app.py

import gradio as gr

import copy
import random
import os
import requests
import time
import sys

from huggingface_hub import snapshot_download
from llama_cpp import Llama

repo_name = "kirp/TinyLlama-1.1B-Chat-v0.2-gguf"
# model_name = "ggml-model-q4_k_m.gguf"
model_name = "ggml-model-q2_k.gguf"


snapshot_download(repo_id=repo_name, local_dir=".", allow_patterns=model_name)

model = Llama(
    model_path=model_name,
    n_ctx=2048,
    n_parts=1,
)


template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"

def generate(
    input=None,
    temperature=0.1,
    top_p=0.75,
    top_k=40,
    max_new_tokens=512,
):
    prompt = template.format(input)
    output = model(prompt, 
                   temperature = temperature,
                   top_k = top_k,
                   top_p = top_p,
                   max_tokens = max_new_tokens + len(input),
                   stop=["<|im_end|>"], 
                   echo=True)
    output = output["choices"][0]['text']
    return output.split("assistant\n")[1]

g = gr.Interface(
    fn=generate,
    inputs=[
        gr.components.Textbox(
            lines=2, label="Prompt", placeholder="Tell me about huggingface."
        ),
        gr.components.Slider(minimum=0, maximum=1, value=0.7, label="Temperature"),
        gr.components.Slider(minimum=0, maximum=1, value=0.8, label="Top p"),
        gr.components.Slider(minimum=0, maximum=100, step=1, value=50, label="Top k"),
        gr.components.Slider(
            minimum=1, maximum=2048, step=1, value=512, label="Max tokens"
        ),
    ],
    outputs=[
        gr.Textbox(
            lines=10,
            label="Output",
        )
    ],
    title = "TinyLlama 1.1B Chat GGUF",
    description =  """
                    original model: [PY007/TinyLlama-1.1B-Chat-v0.2](https://huggingface.co/PY007/TinyLlama-1.1B-Chat-v0.2)
                    quantized_model: [kirp/TinyLlama-1.1B-Chat-v0.2-gguf](https://huggingface.co/kirp/TinyLlama-1.1B-Chat-v0.2-gguf)
                    """
)
g.queue(concurrency_count=2)
g.launch()