app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer
import time
import numpy as np
from torch.nn import functional as F
import os
from threading import Thread

print(f"Starting to load the model to memory")
tok = AutoTokenizer.from_pretrained("stabilityai/stablelm-2-zephyr-1_6b", trust_remote_code=True)
tok.padding_side = "left"
tok.pad_token_id = tokenizer.eos_token_id
# using CUDA for an optimal experience
slot="<slot_for_user_input_design_by_xm>"
chat=[{"role": "user", "content": slot}]
sample_input = tok.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
input_start_id=sample_input.find(slot)
prefix=sample_input[:input_start_id]
suffix=sample_input[input_start_id+len(slot):]
print(tok.encode(prefix,return_tensors="pt")[0])
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
m = AutoModelForCausalLM.from_pretrained(
    "stabilityai/stablelm-2-zephyr-1_6b", torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, trust_remote_code=True)
embedding_func=m.get_input_embeddings()
embedding_func.weight.requires_grad=False
m = m.to(device)
print(f"Sucessfully loaded the model to the memory")

start_message = ""

def user(message, history):
    # Append the user's message to the conversation history
    return "", history + [[message, ""]]

#def defense(message):
    # to determine whether the query is malicious

def chat(message, history):
    chat = []
    for item in history:
        chat.append({"role": "user", "content": item[0]})
        if item[1] is not None:
            chat.append({"role": "assistant", "content": item[1]})
    chat.append({"role": "user", "content": message})
    messages = tok.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
    # Tokenize the messages string
    model_inputs = tok([messages], return_tensors="pt").to(device)
    streamer = TextIteratorStreamer(
        tok, timeout=10., skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = dict(
        model_inputs,
        streamer=streamer,
        max_new_tokens=1024,
        do_sample=True,
        top_p=0.90,
        temperature=0.6,
        num_beams=1
    )
    t = Thread(target=m.generate, kwargs=generate_kwargs)
    t.start()

    # Initialize an empty string to store the generated text
    partial_text = ""
    for new_text in streamer:
        print(new_text)
        partial_text += new_text
        # Yield an empty string to cleanup the message textbox and the updated conversation history
        yield partial_text

demo = gr.ChatInterface(fn=chat, examples=["hello", "hola", "merhaba"], title="Gradient Cuff Vicuna-7B-V1.5")
demo.launch()