app.py

import os
import gradio as gr
import copy
from llama_cpp import Llama
from huggingface_hub import hf_hub_download

# Initialize Llama model from Hugging Face
llm = Llama(
    model_path=hf_hub_download(
        repo_id=os.environ.get("REPO_ID", "mradermacher/Atlas-Chat-2B-GGUF"),
        filename=os.environ.get("MODEL_FILE", "Atlas-Chat-2B.Q8_0.gguf"),
    ),
    n_ctx=2048,
)

# Training prompt format for Atlas-Chat style conversation
training_prompt = """<start_of_turn>user
{}<end_of_turn>
<start_of_turn>model
{}<end_of_turn>"""

EOS_TOKEN = "<end_of_turn>"

# Function to generate the text response based on conversation history
def generate_text(
    message,
    history: list[tuple[str, str]],
    max_tokens,
    temperature,
    top_p,
):
    temp = ""
    input_prompt = ""
    
    # Loop through the conversation history and add each turn to the prompt
    for user_input, assistant_response in history:
        input_prompt += training_prompt.format(user_input, assistant_response)
    
    # Add the current message to the prompt
    input_prompt += training_prompt.format(message, "")
    
    # Generate the output using the model
    output = llm(
        input_prompt,
        temperature=temperature,
        top_p=top_p,
        top_k=40,
        repeat_penalty=1.1,
        max_tokens=max_tokens,
        stop=[
            EOS_TOKEN,
            "<|endoftext|>"
        ],
        stream=True,
    )
    
    # Stream and yield the model’s output
    for out in output:
        stream = copy.deepcopy(out)
        temp += stream["choices"][0]["text"]
        yield temp

# Define the Gradio interface
demo = gr.ChatInterface(
    generate_text,
    title="using Atlas-Chat-2B | I had to switch to the 2B model because the 9B was too much for this space!",
    description="Running LLM with https://github.com/abetlen/llama-cpp-python",
    examples=[
        ['How to setup a human base on Mars? Give short answer.'],
        ['Explain theory of relativity to me like I’m 8 years old.'],
        ['شكون لي صنعك؟'],
        ['أشنو كايمييز المملكة المغربية'],
        ['شنو كيتسمى المنتخب المغربي؟']
    ],
    cache_examples=False,
    retry_btn=None,
    undo_btn="Delete Previous",
    clear_btn="Clear",
    additional_inputs=[
        gr.Slider(minimum=1, maximum=768, value=256, step=1, label="Max new 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 (nucleus sampling)",
        ),
    ],
)

# Launch the Gradio demo interface
if __name__ == "__main__":
    demo.launch()