app.py

import os
import math
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
import gradio as gr
import sentencepiece

title = "Welcome to Tonic's 🐋🐳Orca-2-13B (in 8bit)!"
description = "You can use [🐋🐳microsoft/Orca-2-13b](https://huggingface.co/microsoft/Orca-2-13b) via API using Gradio by scrolling down and clicking Use 'Via API' or privately by [cloning this space on huggingface](https://huggingface.co/spaces/Tonic1/TonicsOrca2?duplicate=true) . [Join my active builders' server on discord](https://discord.gg/VqTxc76K3u). Big thanks to the HuggingFace Organisation for the Community Grant."

# os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:50'
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model_name = "microsoft/Orca-2-13b"
# offload_folder = './model_weights'

# if not os.path.exists(offload_folder):
#     os.makedirs(offload_folder)

tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
model = transformers.AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_8bit=True)

class OrcaChatBot:
    def __init__(self, model, tokenizer, system_message="You are Orca, an AI language model created by Microsoft. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior."):
        self.model = model
        self.tokenizer = tokenizer
        self.system_message = system_message
        self.conversation_history = []

    def update_conversation_history(self, user_message, assistant_message):
        self.conversation_history.append(("user", user_message))
        self.conversation_history.append(("assistant", assistant_message))

    
    def format_prompt(self):
        prompt = f"<|im_start|>assistant\n{self.system_message}<|im_end|>\n"
        for role, message in self.conversation_history:
            if message.strip():
                prompt += f"<|im_start|>{role}\n{message}<|im_end|>\n"
#               if role == "assistant":
#                    prompt += f"<|im_end|>\n"
        prompt += "<|im_start|> assistant\n"
        return prompt

    def predict(self, user_message, temperature=0.4, max_new_tokens=70, top_p=0.99, repetition_penalty=1.9):
        self.update_conversation_history(user_message, "")
        prompt = self.format_prompt()
        inputs = self.tokenizer(prompt, return_tensors='pt', add_special_tokens=False)
        input_ids = inputs["input_ids"].to(self.model.device)

        output_ids = self.model.generate(
            input_ids,
            max_length=input_ids.shape[1] + max_new_tokens,
            temperature=temperature,
            top_p=top_p,
            repetition_penalty=repetition_penalty,
            pad_token_id=self.tokenizer.eos_token_id,
            do_sample=True
    )

        response = self.tokenizer.decode(output_ids[0], skip_special_tokens=True)
        self.update_conversation_history("", response)
        return response
        
Orca_bot = OrcaChatBot(model, tokenizer)

def gradio_predict(user_message, system_message, max_new_tokens, temperature, top_p, repetition_penalty):
    full_message = f"{system_message}\n{user_message}" if system_message else user_message
    return Orca_bot.predict(full_message, temperature, max_new_tokens, top_p, repetition_penalty)

iface = gr.Interface(
    fn=gradio_predict,
    title=title,
    description=description,
    inputs=[
        gr.Textbox(label="Your Message", type="text", lines=3),
        gr.Textbox(label="Introduce a Character Here or Set a Scene (system prompt)", type="text", lines=2),
        gr.Slider(label="Max new tokens", value=420, minimum=25, maximum=2056, step=1),
        gr.Slider(label="Temperature", value=0.1, minimum=0.05, maximum=1.0, step=0.05),
        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"
)

iface.launch()