app.py

import spaces
import json
import subprocess
import gradio as gr
from huggingface_hub import hf_hub_download

subprocess.run('pip install llama-cpp-python==0.2.75 --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124', shell=True)
subprocess.run('pip install llama-cpp-agent==0.2.10', shell=True)

hf_hub_download(
    repo_id="bartowski/Meta-Llama-3-70B-Instruct-GGUF",
    filename="Meta-Llama-3-70B-Instruct-Q3_K_M.gguf",
    local_dir = "./models"
)
hf_hub_download(
    repo_id="bartowski/Llama-3-8B-Synthia-v3.5-GGUF",
    filename="Llama-3-8B-Synthia-v3.5-f16.gguf",
    local_dir = "./models"
)
hf_hub_download(
    repo_id="bartowski/Mistral-7B-Instruct-v0.3-GGUF",
    filename="Mistral-7B-Instruct-v0.3-f32.gguf",
    local_dir = "./models"
)

css = """
.message-row {
    justify-content: space-evenly !important;
}
.message-bubble-border {
    border-radius: 6px !important;
}
.dark.message-bubble-border {
    border-color: #343140 !important;
}
.dark.user {
    background: #1e1c26 !important;
}
.dark.assistant.dark, .dark.pending.dark {
    background: #16141c !important;
}
"""

def get_messages_formatter_type(model_name):
    from llama_cpp_agent import MessagesFormatterType
    if "Llama" in model_name:
        return MessagesFormatterType.LLAMA_3
    elif "Mistral" in model_name:
        return MessagesFormatterType.MISTRAL
    else:
        raise ValueError(f"Unsupported model: {model_name}")

@spaces.GPU(duration=120)
def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
    top_k,
    repeat_penalty,
    model,
):
    from llama_cpp import Llama
    from llama_cpp_agent import LlamaCppAgent
    from llama_cpp_agent.providers import LlamaCppPythonProvider
    from llama_cpp_agent.chat_history import BasicChatHistory
    from llama_cpp_agent.chat_history.messages import Roles

    chat_template = get_messages_formatter_type(model)

    llm = Llama(
        model_path=f"models/{model}",
        flash_attn=True,
        n_threads=40,
        n_gpu_layers=81,
        n_batch=1024,
        n_ctx=8192,
    )
    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

PLACEHOLDER = """
<div class="container" style="max-width: 600px; margin: 0 auto; padding: 30px; background-color: transparent; box-shadow: 0 0 10px rgba(0, 0, 0, 0.1);">
    <h1 style="font-size: 28px; margin-bottom: 15px;">llama-cpp-agent: Simplify LLM Interactions</h1>
    <p style="font-size: 16px; line-height: 1.5; margin-bottom: 15px;">The llama-cpp-agent framework simplifies interactions with Large Language Models (LLMs), providing an interface for chatting, executing function calls, generating structured output, performing retrieval augmented generation, and processing text using agentic chains with tools.</p>
    <p style="font-size: 16px; line-height: 1.5; margin-bottom: 15px;">The framework uses guided sampling to constrain model output to user-defined structures, enabling models not fine-tuned for function calling and JSON output to do so. It is compatible with llama.cpp server, llama-cpp-python and its server, TGI, and vllm servers.</p>
</div>
"""

demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        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",
        ),
        gr.Dropdown([
                'Meta-Llama-3-70B-Instruct-Q3_K_M.gguf',
                'Llama-3-8B-Synthia-v3.5-f16.gguf',
                'Mistral-7B-Instruct-v0.3-f32.gguf'
            ],
            value="Meta-Llama-3-70B-Instruct-Q3_K_M.gguf",
            label="Model"
        ),
    ],
    theme=gr.themes.Soft(primary_hue="violet", secondary_hue="violet", neutral_hue="gray",font=[gr.themes.GoogleFont("Exo"), "ui-sans-serif", "system-ui", "sans-serif"]).set(
        body_background_fill_dark="#16141c",
        block_background_fill_dark="#16141c",
        block_border_width="1px",
        block_title_background_fill_dark="#1e1c26",
        input_background_fill_dark="#292733",
        button_secondary_background_fill_dark="#24212b",
        border_color_primary_dark="#343140",
        background_fill_secondary_dark="#16141c",
        color_accent_soft_dark="transparent"
    ),
    css=css,
    retry_btn="Retry",
    undo_btn="Undo",
    clear_btn="Clear",
    submit_btn="Send",
    description="Llama-cpp-agent: Chat multi llm selection",
    chatbot=gr.Chatbot(scale=1, placeholder=PLACEHOLDER)
)

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