messages_formatter_type=chat_template,
        debug_output=True
    )
    
    settings = provider.get_provider_default_settings()
    settings.temperature = temperature
    settings.top_k = top_k
    settings.top_p = top_p
    settings.max_tokens = max_tokens
    settings.repeat_penalty = repeat_penalty
    settings.stream = True

    messages = BasicChatHistory()

    for msn in history:
        user = {
            'role': Roles.user,
            'content': msn[0]
        }
        assistant = {
            'role': Roles.assistant,
            'content': msn[1]
        }
        messages.add_message(user)
        messages.add_message(assistant)
    
    stream = agent.get_chat_response(
        message,
        llm_sampling_settings=settings,
        chat_history=messages,
        returns_streaming_generator=True,
        print_output=False
    )
    
    outputs = ""
    for output in stream:
        outputs += output
        yield outputs

description = """<p align="center">★画面下のAdditional Inputから、使用したいモデルと、チャットテンプレートを選択してください。★</p>
<p align="center">★使用時の注意：現在GGUF版は著明な性能低下が確認されていますので、本来の性能が発揮されていません。本来の性能は公式推奨の推論方法ご利用ください★</p>
<p><center>
<a href="https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GGUF/" target="_blank">Tanuki-8x8B-dpo-v1.0-IQ4_NL.gguf</a><br>
<a href="https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GGUF/" target="_blank">Tanuki-8x8B-dpo-v1.0-Q6_K.gguf</a><br>
<a href="https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GGUF/" target="_blank">Tanuki-8x8B-dpo-v1.0-IQ3_M.gguf</a><br>
</center></p>
"""

templates = [
    "MISTRAL", "CHATML", "VICUNA", "LLAMA_2", "SYNTHIA",
    "NEURAL_CHAT", "SOLAR", "OPEN_CHAT", "ALPACA", "CODE_DS",
    "B22", "LLAMA_3", "PHI_3"
]

demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Dropdown([
                'Tanuki-8x8B-dpo-v1.0-IQ4_NL.gguf',
                'Tanuki-8x8B-dpo-v1.0-Q6_K.gguf',
                'Tanuki-8x8B-dpo-v1.0-IQ3_M.gguf'
            ],
            value="Tanuki-8x8B-dpo-v1.0-IQ4_NL.gguf",
            label="Model"
        ),
        gr.Dropdown(
            choices=templates,
            value="ALPACA",
            label="Template"
        ),
        gr.Textbox(value="以下は、タスクを説明する指示です。要求を適切に満たす応答を書きなさい。", label="System message"),
        gr.Slider(minimum=1, maximum=4096, value=2048, step=1, label="Max 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",
        ),
        gr.Slider(
            minimum=0,
            maximum=100,
            value=40,
            step=1,
            label="Top-k",
        ),
        gr.Slider(
            minimum=0.0,
            maximum=2.0,
            value=1.1,
            step=0.1,
            label="Repetition penalty",
        ),
    ],
    examples=example_inputs,
    cache_examples=False,
    retry_btn="Retry",
    undo_btn="Undo",
    clear_btn="Clear",
    submit_btn="Send",
    title="Chat with various models using llama.cpp", 
    description=description,
    chatbot=gr.Chatbot(
        scale=1, 
        likeable=False,
        show_copy_button=True
    )
)

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