app.py

import gradio as gr
import os 
from pathlib import Path
import argparse
model_file = "Yi-6B.q4_k_m.gguf"
if not os.path.isfile("Yi-6B.q4_k_m.gguf"):
    os.system("wget -c https://huggingface.co/SamPurkis/Yi-6B-GGUF/resolve/main/Yi-6B.q4_k_m.gguf")

DEFAULT_MODEL_PATH = model_file
parser = argparse.ArgumentParser()
parser.add_argument("-m", "--model", default=DEFAULT_MODEL_PATH, type=Path, help="model path")
parser.add_argument("--mode", default="chat", type=str, choices=["chat", "generate"], help="inference mode")
parser.add_argument("-l", "--max_length", default=2048, type=int, help="max total length including prompt and output")
parser.add_argument("-c", "--max_context_length", default=512, type=int, help="max context length")
parser.add_argument("--top_k", default=0, type=int, help="top-k sampling")
parser.add_argument("--top_p", default=0.7, type=float, help="top-p sampling")
parser.add_argument("--temp", default=0.95, type=float, help="temperature")
parser.add_argument("--repeat_penalty", default=1.1, type=float, help="penalize repeat sequence of tokens")
parser.add_argument("-t", "--threads", default=0, type=int, help="number of threads for inference")
parser.add_argument("--plain", action="store_true", help="display in plain text without markdown support")
args = parser.parse_args()

from llama_cpp import Llama
llm = Llama(model_path=model_file)



def predict(input, system_prompt, chatbot, max_length, ctx_length, top_p, temperature, history):
    chatbot.append((input, ""))
    response = ""
    history.append(input)

    generation_kwargs = dict(
        max_length=max_length,
        max_context_length=ctx_length,
        do_sample=temperature > 0,
        top_k=40,
        top_p=top_p,
        temperature=temperature,
        repetition_penalty=1.1,
        num_threads=0,
        stream=True,
    )
    output = llm(input)
    response = output['choices'][0]['text']

    for response_piece in response:
        response += response_piece
        chatbot[-1] = (chatbot[-1][0], response)

        yield chatbot, history

    history.append(response)
    yield chatbot, history


def reset_user_input():
    return gr.update(value="")


def reset_state():
    return [], []


with gr.Blocks() as demo:
    gr.HTML("""<h1 align="center">01-Yi 6B</h1>""")

    chatbot = gr.Chatbot()
    with gr.Row():
        with gr.Column(scale=4):
            system_prompt = gr.Textbox(show_label=False, placeholder="system prompt ...", lines=2)
            user_input = gr.Textbox(show_label=False, placeholder="Input...", lines=6)
            submitBtn = gr.Button("Submit", variant="primary")
        with gr.Column(scale=1):
            max_length = gr.Slider(0, 32048, value=args.max_length, step=1.0, label="Maximum Length", interactive=True)
            ctx_length = gr.Slider(0, 4096, value=512, step=1.0, label="Maximum Context Length", interactive=True)
            top_p = gr.Slider(0, 1, value=args.top_p, step=0.01, label="Top P", interactive=True)
            temperature = gr.Slider(0, 1, value=args.temp, step=0.01, label="Temperature", interactive=True)
            emptyBtn = gr.Button("Clear History")

    history = gr.State([])

    submitBtn.click(
        predict, [user_input, system_prompt, chatbot, max_length, ctx_length, top_p, temperature, history], [chatbot, history], show_progress=True
    )
    submitBtn.click(reset_user_input, [], [user_input])

    emptyBtn.click(reset_state, outputs=[chatbot, history], show_progress=True)

demo.queue().launch(share=False, inbrowser=True)