app.py

import optimum
import transformers
from transformers import AutoConfig, AutoTokenizer, AutoModel, AutoModelForCausalLM

import torch
import gradio as gr
import json
import os
import shutil
import requests


title = "👋🏻Welcome to Tonic's 💫🌠Starling 7B"
description = "You can use [💫🌠Starling 7B](https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha) or duplicate it for local use or on Hugging Face! [Join me on Discord to build together](https://discord.gg/VqTxc76K3u)."
examples = [
    [
        "The following dialogue is a conversation between Emmanuel Macron and Elon Musk:",  # user_message
        "[Emmanuel Macron]: Hello Mr. Musk. Thank you for receiving me today.",  # assistant_message
        0.9,  # temperature
        450,  # max_new_tokens
        0.90,  # top_p
        1.9,  # repetition_penalty
    ]
]

model_name = "berkeley-nest/Starling-LM-7B-alpha"

device = "cuda" if torch.cuda.is_available() else "cpu"
temperature=0.4
max_new_tokens=240
top_p=0.92
repetition_penalty=1.7


tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
model = transformers.AutoModelForCausalLM.from_pretrained(model_name,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    load_in_4bit=True
)
model.eval()

class StarlingBot:
    def __init__(self, system_prompt="The following dialogue is a conversation"):
        self.system_prompt = system_prompt

    def predict(self, user_message, assistant_message, system_prompt, do_sample, temperature=0.4, max_new_tokens=700, top_p=0.99, repetition_penalty=1.9):
        try:
            conversation = f" <s> [INST] {self.system_prompt} [INST]  {assistant_message if assistant_message else ''} </s> [/INST]  {user_message}  </s> "
            input_ids = tokenizer.encode(conversation, return_tensors="pt", add_special_tokens=False)
            input_ids = input_ids.to(device)
            response = model.generate(
                input_ids=input_ids, 
                use_cache=False, 
                early_stopping=False, 
                bos_token_id=model.config.bos_token_id, 
                eos_token_id=model.config.eos_token_id, 
                pad_token_id=model.config.eos_token_id, 
                temperature=temperature, 
                do_sample=True, 
                max_new_tokens=max_new_tokens, 
                top_p=top_p, 
                repetition_penalty=repetition_penalty
            )
            response_text = tokenizer.decode(response[0], skip_special_tokens=True)
#           response_text = response.split("<|assistant|>\n")[-1]
            return response_text       
            finally:
                del input_ids, attention_mask, output_ids
                gc.collect()
                torch.cuda.empty_cache()

starling_bot = StarlingBot()

iface = gr.Interface(
    fn=starling_bot.predict,
    title=title,
    description=description,
    examples=examples,
    inputs=[
        gr.Textbox(label="🌟🤩User Message", type="text", lines=5),
        gr.Textbox(label="💫🌠Starling Assistant Message or Instructions ", lines=2),
        gr.Textbox(label="💫🌠Starling System Prompt or Instruction", lines=2),
        gr.Checkbox(label="Advanced", value=False),
        gr.Slider(label="Temperature", value=0.7, minimum=0.05, maximum=1.0, step=0.05),
        gr.Slider(label="Max new tokens", value=100, minimum=25, maximum=256, step=1),
        gr.Slider(label="Top-p (nucleus sampling)", value=0.90, minimum=0.01, maximum=0.99, step=0.05),
        gr.Slider(label="Repetition penalty", value=1.9, minimum=1.0, maximum=2.0, step=0.05)
    ],
    outputs="text",
    theme="ParityError/Anime"
)