text-generation-webui/modules/chat.py

462 lines
18 KiB
Python
Raw Normal View History

2023-02-23 11:28:30 -05:00
import base64
import copy
import io
import json
import re
from datetime import datetime
from pathlib import Path
import yaml
2023-04-04 22:03:58 -04:00
from PIL import Image
2023-02-23 12:41:42 -05:00
import modules.extensions as extensions_module
2023-02-25 07:23:02 -05:00
import modules.shared as shared
from modules.extensions import apply_extensions
2023-04-07 13:42:03 -04:00
from modules.html_generator import (chat_html_wrapper, fix_newlines,
2023-04-04 22:03:58 -04:00
make_thumbnail)
2023-03-17 10:42:25 -04:00
from modules.text_generation import (encode, generate_reply,
get_max_prompt_length)
2023-02-23 11:03:52 -05:00
2023-02-25 07:23:02 -05:00
def generate_chat_prompt(user_input, max_new_tokens, name1, name2, context, chat_prompt_size, **kwargs):
is_instruct = kwargs['is_instruct'] if 'is_instruct' in kwargs else False
end_of_turn = kwargs['end_of_turn'] if 'end_of_turn' in kwargs else ''
impersonate = kwargs['impersonate'] if 'impersonate' in kwargs else False
also_return_rows = kwargs['also_return_rows'] if 'also_return_rows' in kwargs else False
rows = [f"{context.strip()}\n"]
2023-04-05 10:49:59 -04:00
# Finding the maximum prompt size
if shared.soft_prompt:
chat_prompt_size -= shared.soft_prompt_tensor.shape[1]
max_length = min(get_max_prompt_length(max_new_tokens), chat_prompt_size)
2023-04-05 10:49:59 -04:00
if is_instruct:
prefix1 = f"{name1}\n"
prefix2 = f"{name2}\n"
else:
prefix1 = f"{name1}: "
prefix2 = f"{name2}: "
i = len(shared.history['internal']) - 1
while i >= 0 and len(encode(''.join(rows), max_new_tokens)[0]) < max_length:
2023-04-05 10:49:59 -04:00
rows.insert(1, f"{prefix2}{shared.history['internal'][i][1].strip()}{end_of_turn}\n")
string = shared.history['internal'][i][0]
if string not in ['', '<|BEGIN-VISIBLE-CHAT|>']:
rows.insert(1, f"{prefix1}{string.strip()}{end_of_turn}\n")
i -= 1
2023-04-05 10:49:59 -04:00
if impersonate:
rows.append(f"{prefix1.strip() if not is_instruct else prefix1}")
limit = 2
else:
# Adding the user message
2023-04-06 00:43:10 -04:00
user_input = fix_newlines(user_input)
2023-03-24 16:17:29 -04:00
if len(user_input) > 0:
2023-04-05 10:49:59 -04:00
rows.append(f"{prefix1}{user_input}{end_of_turn}\n")
# Adding the Character prefix
rows.append(apply_extensions(f"{prefix2.strip() if not is_instruct else prefix2}", "bot_prefix"))
limit = 3
while len(rows) > limit and len(encode(''.join(rows), max_new_tokens)[0]) >= max_length:
rows.pop(1)
2023-02-24 06:31:30 -05:00
prompt = ''.join(rows)
if also_return_rows:
return prompt, rows
else:
return prompt
2023-04-03 00:02:11 -04:00
def extract_message_from_reply(reply, name1, name2, stop_at_newline):
next_character_found = False
2023-04-03 00:02:11 -04:00
if stop_at_newline:
2023-03-05 14:08:50 -05:00
lines = reply.split('\n')
reply = lines[0].strip()
if len(lines) > 1:
next_character_found = True
else:
for string in [f"\n{name1}:", f"\n{name2}:"]:
idx = reply.find(string)
if idx != -1:
reply = reply[:idx]
next_character_found = True
# If something like "\nYo" is generated just before "\nYou:"
# is completed, trim it
if not next_character_found:
for string in [f"\n{name1}:", f"\n{name2}:"]:
for j in range(len(string) - 1, 0, -1):
if reply[-j:] == string[:j]:
reply = reply[:-j]
break
2023-04-06 00:43:10 -04:00
else:
continue
break
reply = fix_newlines(reply)
return reply, next_character_found
def chatbot_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn, regenerate=False):
if mode == 'instruct':
stopping_strings = [f"\n{name1}", f"\n{name2}"]
else:
stopping_strings = [f"\n{name1}:", f"\n{name2}:"]
2023-04-07 10:07:03 -04:00
# Defining some variables
cumulative_reply = ''
just_started = True
2023-03-05 14:08:50 -05:00
name1_original = name1
2023-04-07 10:07:03 -04:00
visible_text = custom_generate_chat_prompt = None
eos_token = '\n' if generate_state['stop_at_newline'] else None
if 'pygmalion' in shared.model_name.lower():
name1 = "You"
2023-02-24 23:08:17 -05:00
# Check if any extension wants to hijack this function call
for extension, _ in extensions_module.iterator():
if hasattr(extension, 'input_hijack') and extension.input_hijack['state']:
2023-02-25 00:36:01 -05:00
extension.input_hijack['state'] = False
2023-02-24 23:08:17 -05:00
text, visible_text = extension.input_hijack['value']
if custom_generate_chat_prompt is None and hasattr(extension, 'custom_generate_chat_prompt'):
custom_generate_chat_prompt = extension.custom_generate_chat_prompt
2023-02-24 23:08:17 -05:00
if visible_text is None:
visible_text = text
text = apply_extensions(text, "input")
2023-04-07 10:07:03 -04:00
# Generating the prompt
kwargs = {'end_of_turn': end_of_turn, 'is_instruct': mode == 'instruct'}
if custom_generate_chat_prompt is None:
prompt = generate_chat_prompt(text, generate_state['max_new_tokens'], name1, name2, context, generate_state['chat_prompt_size'], **kwargs)
2023-02-24 23:08:17 -05:00
else:
prompt = custom_generate_chat_prompt(text, generate_state['max_new_tokens'], name1, name2, context, generate_state['chat_prompt_size'], **kwargs)
2023-02-24 06:31:30 -05:00
# Yield *Is typing...*
2023-03-12 01:42:29 -05:00
if not regenerate:
yield shared.history['visible'] + [[visible_text, shared.processing_message]]
2023-02-24 06:31:30 -05:00
# Generate
for i in range(generate_state['chat_generation_attempts']):
reply = None
for reply in generate_reply(f"{prompt}{' ' if len(cumulative_reply) > 0 else ''}{cumulative_reply}", generate_state, eos_token=eos_token, stopping_strings=stopping_strings):
2023-03-24 01:03:30 -04:00
reply = cumulative_reply + reply
2023-02-24 23:31:01 -05:00
# Extracting the reply
reply, next_character_found = extract_message_from_reply(reply, name1, name2, generate_state['stop_at_newline'])
2023-03-05 15:01:47 -05:00
visible_reply = re.sub("(<USER>|<user>|{{user}})", name1_original, reply)
visible_reply = apply_extensions(visible_reply, "output")
2023-02-24 23:31:01 -05:00
# We need this global variable to handle the Stop event,
# otherwise gradio gets confused
if shared.stop_everything:
return shared.history['visible']
if just_started:
just_started = False
shared.history['internal'].append(['', ''])
shared.history['visible'].append(['', ''])
shared.history['internal'][-1] = [text, reply]
shared.history['visible'][-1] = [visible_text, visible_reply]
if not shared.args.no_stream:
2023-02-24 23:31:01 -05:00
yield shared.history['visible']
if next_character_found:
break
2023-03-05 14:08:50 -05:00
if reply is not None:
cumulative_reply = reply
2023-03-24 01:03:30 -04:00
yield shared.history['visible']
def impersonate_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn):
if mode == 'instruct':
stopping_strings = [f"\n{name1}", f"\n{name2}"]
else:
stopping_strings = [f"\n{name1}:", f"\n{name2}:"]
2023-04-07 10:07:03 -04:00
# Defining some variables
cumulative_reply = ''
eos_token = '\n' if generate_state['stop_at_newline'] else None
if 'pygmalion' in shared.model_name.lower():
name1 = "You"
prompt = generate_chat_prompt(text, generate_state['max_new_tokens'], name1, name2, context, generate_state['chat_prompt_size'], impersonate=True, end_of_turn=end_of_turn)
2023-02-24 06:44:54 -05:00
# Yield *Is typing...*
yield shared.processing_message
2023-03-24 01:03:30 -04:00
for i in range(generate_state['chat_generation_attempts']):
reply = None
for reply in generate_reply(f"{prompt}{' ' if len(cumulative_reply) > 0 else ''}{cumulative_reply}", generate_state, eos_token=eos_token, stopping_strings=stopping_strings):
2023-03-24 01:03:30 -04:00
reply = cumulative_reply + reply
reply, next_character_found = extract_message_from_reply(reply, name1, name2, generate_state['stop_at_newline'])
yield reply
2023-02-24 23:42:19 -05:00
if next_character_found:
break
2023-03-24 01:03:30 -04:00
if reply is not None:
cumulative_reply = reply
2023-03-24 01:03:30 -04:00
yield reply
def cai_chatbot_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn):
2023-04-06 00:43:10 -04:00
for history in chatbot_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn):
2023-04-05 10:49:59 -04:00
yield chat_html_wrapper(history, name1, name2, mode)
def regenerate_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn):
if (len(shared.history['visible']) == 1 and not shared.history['visible'][0][0]) or len(shared.history['internal']) == 0:
2023-04-05 10:49:59 -04:00
yield chat_html_wrapper(shared.history['visible'], name1, name2, mode)
else:
2023-02-23 13:11:18 -05:00
last_visible = shared.history['visible'].pop()
last_internal = shared.history['internal'].pop()
# Yield '*Is typing...*'
yield chat_html_wrapper(shared.history['visible'] + [[last_visible[0], shared.processing_message]], name1, name2, mode)
for history in chatbot_wrapper(last_internal[0], generate_state, name1, name2, context, mode, end_of_turn, regenerate=True):
2023-04-05 10:49:59 -04:00
shared.history['visible'][-1] = [last_visible[0], history[-1][1]]
yield chat_html_wrapper(shared.history['visible'], name1, name2, mode)
2023-04-05 10:49:59 -04:00
def remove_last_message(name1, name2, mode):
if len(shared.history['visible']) > 0 and shared.history['internal'][-1][0] != '<|BEGIN-VISIBLE-CHAT|>':
2023-02-23 13:11:18 -05:00
last = shared.history['visible'].pop()
shared.history['internal'].pop()
else:
last = ['', '']
2023-04-05 10:49:59 -04:00
return chat_html_wrapper(shared.history['visible'], name1, name2, mode), last[0]
def send_last_reply_to_input():
2023-02-23 13:11:18 -05:00
if len(shared.history['internal']) > 0:
return shared.history['internal'][-1][1]
else:
return ''
2023-04-05 10:49:59 -04:00
def replace_last_reply(text, name1, name2, mode):
2023-02-23 13:11:18 -05:00
if len(shared.history['visible']) > 0:
2023-04-05 10:49:59 -04:00
shared.history['visible'][-1][1] = text
2023-02-23 13:11:18 -05:00
shared.history['internal'][-1][1] = apply_extensions(text, "input")
2023-04-05 10:49:59 -04:00
return chat_html_wrapper(shared.history['visible'], name1, name2, mode)
def clear_html():
2023-04-05 10:49:59 -04:00
return chat_html_wrapper([], "", "")
2023-02-23 13:11:18 -05:00
2023-04-05 10:49:59 -04:00
def clear_chat_log(name1, name2, greeting, mode):
shared.history['visible'] = []
shared.history['internal'] = []
if greeting != '':
shared.history['internal'] += [['<|BEGIN-VISIBLE-CHAT|>', greeting]]
shared.history['visible'] += [['', apply_extensions(greeting, "output")]]
2023-04-05 10:49:59 -04:00
return chat_html_wrapper(shared.history['visible'], name1, name2, mode)
2023-04-05 10:49:59 -04:00
def redraw_html(name1, name2, mode):
return chat_html_wrapper(shared.history['visible'], name1, name2, mode)
2023-04-05 10:49:59 -04:00
def tokenize_dialogue(dialogue, name1, name2, mode):
2023-04-03 00:15:20 -04:00
history = []
2023-04-07 10:07:03 -04:00
messages = []
dialogue = re.sub('<START>', '', dialogue)
dialogue = re.sub('<start>', '', dialogue)
dialogue = re.sub('(\n|^)[Aa]non:', '\\1You:', dialogue)
dialogue = re.sub('(\n|^)\[CHARACTER\]:', f'\\g<1>{name2}:', dialogue)
idx = [m.start() for m in re.finditer(f"(^|\n)({re.escape(name1)}|{re.escape(name2)}):", dialogue)]
if len(idx) == 0:
2023-04-03 00:15:20 -04:00
return history
for i in range(len(idx) - 1):
messages.append(dialogue[idx[i]:idx[i + 1]].strip())
messages.append(dialogue[idx[-1]:].strip())
entry = ['', '']
for i in messages:
if i.startswith(f'{name1}:'):
entry[0] = i[len(f'{name1}:'):].strip()
elif i.startswith(f'{name2}:'):
entry[1] = i[len(f'{name2}:'):].strip()
if not (len(entry[0]) == 0 and len(entry[1]) == 0):
2023-04-03 00:15:20 -04:00
history.append(entry)
entry = ['', '']
print("\033[1;32;1m\nDialogue tokenized to:\033[0;37;0m\n", end='')
2023-04-03 00:15:20 -04:00
for row in history:
for column in row:
print("\n")
for line in column.strip().split('\n'):
print("| " + line + "\n")
print("|\n")
print("------------------------------")
2023-04-03 00:15:20 -04:00
return history
def save_history(timestamp=True):
if timestamp:
fname = f"{shared.character}_{datetime.now().strftime('%Y%m%d-%H%M%S')}.json"
else:
fname = f"{shared.character}_persistent.json"
if not Path('logs').exists():
Path('logs').mkdir()
2023-03-12 00:40:01 -05:00
with open(Path(f'logs/{fname}'), 'w', encoding='utf-8') as f:
2023-02-23 13:11:18 -05:00
f.write(json.dumps({'data': shared.history['internal'], 'data_visible': shared.history['visible']}, indent=2))
return Path(f'logs/{fname}')
def load_history(file, name1, name2):
file = file.decode('utf-8')
try:
j = json.loads(file)
if 'data' in j:
2023-02-23 13:11:18 -05:00
shared.history['internal'] = j['data']
if 'data_visible' in j:
2023-02-23 13:11:18 -05:00
shared.history['visible'] = j['data_visible']
else:
2023-02-23 13:11:18 -05:00
shared.history['visible'] = copy.deepcopy(shared.history['internal'])
except:
2023-02-23 13:11:18 -05:00
shared.history['internal'] = tokenize_dialogue(file, name1, name2)
shared.history['visible'] = copy.deepcopy(shared.history['internal'])
def replace_character_names(text, name1, name2):
text = text.replace('{{user}}', name1).replace('{{char}}', name2)
return text.replace('<USER>', name1).replace('<BOT>', name2)
def build_pygmalion_style_context(data):
context = ""
if 'char_persona' in data and data['char_persona'] != '':
context += f"{data['char_name']}'s Persona: {data['char_persona']}\n"
if 'world_scenario' in data and data['world_scenario'] != '':
context += f"Scenario: {data['world_scenario']}\n"
context = f"{context.strip()}\n<START>\n"
return context
def generate_pfp_cache(character):
cache_folder = Path("cache")
if not cache_folder.exists():
cache_folder.mkdir()
for path in [Path(f"characters/{character}.{extension}") for extension in ['png', 'jpg', 'jpeg']]:
if path.exists():
2023-04-04 22:03:58 -04:00
img = make_thumbnail(Image.open(path))
img.save(Path('cache/pfp_character.png'), format='PNG')
return img
return None
2023-04-05 13:22:32 -04:00
def load_character(character, name1, name2, mode):
shared.character = character
2023-02-23 13:11:18 -05:00
shared.history['internal'] = []
shared.history['visible'] = []
2023-04-05 10:49:59 -04:00
context = greeting = end_of_turn = ""
greeting_field = 'greeting'
picture = None
# Deleting the profile picture cache, if any
if Path("cache/pfp_character.png").exists():
Path("cache/pfp_character.png").unlink()
if character != 'None':
2023-04-05 13:22:32 -04:00
folder = 'characters' if not mode == 'instruct' else 'characters/instruction-following'
picture = generate_pfp_cache(character)
for extension in ["yml", "yaml", "json"]:
2023-04-05 10:49:59 -04:00
filepath = Path(f'{folder}/{character}.{extension}')
if filepath.exists():
break
file_contents = open(filepath, 'r', encoding='utf-8').read()
data = json.loads(file_contents) if extension == "json" else yaml.safe_load(file_contents)
if 'your_name' in data and data['your_name'] != '':
name1 = data['your_name']
name2 = data['name'] if 'name' in data else data['char_name']
for field in ['context', 'greeting', 'example_dialogue', 'char_persona', 'char_greeting', 'world_scenario']:
if field in data:
data[field] = replace_character_names(data[field], name1, name2)
if 'context' in data:
context = f"{data['context'].strip()}\n\n"
2023-04-05 10:49:59 -04:00
elif "char_persona" in data:
context = build_pygmalion_style_context(data)
greeting_field = 'char_greeting'
2023-04-05 10:49:59 -04:00
if 'example_dialogue' in data:
context += f"{data['example_dialogue'].strip()}\n"
2023-04-05 10:49:59 -04:00
if greeting_field in data:
greeting = data[greeting_field]
2023-04-05 10:49:59 -04:00
if 'end_of_turn' in data:
end_of_turn = data['end_of_turn']
else:
context = shared.settings['context']
name2 = shared.settings['name2']
greeting = shared.settings['greeting']
2023-04-05 10:49:59 -04:00
end_of_turn = shared.settings['end_of_turn']
2023-02-23 13:11:18 -05:00
if Path(f'logs/{shared.character}_persistent.json').exists():
load_history(open(Path(f'logs/{shared.character}_persistent.json'), 'rb').read(), name1, name2)
elif greeting != "":
shared.history['internal'] += [['<|BEGIN-VISIBLE-CHAT|>', greeting]]
shared.history['visible'] += [['', apply_extensions(greeting, "output")]]
2023-04-05 13:22:32 -04:00
return name1, name2, picture, greeting, context, end_of_turn, chat_html_wrapper(shared.history['visible'], name1, name2, mode, reset_cache=True)
def load_default_history(name1, name2):
2023-04-05 13:22:32 -04:00
load_character("None", name1, name2, "chat")
def upload_character(json_file, img, tavern=False):
json_file = json_file if type(json_file) == str else json_file.decode('utf-8')
data = json.loads(json_file)
outfile_name = data["char_name"]
i = 1
while Path(f'characters/{outfile_name}.json').exists():
outfile_name = f'{data["char_name"]}_{i:03d}'
i += 1
if tavern:
outfile_name = f'TavernAI-{outfile_name}'
2023-03-12 00:21:30 -05:00
with open(Path(f'characters/{outfile_name}.json'), 'w', encoding='utf-8') as f:
f.write(json_file)
if img is not None:
img = Image.open(io.BytesIO(img))
img.save(Path(f'characters/{outfile_name}.png'))
print(f'New character saved to "characters/{outfile_name}.json".')
return outfile_name
def upload_tavern_character(img, name1, name2):
_img = Image.open(io.BytesIO(img))
_img.getexif()
decoded_string = base64.b64decode(_img.info['chara'])
_json = json.loads(decoded_string)
_json = {"char_name": _json['name'], "char_persona": _json['description'], "char_greeting": _json["first_mes"], "example_dialogue": _json['mes_example'], "world_scenario": _json['scenario']}
return upload_character(json.dumps(_json), img, tavern=True)
2023-04-05 10:49:59 -04:00
def upload_your_profile_picture(img, name1, name2, mode):
cache_folder = Path("cache")
if not cache_folder.exists():
cache_folder.mkdir()
if img is None:
if Path("cache/pfp_me.png").exists():
Path("cache/pfp_me.png").unlink()
else:
2023-04-04 22:03:58 -04:00
img = make_thumbnail(img)
img.save(Path('cache/pfp_me.png'))
print('Profile picture saved to "cache/pfp_me.png"')
2023-04-04 22:15:57 -04:00
2023-04-05 10:49:59 -04:00
return chat_html_wrapper(shared.history['visible'], name1, name2, mode, reset_cache=True)