text-generation-webui/server.py

import re
import time
import glob
from sys import exit
import torch
import argparse
import json
from pathlib import Path
import gradio as gr
import transformers
from html_generator import *
from transformers import AutoTokenizer, AutoModelForCausalLM
import warnings
import gc
from tqdm import tqdm


transformers.logging.set_verbosity_error()

parser = argparse.ArgumentParser()
parser.add_argument('--model', type=str, help='Name of the model to load by default.')
parser.add_argument('--notebook', action='store_true', help='Launch the web UI in notebook mode, where the output is written to the same text box as the input.')
parser.add_argument('--chat', action='store_true', help='Launch the web UI in chat mode.')
parser.add_argument('--cai-chat', action='store_true', help='Launch the web UI in chat mode with a style similar to Character.AI\'s. If the file profile.png or profile.jpg exists in the same folder as server.py, this image will be used as the bot\'s profile picture.')
parser.add_argument('--cpu', action='store_true', help='Use the CPU to generate text.')
parser.add_argument('--load-in-8bit', action='store_true', help='Load the model with 8-bit precision.')
parser.add_argument('--auto-devices', action='store_true', help='Automatically split the model across the available GPU(s) and CPU.')
parser.add_argument('--disk', action='store_true', help='If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk.')
parser.add_argument('--max-gpu-memory', type=int, help='Maximum memory in GiB to allocate to the GPU when loading the model. This is useful if you get out of memory errors while trying to generate text. Must be an integer number.')
parser.add_argument('--no-listen', action='store_true', help='Make the web UI unreachable from your local network.')
parser.add_argument('--no-stream', action='store_true', help='Don\'t stream the text output in real time. This slightly improves the text generation performance.')
parser.add_argument('--settings', type=str, help='Load the default interface settings from this json file. See settings-template.json for an example.')
args = parser.parse_args()

loaded_preset = None
available_models = sorted(set([item.replace('.pt', '') for item in map(lambda x : str(x.name), list(Path('models/').glob('*'))+list(Path('torch-dumps/').glob('*'))) if not item.endswith('.txt')]), key=str.lower)
available_presets = sorted(set(map(lambda x : str(x.name).split('.')[0], Path('presets').glob('*.txt'))), key=str.lower)

settings = {
    'max_new_tokens': 200,
    'max_new_tokens_min': 1,
    'max_new_tokens_max': 2000,
    'preset': 'NovelAI-Sphinx Moth',
    'name1': 'Person 1',
    'name2': 'Person 2',
    'name1_pygmalion': 'You',
    'name2_pygmalion': 'Kawaii',
    'context': 'This is a conversation between two people.',
    'context_pygmalion': 'This is a conversation between two people.\n<START>',
    'prompt': 'Common sense questions and answers\n\nQuestion: \nFactual answer:',
    'prompt_gpt4chan': '-----\n--- 865467536\nInput text\n--- 865467537\n',
    'stop_at_newline': True,
}

if args.settings is not None and Path(args.settings).exists():
    with open(Path(args.settings), 'r') as f:
        new_settings = json.load(f)
    for item in new_settings:
        if item in settings:
            settings[item] = new_settings[item]

def load_model(model_name):
    print(f"Loading {model_name}...")
    t0 = time.time()

    # Default settings
    if not (args.cpu or args.load_in_8bit or args.auto_devices or args.disk or args.max_gpu_memory is not None):
        if Path(f"torch-dumps/{model_name}.pt").exists():
            print("Loading in .pt format...")
            model = torch.load(Path(f"torch-dumps/{model_name}.pt"))
        elif model_name.lower().startswith(('gpt-neo', 'opt-', 'galactica')) and any(size in model_name.lower() for size in ('13b', '20b', '30b')):
            model = AutoModelForCausalLM.from_pretrained(Path(f"models/{model_name}"), device_map='auto', load_in_8bit=True)
        else:
            model = AutoModelForCausalLM.from_pretrained(Path(f"models/{model_name}"), low_cpu_mem_usage=True, torch_dtype=torch.float16).cuda()

    # Custom
    else:
        settings = ["low_cpu_mem_usage=True"]
        command = "AutoModelForCausalLM.from_pretrained"

        if args.cpu:
            settings.append("torch_dtype=torch.float32")
        else:
            settings.append("device_map='auto'")
            if args.max_gpu_memory is not None:
                settings.append(f"max_memory={{0: '{args.max_gpu_memory}GiB', 'cpu': '99GiB'}}")
            if args.disk:
                settings.append("offload_folder='cache'")
            if args.load_in_8bit:
                settings.append("load_in_8bit=True")
            else:
                settings.append("torch_dtype=torch.float16")

        settings = ', '.join(set(settings))
        command = f"{command}(Path(f'models/{model_name}'), {settings})"
        model = eval(command)

    # Loading the tokenizer
    if model_name.lower().startswith(('gpt4chan', 'gpt-4chan', '4chan')) and Path(f"models/gpt-j-6B/").exists():
        tokenizer = AutoTokenizer.from_pretrained(Path("models/gpt-j-6B/"))
    else:
        tokenizer = AutoTokenizer.from_pretrained(Path(f"models/{model_name}/"))
    tokenizer.truncation_side = 'left'

    print(f"Loaded the model in {(time.time()-t0):.2f} seconds.")
    return model, tokenizer

# Removes empty replies from gpt4chan outputs
def fix_gpt4chan(s):
    for i in range(10):
        s = re.sub("--- [0-9]*\n>>[0-9]*\n---", "---", s)
        s = re.sub("--- [0-9]*\n *\n---", "---", s)
        s = re.sub("--- [0-9]*\n\n\n---", "---", s)
    return s

# Fix the LaTeX equations in galactica
def fix_galactica(s):
    s = s.replace(r'\[', r'$')
    s = s.replace(r'\]', r'$')
    s = s.replace(r'\(', r'$')
    s = s.replace(r'\)', r'$')
    s = s.replace(r'$$', r'$')
    return s

def encode(prompt, tokens):
    if not args.cpu:
        torch.cuda.empty_cache()
        input_ids = tokenizer.encode(str(prompt), return_tensors='pt', truncation=True, max_length=2048-tokens).cuda()
    else:
        input_ids = tokenizer.encode(str(prompt), return_tensors='pt', truncation=True, max_length=2048-tokens)
    return input_ids

def decode(output_ids):
    reply = tokenizer.decode(output_ids, skip_special_tokens=True)
    reply = reply.replace(r'<|endoftext|>', '')
    return reply

def formatted_outputs(reply, model_name):
    if model_name.lower().startswith('galactica'):
        reply = fix_galactica(reply)
        return reply, reply, generate_basic_html(reply)
    elif model_name.lower().startswith('gpt4chan'):
        reply = fix_gpt4chan(reply)
        return reply, 'Only applicable for GALACTICA models.', generate_4chan_html(reply)
    else:
        return reply, 'Only applicable for GALACTICA models.', generate_basic_html(reply)

def generate_reply(question, tokens, inference_settings, selected_model, eos_token=None):
    global model, tokenizer, model_name, loaded_preset, preset

    if selected_model != model_name:
        model_name = selected_model
        model = None
        tokenizer = None
        if not args.cpu:
            gc.collect()
            torch.cuda.empty_cache()
        model, tokenizer = load_model(model_name)
    if inference_settings != loaded_preset:
        with open(Path(f'presets/{inference_settings}.txt'), 'r') as infile:
            preset = infile.read()
        loaded_preset = inference_settings

    cuda = "" if args.cpu else ".cuda()"
    n = None if eos_token is None else tokenizer.encode(eos_token, return_tensors='pt')[0][-1]

    # Generate the entire reply at once
    if args.no_stream:
        input_ids = encode(question, tokens)
        output = eval(f"model.generate(input_ids, eos_token_id={n}, {preset}){cuda}")
        reply = decode(output[0])
        yield formatted_outputs(reply, model_name)

    # Generate the reply 1 token at a time
    else:
        yield formatted_outputs(question, model_name)
        input_ids = encode(question, 1)
        preset = preset.replace('max_new_tokens=tokens', 'max_new_tokens=1')
        for i in tqdm(range(tokens)):
            output = eval(f"model.generate(input_ids, {preset}){cuda}")
            reply = decode(output[0])
            if eos_token is not None and reply[-1] == eos_token:
                break

            yield formatted_outputs(reply, model_name)
            input_ids = output

# Choosing the default model
if args.model is not None:
    model_name = args.model
else:
    if len(available_models) == 0:
        print("No models are available! Please download at least one.")
        exit(0)
    elif len(available_models) == 1:
        i = 0
    else:
        print("The following models are available:\n")
        for i,model in enumerate(available_models):
            print(f"{i+1}. {model}")
        print(f"\nWhich one do you want to load? 1-{len(available_models)}\n")
        i = int(input())-1
        print()
    model_name = available_models[i]
model, tokenizer = load_model(model_name)

# UI settings
if model_name.lower().startswith('gpt4chan'):
    default_text = settings['prompt_gpt4chan']
else:
    default_text = settings['prompt']

description = f"\n\n# Text generation lab\nGenerate text using Large Language Models.\n"
css = ".my-4 {margin-top: 0} .py-6 {padding-top: 2.5rem}"
if args.chat or args.cai_chat:
    history = []

    # This gets the new line characters right.
    def clean_chat_message(text):
        text = text.replace('\n', '\n\n')
        text = re.sub(r"\n{3,}", "\n\n", text)
        text = text.strip()
        return text

    def generate_chat_prompt(text, tokens, name1, name2, context):
        text = clean_chat_message(text)

        rows = [f"{context}\n\n"]
        i = len(history)-1
        while i >= 0 and len(encode(''.join(rows), tokens)[0]) < 2048-tokens:
            rows.insert(1, f"{name2}: {history[i][1].strip()}\n")
            rows.insert(1, f"{name1}: {history[i][0].strip()}\n")
            i -= 1
        rows.append(f"{name1}: {text}\n")
        rows.append(f"{name2}:")

        while len(rows) > 3 and len(encode(''.join(rows), tokens)[0]) >= 2048-tokens:
            rows.pop(1)
            rows.pop(1)

        question = ''.join(rows)
        return question

    def chatbot_wrapper(text, tokens, inference_settings, selected_model, name1, name2, context, check):
        question = generate_chat_prompt(text, tokens, name1, name2, context)
        history.append(['', ''])
        eos_token = '\n' if check else None
        for i in generate_reply(question, tokens, inference_settings, selected_model, eos_token=eos_token):
            reply = i[0]
            next_character_found = False

            if check:
                idx = reply.rfind(question[-1024:])
                reply = reply[idx+min(1024, len(question)):].split('\n')[0].strip()
            else:
                idx = reply.rfind(question[-1024:])
                reply = reply[idx+min(1024, len(question)):]
                idx = reply.find(f"\n{name1}:")
                if idx != -1:
                    reply = reply[:idx]
                    next_character_found = True
                reply = clean_chat_message(reply)

            history[-1] = [text, reply]
            if next_character_found:
                break

            # Prevent the chat log from flashing if something like "\nYo" is generated just
            # before "\nYou:" is completed
            tmp = f"\n{name1}:"
            next_character_substring_found = False
            for j in range(1, len(tmp)):
                if reply[-j:] == tmp[:j]:
                    next_character_substring_found = True

            if not next_character_substring_found:
                yield history

        yield history

    def cai_chatbot_wrapper(text, tokens, inference_settings, selected_model, name1, name2, context, check):
        for history in chatbot_wrapper(text, tokens, inference_settings, selected_model, name1, name2, context, check):
            yield generate_chat_html(history, name1, name2)

    def remove_last_message(name1, name2):
        history.pop()
        if args.cai_chat:
            return generate_chat_html(history, name1, name2)
        else:
            return history

    def clear():
        global history
        history = []

    def clear_html():
        return generate_chat_html([], "", "")

    def redraw_html(name1, name2):
        global history
        return generate_chat_html(history, name1, name2)

    def save_history():
        if not Path('logs').exists():
            Path('logs').mkdir()
        with open(Path('logs/conversation.json'), 'w') as f:
            f.write(json.dumps({'data': history}))
        return Path('logs/conversation.json')

    def load_history(file):
        global history
        history = json.loads(file.decode('utf-8'))['data']

    if 'pygmalion' in model_name.lower():
        context_str = settings['context_pygmalion']
        name1_str = settings['name1_pygmalion']
        name2_str = settings['name2_pygmalion']
    else:
        context_str = settings['context']
        name1_str = settings['name1']
        name2_str = settings['name2']

    with gr.Blocks(css=css+".h-\[40vh\] {height: 66.67vh} .gradio-container {max-width: 800px; margin-left: auto; margin-right: auto}", analytics_enabled=False) as interface:
        if args.cai_chat:
            display1 = gr.HTML(value=generate_chat_html([], "", ""))
        else:
            display1 = gr.Chatbot()
        textbox = gr.Textbox(lines=2, label='Input')
        btn = gr.Button("Generate")
        with gr.Row():
            btn2 = gr.Button("Clear history")
            stop = gr.Button("Stop")
            btn3 = gr.Button("Remove last message")

        length_slider = gr.Slider(minimum=settings['max_new_tokens_min'], maximum=settings['max_new_tokens_max'], step=1, label='max_new_tokens', value=settings['max_new_tokens'])
        with gr.Row():
            with gr.Column():
                model_menu = gr.Dropdown(choices=available_models, value=model_name, label='Model')
            with gr.Column():
                preset_menu = gr.Dropdown(choices=available_presets, value=settings['preset'], label='Settings preset')

        name1 = gr.Textbox(value=name1_str, lines=1, label='Your name')
        name2 = gr.Textbox(value=name2_str, lines=1, label='Bot\'s name')
        context = gr.Textbox(value=context_str, lines=2, label='Context')
        with gr.Row():
            check = gr.Checkbox(value=settings['stop_at_newline'], label='Stop generating at new line character?')
        with gr.Row():
            with gr.Column():
                gr.Markdown("Upload chat history")
                upload = gr.File(type='binary')
            with gr.Column():
                gr.Markdown("Download chat history")
                save_btn = gr.Button(value="Click me")
                download = gr.File()

        if args.cai_chat:
            gen_event = btn.click(cai_chatbot_wrapper, [textbox, length_slider, preset_menu, model_menu, name1, name2, context, check], display1, show_progress=args.no_stream, api_name="textgen")
            gen_event2 = textbox.submit(cai_chatbot_wrapper, [textbox, length_slider, preset_menu, model_menu, name1, name2, context, check], display1, show_progress=args.no_stream)
            btn2.click(clear_html, [], display1, show_progress=False)
        else:
            gen_event = btn.click(chatbot_wrapper, [textbox, length_slider, preset_menu, model_menu, name1, name2, context, check], display1, show_progress=args.no_stream, api_name="textgen")
            gen_event2 = textbox.submit(chatbot_wrapper, [textbox, length_slider, preset_menu, model_menu, name1, name2, context, check], display1, show_progress=args.no_stream)
            btn2.click(lambda x: "", display1, display1, show_progress=False)

        btn2.click(clear)
        btn3.click(remove_last_message, [name1, name2], display1, show_progress=False)
        btn.click(lambda x: "", textbox, textbox, show_progress=False)
        textbox.submit(lambda x: "", textbox, textbox, show_progress=False)
        stop.click(None, None, None, cancels=[gen_event, gen_event2])

        save_btn.click(save_history, inputs=[], outputs=[download])
        upload.upload(load_history, [upload], [])
        upload.upload(redraw_html, [name1, name2], [display1])


elif args.notebook:
    with gr.Blocks(css=css, analytics_enabled=False) as interface:
        gr.Markdown(description)
        with gr.Tab('Raw'):
            textbox = gr.Textbox(value=default_text, lines=23)
        with gr.Tab('Markdown'):
            markdown = gr.Markdown()
        with gr.Tab('HTML'):
            html = gr.HTML()
        btn = gr.Button("Generate")
        stop = gr.Button("Stop")

        length_slider = gr.Slider(minimum=settings['max_new_tokens_min'], maximum=settings['max_new_tokens_max'], step=1, label='max_new_tokens', value=settings['max_new_tokens'])
        with gr.Row():
            with gr.Column():
                model_menu = gr.Dropdown(choices=available_models, value=model_name, label='Model')
            with gr.Column():
                preset_menu = gr.Dropdown(choices=available_presets, value=settings['preset'], label='Settings preset')

        gen_event = btn.click(generate_reply, [textbox, length_slider, preset_menu, model_menu], [textbox, markdown, html], show_progress=args.no_stream, api_name="textgen")
        gen_event2 = textbox.submit(generate_reply, [textbox, length_slider, preset_menu, model_menu], [textbox, markdown, html], show_progress=args.no_stream)
        stop.click(None, None, None, cancels=[gen_event, gen_event2])

else:
    with gr.Blocks(css=css, analytics_enabled=False) as interface:
        gr.Markdown(description)
        with gr.Row():
            with gr.Column():
                textbox = gr.Textbox(value=default_text, lines=15, label='Input')
                length_slider = gr.Slider(minimum=settings['max_new_tokens_min'], maximum=settings['max_new_tokens_max'], step=1, label='max_new_tokens', value=settings['max_new_tokens'])
                preset_menu = gr.Dropdown(choices=available_presets, value=settings['preset'], label='Settings preset')
                model_menu = gr.Dropdown(choices=available_models, value=model_name, label='Model')
                btn = gr.Button("Generate")
                with gr.Row():
                    with gr.Column():
                        cont = gr.Button("Continue")
                    with gr.Column():
                        stop = gr.Button("Stop")
            with gr.Column():
                with gr.Tab('Raw'):
                    output_textbox = gr.Textbox(lines=15, label='Output')
                with gr.Tab('Markdown'):
                    markdown = gr.Markdown()
                with gr.Tab('HTML'):
                    html = gr.HTML()

        gen_event = btn.click(generate_reply, [textbox, length_slider, preset_menu, model_menu], [output_textbox, markdown, html], show_progress=args.no_stream, api_name="textgen")
        gen_event2 = textbox.submit(generate_reply, [textbox, length_slider, preset_menu, model_menu], [output_textbox, markdown, html], show_progress=args.no_stream)
        cont_event = cont.click(generate_reply, [output_textbox, length_slider, preset_menu, model_menu], [output_textbox, markdown, html], show_progress=args.no_stream)
        stop.click(None, None, None, cancels=[gen_event, gen_event2, cont_event])

interface.queue()
if args.no_listen:
    interface.launch(share=False)
else:
    interface.launch(share=False, server_name="0.0.0.0")