app.py

#!/usr/bin/env python#

from llama_cpp import Llama
from time import time
import gradio as gr
import psutil

# load like this - use tne variable everywhere 
model_path=os.getenv("MODEL_PATH")
# show warning, when empty and briefs description of how to set it 

# Initing things                
print(f"debug: init model: {model_path}")
#llm = Llama(model_path="./model.bin")                             # LLaMa model
print("! INITING DONE !")

# Preparing things to work
title = "# Demo for 7B Models - Quantized"
descr = '''
Quantized to run in the free tier hosting. 
Have a quick way to test models or share them with others without hassle.
It runs slow, as it's on cpu. Usable for basic tests.
It uses quantized models in gguf-Format and llama.cpp to run them.

Powered by ...'''

print(f"DEBUG: Memory free: {psutil.virtual_memory().free / (1024.0 ** 3)} GiB")
print(f"DEBUG: Memory: {psutil.virtual_memory().total / (1024.0 ** 3)} GiB")

import os
from threading import Thread
from typing import Iterator

import gradio as gr
import spaces
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer

DESCRIPTION = "# Mistral-7B"

if torch.cuda.is_available():
    DESCRIPTION += "\n<p>This space is using CPU only. Use a different one if you want to go fast and use GPU. </p>"

MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))


#download model here
# check localstorage, if no there, load, else use existing.
# check gradio - how does it dl? is there a function we can use?

if torch.cuda.is_available():
    model_id = "mistralai/Mistral-7B-Instruct-v0.1"
    model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
    tokenizer = AutoTokenizer.from_pretrained(model_id)

# we need to make sure we only run one thread or we probably run out of ram
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    max_new_tokens: int = 1024,
    temperature: float = 0.6,
    top_p: float = 0.9,
    top_k: int = 50,
    repetition_penalty: float = 1.2,
) -> Iterator[str]:
    conversation = []
    for user, assistant in chat_history:
        conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
    conversation.append({"role": "user", "content": message})

    input_ids = tokenizer.apply_chat_template(conversation, return_tensors="pt")
    if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
        input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
        gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
    input_ids = input_ids.to(model.device)

    streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = dict(
        {"input_ids": input_ids},
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        num_beams=1,
        repetition_penalty=repetition_penalty,
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    outputs = []
    for text in streamer:
        outputs.append(text)
        yield "".join(outputs)


chat_interface = gr.ChatInterface(
    fn=generate,
    additional_inputs=[
        gr.Slider(
            label="Max new tokens",
            minimum=1,
            maximum=MAX_MAX_NEW_TOKENS,
            step=1,
            value=DEFAULT_MAX_NEW_TOKENS,
        ),
        gr.Slider(
            label="Temperature",
            minimum=0.1,
            maximum=4.0,
            step=0.1,
            value=0.6,
        ),
        gr.Slider(
            label="Top-p (nucleus sampling)",
            minimum=0.05,
            maximum=1.0,
            step=0.05,
            value=0.9,
        ),
        gr.Slider(
            label="Top-k",
            minimum=1,
            maximum=1000,
            step=1,
            value=50,
        ),
        gr.Slider(
            label="Repetition penalty",
            minimum=1.0,
            maximum=2.0,
            step=0.05,
            value=1.2,
        ),
    ],
    stop_btn=None,
    # add more eval examples, like a long list taken from teknium and others maybe group by type
    examples=[
        ["Hello there! How are you doing?"],
        ["Can you explain briefly to me what is the Python programming language?"],
        ["Explain the plot of Cinderella in a sentence."],
        ["How many hours does it take a man to eat a Helicopter?"],
        ["Write a 100-word article on 'Benefits of Open-Source in AI research'"],
    ],
)

with gr.Blocks(css="style.css") as demo:
    gr.Markdown(title)
    gr.Markdown(descr)
    gr.DuplicateButton(
        value="Duplicate Space for private use",
        elem_id="duplicate-button",
        visible=os.getenv("SHOW_DUPLICATE_BUTTON") == "1",
        # add 
    )
    chat_interface.render()

if __name__ == "__main__":
    demo.queue(max_size=20).launch()