app.py

import gradio as gr
import torch
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    TextIteratorStreamer,
    BitsAndBytesConfig,
)
import os
from threading import Thread
import spaces
import time

token = os.environ["HF_TOKEN"]

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16
)

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3-8B-Instruct", quantization_config=quantization_config, token=token
)
tok = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct", token=token)

if torch.cuda.is_available():
    device = torch.device("cuda")
    print(f"Using GPU: {torch.cuda.get_device_name(device)}")
else:
    device = torch.device("cpu")
    print("Using CPU")

# model = model.to(device)
# Dispatch Errors


@spaces.GPU
def chat(message, history, temperature, top_p, top_k, max_tokens):
    start_time = time.time()
    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)
    model_inputs = tok([messages], return_tensors="pt").to(device)
    streamer = TextIteratorStreamer(
        tok, timeout=10.0, skip_prompt=True, skip_special_tokens=True
    )
    generate_kwargs = dict(
        model_inputs,
        streamer=streamer,
        max_new_tokens=max_tokens,
        do_sample=True,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    partial_text = ""
    first_token_time = None
    for new_text in streamer:
        if not first_token_time:
            first_token_time = time.time() - start_time
        partial_text += new_text
        yield partial_text

    total_time = time.time() - start_time
    tokens = len(tok.tokenize(partial_text))
    tokens_per_second = tokens / total_time if total_time > 0 else 0

    timing_info = f"\n\nTime taken to first token: {first_token_time:.2f} seconds\nTokens per second: {tokens_per_second:.2f}"
    gr.Info(timing_info)
    yield partial_text 


demo = gr.ChatInterface(
    fn=chat,
    examples=[["Write me a poem about Machine Learning."]],
    # multimodal=True,
    additional_inputs_accordion=gr.Accordion(
        label="⚙️ Parameters", open=False, render=False
    ),
    additional_inputs=[
        gr.Slider(
            minimum=0, maximum=1, step=0.1, value=0.9, label="Temperature", render=False
        ),
        gr.Slider(
            minimum=0, maximum=1, step=0.1, value=0.95, label="top_p", render=False
        ),
        gr.Slider(
            minimum=1, maximum=10000, step=5, value=1000, label="top_k", render=False
        ),
        gr.Slider(
            minimum=128,
            maximum=4096,
            step=1,
            value=1024,
            label="Max new tokens",
            render=False,
        ),
    ],
    stop_btn="Stop Generation",
    title="Chat With LLMs",
    description="Now Running [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) in 4bit"
)
demo.launch()