app.py

import gradio as gr
import codecs
from datetime import datetime
import gc
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":None,"model_name":None}
output = {MODEL_NAMES[0]:"",MODEL_NAMES[1]:""}


print (DEVICE)

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

def infer(
        prompt,
        model_idx = 0,
        max_new_tokens=10,
        temperature=0.1,
        top_p=1.0,
        repetition_penalty = 1.0,
        stop="\n",
        num_completions=1,
        seed=42,
):
    global output
    global models

    print("Loading Models\n")
    model_name = MODEL_NAMES[model_idx]
    if (models["model_name"] == None or models["model_name"] != model_name):
        models = {"model":None,"model_name":None}
        gc.collect()
        if (DEVICE == "cuda"):
            torch.cuda.empty_cache()
        tokenizer = BloomTokenizerFast.from_pretrained(model_name)
        model = DistributedBloomForCausalLM.from_pretrained(model_name, torch_dtype=TORCH_DTYPE, request_timeout=300)
        model = model.to(DEVICE)
        models["model"] = tokenizer, model
        models["model_name"] = model_name
    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")
    
    flag = False
    token_cnt = 0
    with models["model"][1].inference_session(max_length=512) as sess:
        print(f"Encode Input Prompt")
        output[model_name] = ""
        inputs = models["model"][0](prompt, return_tensors="pt")["input_ids"].to(DEVICE)
        n_input_tokens = inputs.shape[1]
        done = False
        print(f"Start Inference ({sess})")
        while not done:
            outputs = models["model"][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"][0].decode(outputs[0, n_input_tokens:])
            token_cnt += 1
            print("\n["+ str(model_name) + "]" + output[model_name], end="", flush=True)
            yield output[model_name]
            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)")
                    done = True
            if flag or (token_cnt >= max_new_tokens):
                print(f"\nDONE (max tokens)")
                done = True
            inputs = None  # Prefix is passed only for the 1st token of the bot's response
            n_input_tokens = 0
        print(f"\nEnd")
    yield output[model_name]

examples = [
    [
        # Question Answering
        '''Please answer the following question:
Question: What is the capital of Germany?
Answer:''',"BLOOMZ" , 3, 0.2, 1.0, 1.0, "\\n,</s>", ["BLOOM","BLOOMZ"]],
    [
        # Chatbot 1
        '''This is a conversation between Alex (an AI based on the 2020 GPT-3 language model), and Fritz (an AI based on the 2021 Jurassic-1 language model). They are exploring each other's capabilities, and trying to ask interesting, complex, and 'ungoogleable' questions of one another, to test the limits of the AI...
Alex: Good morning, Fritz!
Fritz:''',"BLOOM" , 160, 0.85, 0.9, 1.0, "\\n\\n,</s>"],
    [
        # Chatbot 1
        '''This is a conversation between Alex (an AI based on the 2020 GPT-3 language model), and Fritz (an AI based on the 2021 Jurassic-1 language model). They are exploring each other's capabilities, and trying to ask interesting, complex, and 'ungoogleable' questions of one another, to test the limits of the AI...
Alex: Good morning, Fritz!
Fritz:''',"BLOOMZ" , 160, 0.85, 0.9, 1.0, "\\n\\n,</s>"],
    [
        # Expert Answers 
        '''Expert Questions & Helpful Answers
Ask Research Experts
Question:
Are humans good or bad?

Full Answer:''',"BLOOM" , 120, 0.85, 0.9, 1.0, "</s>"],
    [
        # G
        '''You are the writing assistant for Stephen King. You have worked in the fiction/horror genre for 30 years. You are a Pulitzer Prize-winning author, and now you are tasked with developing a skeletal outline for his newest novel, set to be completed in the spring of 2024. Create a title and brief description for the first 5 chapters of this work.\n\nTitle:''',"BLOOM" , 120, 0.85, 0.9, 1.0, "</s>"
    ]
]



iface = gr.Interface(
    fn=infer,
    allow_flagging="never",
    inputs=[
        gr.Textbox(lines=20,label="Input Prompt", max_lines=10),  # prompt
        gr.Radio(["BLOOM","BLOOMZ"], value="BLOOM", type="index", label="Choose 176 billion parameter Model"),
        gr.Slider(1, 256, value=15),  # max_tokens
        gr.Slider(0.0, 1.0, value=0.2),  # temperature
        gr.Slider(0.0, 1.0, value=0.9),  # top_p
        gr.Slider(0.9, 3.0, value=1.0),  # repetition penalty
        gr.Textbox(lines=1, value="\\n\\n,</s>") # stop
    ],
    outputs=gr.Textbox(lines=20, label="Generated Output:"),
    
    examples=examples,
    cache_examples=False,
    title="BLOOM vs BLOOMZ",
    description='''<p>Compare outputs of the BLOOM and BLOOMZ 176 billion parameter models using the Petals network. <b>WARNING:</b> Initial inference may take a long time. Keep the input prompt to a minimum size to speed things up.<p>
    <p>Please consider contributing your unused GPU cycles to the <a href='https://github.com/bigscience-workshop/petals#connect-your-gpu-and-increase-petals-capacity'>petals swarm</a> to help speed up inference. Check the <a href='http://health.petals.ml/'>Health</a> of the Petals Swarm.</p>
    <p>Big thanks to <a href='https://www.rftcapital.com/'>RFT Capital</a> for providing initial compute resources.</p>'''
)

iface.queue(concurrency_count=2)
iface.launch()