text-generation-webui/server.py

241 lines
11 KiB
Python
Raw Normal View History

2022-12-21 17:27:31 +01:00
import re
2023-01-06 05:33:21 +01:00
import time
import glob
2023-01-06 23:56:44 +01:00
from sys import exit
2022-12-21 17:27:31 +01:00
import torch
2023-01-06 23:56:44 +01:00
import argparse
from pathlib import Path
2022-12-21 17:27:31 +01:00
import gradio as gr
import transformers
2023-01-07 03:14:08 +01:00
from html_generator import *
from transformers import AutoTokenizer, T5Tokenizer
from transformers import AutoModelForCausalLM, T5ForConditionalGeneration
2022-12-21 17:27:31 +01:00
2023-01-07 03:14:08 +01:00
2023-01-06 23:56:44 +01:00
parser = argparse.ArgumentParser()
2023-01-07 00:22:26 +01:00
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 webui in notebook mode, where the output is written to the same text box as the input.')
2023-01-08 02:52:46 +01:00
parser.add_argument('--chat', action='store_true', help='Launch the webui in chat mode.')
2023-01-09 14:58:46 +01:00
parser.add_argument('--cpu', action='store_true', help='Use the CPU to generate text.')
2023-01-06 23:56:44 +01:00
args = parser.parse_args()
2023-01-06 06:06:59 +01:00
loaded_preset = None
available_models = sorted(set(map(lambda x : str(x.name).replace('.pt', ''), list(Path('models/').glob('*'))+list(Path('torch-dumps/').glob('*')))))
2023-01-07 03:14:08 +01:00
available_models = [item for item in available_models if not item.endswith('.txt')]
available_presets = sorted(set(map(lambda x : str(x.name).split('.')[0], list(Path('presets').glob('*.txt')))))
2023-01-06 05:33:21 +01:00
2022-12-21 17:27:31 +01:00
def load_model(model_name):
2023-01-06 05:41:52 +01:00
print(f"Loading {model_name}...")
2022-12-21 17:27:31 +01:00
t0 = time.time()
2023-01-06 05:41:52 +01:00
2023-01-09 14:58:46 +01:00
if args.cpu:
dtype = torch.float32
else:
dtype = torch.float16
2023-01-06 06:54:33 +01:00
# Loading the model
2023-01-09 14:58:46 +01:00
if not args.cpu and Path(f"torch-dumps/{model_name}.pt").exists():
2023-01-06 05:41:52 +01:00
print("Loading in .pt format...")
2023-01-09 14:58:46 +01:00
model = torch.load(Path(f"torch-dumps/{model_name}.pt"))
2023-01-06 06:54:33 +01:00
elif model_name.lower().startswith(('gpt-neo', 'opt-', 'galactica')):
2023-01-07 16:21:04 +01:00
if 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)
2023-01-06 06:54:33 +01:00
else:
2023-01-09 14:58:46 +01:00
model = AutoModelForCausalLM.from_pretrained(Path(f"models/{model_name}"), low_cpu_mem_usage=True, torch_dtype=dtype)
2023-01-06 06:06:59 +01:00
elif model_name in ['flan-t5', 't5-large']:
2023-01-09 14:58:46 +01:00
model = T5ForConditionalGeneration.from_pretrained(Path(f"models/{model_name}"))
2023-01-06 06:54:33 +01:00
else:
2023-01-09 14:58:46 +01:00
model = AutoModelForCausalLM.from_pretrained(Path(f"models/{model_name}"), low_cpu_mem_usage=True, torch_dtype=dtype)
2022-12-21 17:27:31 +01:00
2023-01-06 06:54:33 +01:00
# Loading the tokenizer
if model_name.lower().startswith('gpt4chan') and Path(f"models/gpt-j-6B/").exists():
tokenizer = AutoTokenizer.from_pretrained(Path("models/gpt-j-6B/"))
2023-01-09 14:58:46 +01:00
elif model_name in ['flan-t5', 't5-large']:
tokenizer = T5Tokenizer.from_pretrained(Path(f"models/{model_name}/"))
2022-12-21 17:27:31 +01:00
else:
tokenizer = AutoTokenizer.from_pretrained(Path(f"models/{model_name}/"))
2022-12-21 17:27:31 +01:00
2023-01-09 14:58:46 +01:00
# Sending to the GPU
if not (args.cpu or any(size in model_name.lower() for size in ('13b', '20b', '30b'))):
model = model.cuda()
2023-01-06 06:06:59 +01:00
print(f"Loaded the model in {(time.time()-t0):.2f} seconds.")
2022-12-21 17:27:31 +01:00
return model, tokenizer
2023-01-06 06:26:33 +01:00
# Removes empty replies from gpt4chan outputs
2022-12-21 17:27:31 +01:00
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
2023-01-07 05:56:21 +01:00
def fix_galactica(s):
s = s.replace(r'\[', r'$')
s = s.replace(r'\]', r'$')
2023-01-07 16:13:09 +01:00
s = s.replace(r'\(', r'$')
s = s.replace(r'\)', r'$')
s = s.replace(r'$$', r'$')
2023-01-07 05:56:21 +01:00
return s
def generate_reply(question, temperature, max_length, inference_settings, selected_model, eos_token=None):
2023-01-06 06:06:59 +01:00
global model, tokenizer, model_name, loaded_preset, preset
2022-12-21 17:27:31 +01:00
if selected_model != model_name:
model_name = selected_model
model = None
2023-01-08 18:37:43 +01:00
tokenizer = None
2023-01-09 14:58:46 +01:00
if not args.cpu:
torch.cuda.empty_cache()
2022-12-21 17:27:31 +01:00
model, tokenizer = load_model(model_name)
2023-01-06 06:06:59 +01:00
if inference_settings != loaded_preset:
with open(Path(f'presets/{inference_settings}.txt'), 'r') as infile:
2023-01-06 05:33:21 +01:00
preset = infile.read()
2023-01-06 06:06:59 +01:00
loaded_preset = inference_settings
2022-12-21 17:27:31 +01:00
2023-01-09 14:58:46 +01:00
if not args.cpu:
torch.cuda.empty_cache()
input_ids = tokenizer.encode(str(question), return_tensors='pt').cuda()
cuda = ".cuda()"
else:
input_ids = tokenizer.encode(str(question), return_tensors='pt')
cuda = ""
2022-12-21 17:27:31 +01:00
if eos_token is None:
2023-01-09 14:58:46 +01:00
output = eval(f"model.generate(input_ids, {preset}){cuda}")
else:
n = tokenizer.encode(eos_token, return_tensors='pt')[0][1]
2023-01-09 14:58:46 +01:00
output = eval(f"model.generate(input_ids, eos_token_id={n}, {preset}){cuda}")
2023-01-07 03:14:08 +01:00
2023-01-09 14:58:46 +01:00
reply = tokenizer.decode(output[0], skip_special_tokens=True)
2023-01-07 00:22:26 +01:00
if model_name.lower().startswith('galactica'):
2023-01-07 05:56:21 +01:00
reply = fix_galactica(reply)
2023-01-07 03:14:08 +01:00
return reply, reply, 'Only applicable for gpt4chan.'
elif model_name.lower().startswith('gpt4chan'):
2023-01-07 05:56:21 +01:00
reply = fix_gpt4chan(reply)
2023-01-07 03:14:08 +01:00
return reply, 'Only applicable for galactica models.', generate_html(reply)
2023-01-07 00:22:26 +01:00
else:
2023-01-07 03:14:08 +01:00
return reply, 'Only applicable for galactica models.', 'Only applicable for gpt4chan.'
2022-12-21 17:27:31 +01:00
2023-01-06 23:56:44 +01:00
# Choosing the default model
if args.model is not None:
model_name = args.model
else:
2023-01-07 02:05:37 +01:00
if len(available_models) == 0:
2023-01-06 23:56:44 +01:00
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
model_name = available_models[i]
2022-12-21 17:27:31 +01:00
model, tokenizer = load_model(model_name)
2023-01-06 23:56:44 +01:00
# UI settings
2023-01-07 23:11:21 +01:00
if model_name.lower().startswith('gpt4chan'):
2022-12-21 17:27:31 +01:00
default_text = "-----\n--- 865467536\nInput text\n--- 865467537\n"
else:
default_text = "Common sense questions and answers\n\nQuestion: \nFactual answer:"
description = f"""
2023-01-07 23:11:21 +01:00
2023-01-07 00:22:26 +01:00
# Text generation lab
Generate text using Large Language Models.
"""
css=".my-4 {margin-top: 0} .py-6 {padding-top: 2.5rem}"
2023-01-07 00:22:26 +01:00
if args.notebook:
with gr.Blocks(css=css, analytics_enabled=False) as interface:
gr.Markdown(description)
2023-01-07 02:05:37 +01:00
with gr.Tab('Raw'):
textbox = gr.Textbox(value=default_text, lines=23)
with gr.Tab('Markdown'):
markdown = gr.Markdown()
2023-01-07 03:14:08 +01:00
with gr.Tab('HTML'):
html = gr.HTML()
2023-01-07 00:22:26 +01:00
btn = gr.Button("Generate")
2023-01-07 02:05:37 +01:00
with gr.Row():
with gr.Column():
length_slider = gr.Slider(minimum=1, maximum=2000, step=1, label='max_length', value=200)
temp_slider = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Temperature', value=0.7)
2023-01-07 02:05:37 +01:00
with gr.Column():
preset_menu = gr.Dropdown(choices=available_presets, value="NovelAI-Sphinx Moth", label='Preset')
2023-01-07 02:05:37 +01:00
model_menu = gr.Dropdown(choices=available_models, value=model_name, label='Model')
2023-01-07 00:22:26 +01:00
btn.click(generate_reply, [textbox, temp_slider, length_slider, preset_menu, model_menu], [textbox, markdown, html], show_progress=True)
2023-01-09 15:22:12 +01:00
textbox.submit(generate_reply, [textbox, temp_slider, length_slider, preset_menu, model_menu], [textbox, markdown, html], show_progress=True)
2023-01-08 02:52:46 +01:00
elif args.chat:
history = []
def chatbot(text, temperature, max_length, inference_settings, selected_model, name1, name2, context):
question = context+'\n\n'
for i in range(len(history)):
question += f"{name1}: {history[i][0][3:-5].strip()}\n"
question += f"{name2}: {history[i][1][3:-5].strip()}\n"
question += f"{name1}: {text.strip()}\n"
question += f"{name2}:"
reply = generate_reply(question, temperature, max_length, inference_settings, selected_model, eos_token='\n')[0]
2023-01-08 02:52:46 +01:00
reply = reply[len(question):].split('\n')[0].strip()
history.append((text, reply))
return history
def clear():
global history
history = []
with gr.Blocks(css=css+".h-\[40vh\] {height: 50vh}", analytics_enabled=False) as interface:
gr.Markdown(description)
2023-01-08 02:52:46 +01:00
with gr.Row(equal_height=True):
with gr.Column():
with gr.Row(equal_height=True):
with gr.Column():
length_slider = gr.Slider(minimum=1, maximum=2000, step=1, label='max_length', value=200)
2023-01-08 02:52:46 +01:00
preset_menu = gr.Dropdown(choices=available_presets, value="NovelAI-Sphinx Moth", label='Preset')
with gr.Column():
temp_slider = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Temperature', value=0.7)
model_menu = gr.Dropdown(choices=available_models, value=model_name, label='Model')
name1 = gr.Textbox(value='Person 1', lines=1, label='Your name')
name2 = gr.Textbox(value='Person 2', lines=1, label='Bot\'s name')
context = gr.Textbox(value='This is a conversation between two people.', lines=2, label='Context')
with gr.Column():
display1 = gr.Chatbot()
2023-01-08 02:55:54 +01:00
textbox = gr.Textbox(lines=2, label='Input')
2023-01-08 02:52:46 +01:00
btn = gr.Button("Generate")
btn2 = gr.Button("Clear history")
btn.click(chatbot, [textbox, temp_slider, length_slider, preset_menu, model_menu, name1, name2, context], display1, show_progress=True)
2023-01-08 05:10:02 +01:00
textbox.submit(chatbot, [textbox, temp_slider, length_slider, preset_menu, model_menu, name1, name2, context], display1, show_progress=True)
2023-01-08 02:52:46 +01:00
btn2.click(clear)
2023-01-08 05:10:02 +01:00
btn.click(lambda x: "", textbox, textbox, show_progress=False)
2023-01-08 05:33:45 +01:00
textbox.submit(lambda x: "", textbox, textbox, show_progress=False)
2023-01-08 05:10:02 +01:00
btn2.click(lambda x: "", display1, display1)
2023-01-07 00:22:26 +01:00
else:
with gr.Blocks(css=css, analytics_enabled=False) as interface:
gr.Markdown(description)
2023-01-07 02:05:37 +01:00
with gr.Row():
with gr.Column():
textbox = gr.Textbox(value=default_text, lines=15, label='Input')
temp_slider = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Temperature', value=0.7)
length_slider = gr.Slider(minimum=1, maximum=2000, step=1, label='max_length', value=200)
preset_menu = gr.Dropdown(choices=available_presets, value="NovelAI-Sphinx Moth", label='Preset')
2023-01-07 02:05:37 +01:00
model_menu = gr.Dropdown(choices=available_models, value=model_name, label='Model')
btn = gr.Button("Generate")
with gr.Column():
with gr.Tab('Raw'):
output_textbox = gr.Textbox(value=default_text, lines=15, label='Output')
with gr.Tab('Markdown'):
markdown = gr.Markdown()
2023-01-07 03:14:08 +01:00
with gr.Tab('HTML'):
html = gr.HTML()
2023-01-07 02:05:37 +01:00
2023-01-07 03:14:08 +01:00
btn.click(generate_reply, [textbox, temp_slider, length_slider, preset_menu, model_menu], [output_textbox, markdown, html], show_progress=True)
2023-01-09 15:22:12 +01:00
textbox.submit(generate_reply, [textbox, temp_slider, length_slider, preset_menu, model_menu], [output_textbox, markdown, html], show_progress=True)
2022-12-21 17:27:31 +01:00
interface.launch(share=False, server_name="0.0.0.0")