app.py

import gradio as gr
import threading
import codecs
from datetime import datetime
from transformers import BloomTokenizerFast
from petals.client import DistributedBloomForCausalLM
import torch
import time

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
TORCH_DTYPE = torch.bfloat16
MODEL_NAMES = ["bigscience/bloom-petals", "bigscience/bloomz-petals"]

models = {MODEL_NAMES[0]:None,MODEL_NAMES[1]:None}
output = {MODEL_NAMES[0]:"",MODEL_NAMES[1]:""}
kill = threading.Event()

def stop_threads():
    global kill
    print("Force stopping threads")
    kill.set()

def gen_thread(model_name, prompt, max_tokens, temperature, top_p, repetition_penalty, stop):
    global output
    
    if kill.is_set():
        return
    
    flag = False
    token_cnt = 0
    with models[model_name][1].inference_session(max_length=512) as sess:
        print(f"Thread Start -> {threading.get_ident()}")
        output[model_name] = ""
        inputs = models[model_name][0](prompt, return_tensors="pt")["input_ids"].to(DEVICE)
        n_input_tokens = inputs.shape[1]
        done = False
        while not done and not kill.is_set():
            outputs = models[model_name][1].generate(
                inputs, 
                max_new_tokens=1, 
                do_sample=True, 
                top_p=top_p, 
                temperature=temperature, 
                repetition_penalty=repetition_penalty,
                session=sess
            )
            output[model_name] += models[model_name][0].decode(outputs[0, n_input_tokens:])
            token_cnt += 1
            print("\n["+ str(threading.get_ident()) + "]" + output[model_name], end="", flush=True)

            for stop_word in stop:
                stop_word = codecs.getdecoder("unicode_escape")(stop_word)[0]
                if stop_word != '' and stop_word in output[model_name]:
                    print(f"\nDONE (stop) -> {threading.get_ident()}")
                    done = True
            if flag or (token_cnt >= max_tokens):
                print(f"\nDONE (max tokens) -> {threading.get_ident()}")
                done = True
            inputs = None  # Prefix is passed only for the 1st token of the bot's response
            n_input_tokens = 0
        print(f"\nThread End -> {threading.get_ident()}")

def to_md(text):
    return text.replace("\n", "<br />")

threads = list()

def infer(
        prompt,
        model_idx = ["BLOOM","BLOOMZ"],
        max_new_tokens=10,
        temperature=0.1,
        top_p=1.0,
        repetition_penalty = 1.0,
        stop="\n",
        num_completions=1,
        seed=42,
):
    global threads
    global output
    global models

    if len(model_idx) == 0:
        return

    kill.clear()    
    print("Loading Models\n")
    for idx in model_idx:
        model_name = MODEL_NAMES[idx]
        if models[model_name] == None:
            print ("Initializing " + model_name)
            tokenizer = BloomTokenizerFast.from_pretrained(model_name)
            model = DistributedBloomForCausalLM.from_pretrained(model_name, torch_dtype=TORCH_DTYPE)
            model = model.to(DEVICE)
            models[model_name] = tokenizer, model
            output[model_name] = ""

    max_new_tokens = int(max_new_tokens)
    temperature = float(temperature)
    top_p = float(top_p)
    stop =  [x.strip(' ') for x in stop.split(',')]
    repetition_penalty = float(repetition_penalty)
    seed = seed

    assert 1 <= max_new_tokens <= 384
    assert 1 <= num_completions <= 5
    assert 0.0 <= temperature <= 1.0
    assert 0.0 <= top_p <= 1.0
    assert 0.9 <= repetition_penalty <= 3.0

    if temperature == 0.0:
        temperature = 0.01
    if prompt == "":
        prompt = " "
    
    print(f"START -> ({datetime.now()})\n")
    print(f"PROMPT ({datetime.now()}):\n-------\n{prompt}\n")
    for idx in model_idx:
        model_name = MODEL_NAMES[idx]
        x = threading.Thread(target=gen_thread, args=(model_name, prompt, max_new_tokens, temperature, top_p, repetition_penalty, stop))
        threads.append(x)
        x.start()
    
    # Join Threads
    for model_name, thread in enumerate(threads):
        while thread.is_alive():
            thread.join(timeout=0.2)
            yield output[MODEL_NAMES[0]], output[MODEL_NAMES[1]]


examples = [
    [
        # Question Answering
        '''Please answer the following question:
Question: What is the capital of Germany?
Answer:''',["BLOOM","BLOOMZ"] , 3, 0.2, 1.0, 1.0, "\\n,</s>", ["BLOOM","BLOOMZ"]],
    [
        # Natural Language Interface
        '''Given a pair of sentences, choose whether the two sentences agree (entailment)/disagree (contradiction) with each other.
Possible labels: 1. entailment 2. contradiction
Sentence 1: The skier was on the edge of the ramp. Sentence 2: The skier was dressed in winter clothes.
Label: entailment
Sentence 1: The boy skated down the staircase railing. Sentence 2: The boy is a newbie skater.
Label: contradiction
Sentence 1: Two middle-aged people stand by a golf hole. Sentence 2: A couple riding in a golf cart.
Label:''',["BLOOM","BLOOMZ"] , 2, 0.2, 1.0, 1.0, "\\n,</s>"]
]

def clear_prompt():
    return "","",""

with gr.Blocks() as demo:
    gr.Markdown("# <p style='text-align: center;'>BLOOM vs BLOOMZ Comparison</p>")
    gr.Markdown("")
    gr.Markdown("Test Inference on the [BLOOM](https://huggingface.co/bigscience/bloom) and [BLOOMZ](https://huggingface.co/bigscience/bloomz) 176 Billion Parameter models using Petals.  \
        Please consider contributing your unused GPU cycles to the [Petals Swarm](https://github.com/bigscience-workshop/petals) to speed up inference. <br />\n \
        Due to heavy resource requirements of these large models, token generation can take upwards of 3-5 seconds per token. Try to keep Max Tokens to a minimum.")
    gr.Markdown("")
    gr.Markdown("Special thanks to [RFT Capital](https://www.rftcapital.com/) for supporting our experiments with compute time dontations.")
    gr.Markdown("Type a Prompt and then click **Run** to see the output.")
    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(lines=17,label="Prompt",placeholder="Enter Prompt", interactive=True)
            with gr.Box():
                chk_boxes = gr.CheckboxGroup(choices=["BLOOM","BLOOMZ"],value=["BLOOM","BLOOMZ"], type="index", label="Model")
                #min_length = gr.Slider(minimum=0, maximum=256, value=1, label="Minimum Length") #min_length
                max_tokens = gr.Slider(minimum=1, maximum=256, value=15, label="Max Tokens")  # max_tokens
                temperature = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=0.2, label="Temperature")  # temperature
                top_p = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=0.9, label="Top P")  # top_p
                rep_penalty = gr.Slider(minimum=0.9, maximum=3.0, step=0.1, value=1.0, label="Repetition Penalty")  # repetition penalty
                stop = gr.Textbox(lines=1, value="\\n,</s>", label="Stop Token") # stop
        with gr.Column():
            bloom_out = gr.Textbox(lines=7, label="BLOOM OUTPUT:") 
            bloomz_out = gr.Textbox(lines=7,label="BLOOMZ OUTPUT:")
    with gr.Row():
        btn_clear = gr.Button("Clear", variant="secondary")
        btn_run = gr.Button("Run", variant="primary")
        btn_stop = gr.Button("Stop", variant="stop")
        click_run = btn_run.click(infer, inputs=[prompt, chk_boxes, max_tokens, temperature, top_p, rep_penalty, stop], outputs=[bloom_out,bloomz_out])   
        btn_clear.click(clear_prompt, outputs=[prompt, bloom_out, bloomz_out])
        btn_stop.click(stop_threads,cancels=click_run)
    gr.Examples(examples, inputs=[prompt, chk_boxes, max_tokens, temperature, top_p, rep_penalty, stop])

demo.queue(concurrency_count=1)
demo.launch()