app.py

import spaces
from threading import Thread

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

model_id = "facebook/opt-1.3b"
assistant_id = "facebook/opt-125m"

model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
assistant_model = AutoModelForCausalLM.from_pretrained(assistant_id, torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

@spaces.GPU
def run_generation(user_text, use_assistant, temperature, max_new_tokens):
    if temperature < 0.1:
        do_sample = False
    else:
        do_sample = True

    # Get the model and tokenizer, and tokenize the user text.
    model_inputs = tokenizer([user_text], return_tensors="pt").to(model.device)

    # Start generation on a separate thread, so that we don't block the UI. The text is pulled from the streamer
    # in the main thread. Adds timeout to the streamer to handle exceptions in the generation thread.
    streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = dict(
        model_inputs,
        assistant_model=assistant_model if use_assistant else None,
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        do_sample=do_sample,
        top_p=0.95,
        temperature=float(temperature),
        top_k=50,
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    start = time.time()
    t.start()

    # Pull the generated text from the streamer, and update the model output. Return the model output and time
    # spent so far.
    model_output = ""
    for new_text in streamer:
        model_output += new_text
        yield [model_output, round(time.time() - start, 3)]
    return [model_output, round(time.time() - start, 3)]


def reset_textbox():
    return gr.update(value='')


with gr.Blocks() as demo:
    gr.Markdown(
        "# 🤗 Assisted Generation Demo\n"
        f"- Model: {model_id} (BF16, ~3GB)\n"
        f"- Assistant Model: {assistant_id} (BF16, ~0.3GB)\n"
        "- Running on a T4 GPU\n"
        "- Best results are obtained when the model size difference in parameters is >10x"
    )

    with gr.Row():
        with gr.Column(scale=4):
            user_text = gr.Textbox(
                placeholder="A sequence: one, two, three, ",
                label="Prompt"
            )
            model_output = gr.Textbox(label="Model output", lines=10, interactive=False)
            button_submit = gr.Button(value="Submit")

        with gr.Column(scale=1, min_width=200):
            gr.Markdown("### Generation Settings")
            use_assistant = gr.Checkbox(label="Use Assisted Generation", value=True)
            max_new_tokens = gr.Slider(
                minimum=1, maximum=500, value=100, step=1, interactive=True, label="Max New Tokens",
            )
            temperature = gr.Slider(
                minimum=0.0, maximum=2.0, value=0.0, step=0.1, interactive=True, label="Temperature (0.0 = Greedy)",
            )
            gr.Markdown("### Generation time (seconds)")
            generation_time = gr.Textbox(lines=1, interactive=False, show_label=False)

    generate_inputs = [user_text, use_assistant, temperature, max_new_tokens]
    generate_outputs = [model_output, generation_time]
    user_text.submit(run_generation, generate_inputs, generate_outputs)
    button_submit.click(run_generation, generate_inputs, generate_outputs)

    demo.queue(max_size=32).launch()