app.py

# c2-standard-8		spot 9ct/h
# sudo apt-get install git git-lfs pip cmake podman
# git lfs install
#conda
# wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
# bash Miniconda3-latest-Linux-x86_64.sh
# conda create --name dev python=3.10 
# conda activate dev
# conda create --name dev4 python=3.10

##########
# git clone https://huggingface.co/spaces/TobDeBer/Qwen-2-llamacpp
# pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
# pip install huggingface_hub scikit-build-core llama-cpp-agent
# 
import llama_cpp
import os
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

huggingface_token = os.getenv("HUGGINGFACE_TOKEN")

hf_hub_download(
    repo_id="Qwen/Qwen2-0.5B-Instruct-GGUF",
    filename="qwen2-0_5b-instruct-q4_k_m.gguf",
    local_dir="./models"
)

hf_hub_download(
    repo_id="TobDeBer/gpt2-Q4_K_M-GGUF",
    filename="gpt2-q4_k_m.gguf",
    local_dir="./models"
)

hf_hub_download(
    repo_id="TobDeBer/Meta-Llama-3.1-8B-Instruct-Q4_K_M-GGUF",
    filename="meta-llama-3.1-8b-instruct-q4_k_m.gguf",
    local_dir="./models",
    token=huggingface_token
)
# 5GB


# RichardErkhov/ibm-granite_-_granite-7b-base-gguf
# granite-7b-base.Q4_K_M.gguf
# 4GB

# TobDeBer/granite-8b-code-instruct-128k-Q4_K_M-GGUF
# granite-8b-code-instruct-128k-q4_k_m.gguf
# 5GB

llm = None
llm_model = None

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 = ""
    for output in stream:
        outputs += output
        yield outputs

description = """<p align="center">Defaults to Qwen 500M<br>
More models in Advanced Section <br></p>
"""

demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Dropdown([
                'qwen2-0_5b-instruct-q4_k_m.gguf',
				'gpt2-q4_k_m.gguf',
                'meta-llama-3.1-8b-instruct-q4_k_m.gguf',
            ],
            value="qwen2-0_5b-instruct-q4_k_m.gguf",
            label="Model"
        ),
        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 Qwen 2 and friends using llama.cpp", 
    description=description,
    chatbot=gr.Chatbot(
        scale=1, 
        show_copy_button=True
    )
)

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