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
from huggingface_hub import hf_hub_download

# モデルのダウンロード
hf_hub_download(
    repo_id="Aratako/Oumuamua-7b-RP-GGUF",
    filename="Oumuamua-7b-RP_Q4_K_M.gguf",
    local_dir="./models"
)

hf_hub_download(
    repo_id="bartowski/Oumuamua-7b-instruct-v2-GGUF",
    filename="Oumuamua-7b-instruct-v2-Q4_K_M.gguf",
    local_dir="./models"
)

hf_hub_download(
    repo_id="mmnga/umiyuki-Umievo-itr012-Gleipnir-7B-gguf",
    filename="umiyuki-Umievo-itr012-Gleipnir-7B-Q4_K_M.gguf",
    local_dir="./models"
)

hf_hub_download(
    repo_id="Local-Novel-LLM-project/Ninja-V3-GGUF",
    filename="Ninja-V3-Q4_K_M.gguf",
    local_dir="./models"
)

hf_hub_download(
    repo_id="Local-Novel-LLM-project/Kagemusya-7B-v1-GGUF",
    filename="kagemusya-7b-v1Q8_0.gguf",
    local_dir="./models"
)

hf_hub_download(
    repo_id="elyza/Llama-3-ELYZA-JP-8B-GGUF",
    filename="Llama-3-ELYZA-JP-8B-q4_k_m.gguf",
    local_dir="./models"
)


llm = None
llm_model = None

@spaces.GPU(duration=120)
def respond(
    message,
    history: list[tuple[str, str]],
    model,
    system_message,
    max_tokens,
    temperature,
    top_p,
    top_k,
    repeat_penalty,
):
    chat_template = MessagesFormatterType.GEMMA_2

    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 = ""
    try:
        async for output in stream:
            outputs += output
            yield outputs
    except StopAsyncIteration:
        pass  # Handle end of iteration gracefully

description = """<p align="center">Defaults to Oumuamua-7b-RP (you can switch to other models from additional inputs)</p>
<p><center>
<a href="https://huggingface.co/Aratako/Oumuamua-7b-RP-GGUF" target="_blank">[Oumuamua-7b-RP Model]</a>
<a href="https://huggingface.co/bartowski/Oumuamua-7b-instruct-v2-GGUF" target="_blank">[Oumuamua-7b-instruct-v2 Model]</a>
<a href="https://huggingface.co/mmnga/umiyuki-Umievo-itr012-Gleipnir-7B-gguf" target="_blank">[Umievo-itr012-Gleipnir-7B Model]</a>
<a href="https://huggingface.co/Local-Novel-LLM-project/Ninja-V3-GGUF" target="_blank">[Ninja-V3 Model]</a>
<a href="https://huggingface.co/Local-Novel-LLM-project/Kagemusya-7B-v1-GGUF" target="_blank">[Kagemusya-7B-v1 Model]</a>
<a href="https://huggingface.co/elyza/Llama-3-ELYZA-JP-8B-GGUF" target="_blank">[Llama-3-ELYZA-JP-8B Model]</a>
</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([
                'Oumuamua-7b-RP_Q4_K_M.gguf',
                'Oumuamua-7b-instruct-v2-Q4_K_M.gguf',
                'umiyuki-Umievo-itr012-Gleipnir-7B-Q4_K_M.gguf',
                'Ninja-V3-Q4_K_M.gguf',
                'kagemusya-7b-v1Q8_0.gguf',
                'Llama-3-ELYZA-JP-8B-q4_k_m.gguf'
            ],
            value="Oumuamua-7b-RP_Q4_K_M.gguf",
            label="Model"
        ),
        gr.Dropdown(
            choices=templates,
            value="LLAMA_2",
            label="Template"
        ),
        gr.Textbox(value="You are a helpful assistant.", 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",
        ),
    ],
    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()