app_gguf.py

import spaces
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

import llama_cpp
print(llama_cpp.__file__)
print(llama_cpp.__version__)

huggingface_token = os.getenv("HF_TOKEN")

hf_hub_download(
    repo_id="unsloth/Qwen3-Coder-Next-GGUF",
    filename="Qwen3-Coder-Next-Q4_K_S.gguf",

    local_dir="./models"
)

# hf_hub_download(
#     repo_id="bartowski/gemma-2-27b-it-GGUF",
#     filename="gemma-2-27b-it-Q5_K_M.gguf",
#     local_dir="./models"
# )

# hf_hub_download(
#     repo_id="google/gemma-2-2b-it-GGUF",
#     filename="2b_it_v2.gguf",
#     local_dir="./models",
#     token=huggingface_token
# )

# hf_hub_download(
#     repo_id="unsloth/Qwen3-Coder-Next-GGUF",
#     filename="Qwen3-Coder-Next-IQ4_NL.gguf",
#     local_dir="./models",
#     token=huggingface_token
# )




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


import gradio as gr

demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Dropdown(
            [
               
                "unsloth/Qwen3-Coder-Next-GGUF"
            ],
            value="Qwen Coder",
            label="Model",
        ),
        gr.Textbox(
            value="You are a helpful assistant.",
            label="System message",
        ),
        gr.Slider(1, 4096, value=2048, step=1, label="Max tokens"),
        gr.Slider(0.1, 4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p"),
        gr.Slider(0, 100, value=40, step=1, label="Top-k"),
        gr.Slider(0.0, 2.0, value=1.1, step=0.1, label="Repetition penalty"),
    ],
    title="Qwen Coder Next",
    description=description,
)


# demo.launch()

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