chat.py 19 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471
  1. import base64
  2. import copy
  3. import io
  4. import json
  5. import re
  6. from datetime import datetime
  7. from pathlib import Path
  8. import yaml
  9. from PIL import Image
  10. import modules.extensions as extensions_module
  11. import modules.shared as shared
  12. from modules.extensions import apply_extensions
  13. from modules.html_generator import (fix_newlines, chat_html_wrapper,
  14. make_thumbnail)
  15. from modules.text_generation import (encode, generate_reply,
  16. get_max_prompt_length)
  17. def generate_chat_prompt(user_input, max_new_tokens, name1, name2, context, chat_prompt_size, **kwargs):
  18. is_instruct = kwargs['is_instruct'] if 'is_instruct' in kwargs else False
  19. end_of_turn = kwargs['end_of_turn'] if 'end_of_turn' in kwargs else ''
  20. impersonate = kwargs['impersonate'] if 'impersonate' in kwargs else False
  21. also_return_rows = kwargs['also_return_rows'] if 'also_return_rows' in kwargs else False
  22. rows = [f"{context.strip()}\n"]
  23. # Finding the maximum prompt size
  24. if shared.soft_prompt:
  25. chat_prompt_size -= shared.soft_prompt_tensor.shape[1]
  26. max_length = min(get_max_prompt_length(max_new_tokens), chat_prompt_size)
  27. if is_instruct:
  28. prefix1 = f"{name1}\n"
  29. prefix2 = f"{name2}\n"
  30. else:
  31. prefix1 = f"{name1}: "
  32. prefix2 = f"{name2}: "
  33. i = len(shared.history['internal']) - 1
  34. while i >= 0 and len(encode(''.join(rows), max_new_tokens)[0]) < max_length:
  35. rows.insert(1, f"{prefix2}{shared.history['internal'][i][1].strip()}{end_of_turn}\n")
  36. string = shared.history['internal'][i][0]
  37. if string not in ['', '<|BEGIN-VISIBLE-CHAT|>']:
  38. rows.insert(1, f"{prefix1}{string.strip()}{end_of_turn}\n")
  39. i -= 1
  40. if impersonate:
  41. rows.append(f"{prefix1.strip() if not is_instruct else prefix1}")
  42. limit = 2
  43. else:
  44. # Adding the user message
  45. user_input = fix_newlines(user_input)
  46. if len(user_input) > 0:
  47. rows.append(f"{prefix1}{user_input}{end_of_turn}\n")
  48. # Adding the Character prefix
  49. rows.append(apply_extensions(f"{prefix2.strip() if not is_instruct else prefix2}", "bot_prefix"))
  50. limit = 3
  51. while len(rows) > limit and len(encode(''.join(rows), max_new_tokens)[0]) >= max_length:
  52. rows.pop(1)
  53. prompt = ''.join(rows)
  54. if also_return_rows:
  55. return prompt, rows
  56. else:
  57. return prompt
  58. def extract_message_from_reply(reply, name1, name2, stop_at_newline):
  59. next_character_found = False
  60. if stop_at_newline:
  61. lines = reply.split('\n')
  62. reply = lines[0].strip()
  63. if len(lines) > 1:
  64. next_character_found = True
  65. else:
  66. for string in [f"\n{name1}:", f"\n{name2}:"]:
  67. idx = reply.find(string)
  68. if idx != -1:
  69. reply = reply[:idx]
  70. next_character_found = True
  71. # If something like "\nYo" is generated just before "\nYou:"
  72. # is completed, trim it
  73. if not next_character_found:
  74. for string in [f"\n{name1}:", f"\n{name2}:"]:
  75. for j in range(len(string) - 1, 0, -1):
  76. if reply[-j:] == string[:j]:
  77. reply = reply[:-j]
  78. break
  79. else:
  80. continue
  81. break
  82. reply = fix_newlines(reply)
  83. return reply, next_character_found
  84. def chatbot_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn, regenerate=False):
  85. if mode == 'instruct':
  86. stopping_strings = [f"\n{name1}", f"\n{name2}"]
  87. else:
  88. stopping_strings = [f"\n{name1}:", f"\n{name2}:"]
  89. # Defining some variables
  90. cumulative_reply = ''
  91. just_started = True
  92. name1_original = name1
  93. visible_text = custom_generate_chat_prompt = None
  94. eos_token = '\n' if generate_state['stop_at_newline'] else None
  95. if 'pygmalion' in shared.model_name.lower():
  96. name1 = "You"
  97. # Check if any extension wants to hijack this function call
  98. for extension, _ in extensions_module.iterator():
  99. if hasattr(extension, 'input_hijack') and extension.input_hijack['state']:
  100. extension.input_hijack['state'] = False
  101. text, visible_text = extension.input_hijack['value']
  102. if custom_generate_chat_prompt is None and hasattr(extension, 'custom_generate_chat_prompt'):
  103. custom_generate_chat_prompt = extension.custom_generate_chat_prompt
  104. if visible_text is None:
  105. visible_text = text
  106. text = apply_extensions(text, "input")
  107. # Generating the prompt
  108. kwargs = {'end_of_turn': end_of_turn, 'is_instruct': mode == 'instruct'}
  109. if custom_generate_chat_prompt is None:
  110. prompt = generate_chat_prompt(text, generate_state['max_new_tokens'], name1, name2, context, generate_state['chat_prompt_size'], **kwargs)
  111. else:
  112. prompt = custom_generate_chat_prompt(text, generate_state['max_new_tokens'], name1, name2, context, generate_state['chat_prompt_size'], **kwargs)
  113. # Yield *Is typing...*
  114. if not regenerate:
  115. yield shared.history['visible'] + [[visible_text, shared.processing_message]]
  116. # Generate
  117. for i in range(generate_state['chat_generation_attempts']):
  118. reply = None
  119. 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):
  120. reply = cumulative_reply + reply
  121. # Extracting the reply
  122. reply, next_character_found = extract_message_from_reply(reply, name1, name2, generate_state['stop_at_newline'])
  123. visible_reply = re.sub("(<USER>|<user>|{{user}})", name1_original, reply)
  124. visible_reply = apply_extensions(visible_reply, "output")
  125. # We need this global variable to handle the Stop event,
  126. # otherwise gradio gets confused
  127. if shared.stop_everything:
  128. return shared.history['visible']
  129. if just_started:
  130. just_started = False
  131. shared.history['internal'].append(['', ''])
  132. shared.history['visible'].append(['', ''])
  133. shared.history['internal'][-1] = [text, reply]
  134. shared.history['visible'][-1] = [visible_text, visible_reply]
  135. if not shared.args.no_stream:
  136. yield shared.history['visible']
  137. if next_character_found:
  138. break
  139. if reply is not None:
  140. cumulative_reply = reply
  141. yield shared.history['visible']
  142. def impersonate_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn):
  143. if mode == 'instruct':
  144. stopping_strings = [f"\n{name1}", f"\n{name2}"]
  145. else:
  146. stopping_strings = [f"\n{name1}:", f"\n{name2}:"]
  147. # Defining some variables
  148. cumulative_reply = ''
  149. eos_token = '\n' if generate_state['stop_at_newline'] else None
  150. if 'pygmalion' in shared.model_name.lower():
  151. name1 = "You"
  152. 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)
  153. # Yield *Is typing...*
  154. yield shared.processing_message
  155. for i in range(generate_state['chat_generation_attempts']):
  156. reply = None
  157. 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):
  158. reply = cumulative_reply + reply
  159. reply, next_character_found = extract_message_from_reply(reply, name1, name2, generate_state['stop_at_newline'])
  160. yield reply
  161. if next_character_found:
  162. break
  163. if reply is not None:
  164. cumulative_reply = reply
  165. yield reply
  166. def cai_chatbot_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn):
  167. for history in chatbot_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn):
  168. yield chat_html_wrapper(history, name1, name2, mode)
  169. def regenerate_wrapper(text, generate_state, name1, name2, context, mode, end_of_turn):
  170. if (len(shared.history['visible']) == 1 and not shared.history['visible'][0][0]) or len(shared.history['internal']) == 0:
  171. yield chat_html_wrapper(shared.history['visible'], name1, name2, mode)
  172. else:
  173. last_visible = shared.history['visible'].pop()
  174. last_internal = shared.history['internal'].pop()
  175. # Yield '*Is typing...*'
  176. yield chat_html_wrapper(shared.history['visible'] + [[last_visible[0], shared.processing_message]], name1, name2, mode)
  177. for history in chatbot_wrapper(last_internal[0], generate_state, name1, name2, context, mode, end_of_turn, regenerate=True):
  178. shared.history['visible'][-1] = [last_visible[0], history[-1][1]]
  179. yield chat_html_wrapper(shared.history['visible'], name1, name2, mode)
  180. def remove_last_message(name1, name2, mode):
  181. if len(shared.history['visible']) > 0 and shared.history['internal'][-1][0] != '<|BEGIN-VISIBLE-CHAT|>':
  182. last = shared.history['visible'].pop()
  183. shared.history['internal'].pop()
  184. else:
  185. last = ['', '']
  186. return chat_html_wrapper(shared.history['visible'], name1, name2, mode), last[0]
  187. def send_last_reply_to_input():
  188. if len(shared.history['internal']) > 0:
  189. return shared.history['internal'][-1][1]
  190. else:
  191. return ''
  192. def replace_last_reply(text, name1, name2, mode):
  193. if len(shared.history['visible']) > 0:
  194. shared.history['visible'][-1][1] = text
  195. shared.history['internal'][-1][1] = apply_extensions(text, "input")
  196. return chat_html_wrapper(shared.history['visible'], name1, name2, mode)
  197. def clear_html():
  198. return chat_html_wrapper([], "", "")
  199. def clear_chat_log(name1, name2, greeting, mode):
  200. shared.history['visible'] = []
  201. shared.history['internal'] = []
  202. if greeting != '':
  203. shared.history['internal'] += [['<|BEGIN-VISIBLE-CHAT|>', greeting]]
  204. shared.history['visible'] += [['', apply_extensions(greeting, "output")]]
  205. return chat_html_wrapper(shared.history['visible'], name1, name2, mode)
  206. def redraw_html(name1, name2, mode):
  207. return chat_html_wrapper(shared.history['visible'], name1, name2, mode)
  208. def tokenize_dialogue(dialogue, name1, name2, mode):
  209. history = []
  210. messages = []
  211. dialogue = re.sub('<START>', '', dialogue)
  212. dialogue = re.sub('<start>', '', dialogue)
  213. dialogue = re.sub('(\n|^)[Aa]non:', '\\1You:', dialogue)
  214. dialogue = re.sub('(\n|^)\[CHARACTER\]:', f'\\g<1>{name2}:', dialogue)
  215. idx = [m.start() for m in re.finditer(f"(^|\n)({re.escape(name1)}|{re.escape(name2)}):", dialogue)]
  216. if len(idx) == 0:
  217. return history
  218. for i in range(len(idx) - 1):
  219. messages.append(dialogue[idx[i]:idx[i + 1]].strip())
  220. messages.append(dialogue[idx[-1]:].strip())
  221. entry = ['', '']
  222. for i in messages:
  223. if i.startswith(f'{name1}:'):
  224. entry[0] = i[len(f'{name1}:'):].strip()
  225. elif i.startswith(f'{name2}:'):
  226. entry[1] = i[len(f'{name2}:'):].strip()
  227. if not (len(entry[0]) == 0 and len(entry[1]) == 0):
  228. history.append(entry)
  229. entry = ['', '']
  230. print("\033[1;32;1m\nDialogue tokenized to:\033[0;37;0m\n", end='')
  231. for row in history:
  232. for column in row:
  233. print("\n")
  234. for line in column.strip().split('\n'):
  235. print("| " + line + "\n")
  236. print("|\n")
  237. print("------------------------------")
  238. return history
  239. def save_history(timestamp=True):
  240. if timestamp:
  241. fname = f"{shared.character}_{datetime.now().strftime('%Y%m%d-%H%M%S')}.json"
  242. else:
  243. fname = f"{shared.character}_persistent.json"
  244. if not Path('logs').exists():
  245. Path('logs').mkdir()
  246. with open(Path(f'logs/{fname}'), 'w', encoding='utf-8') as f:
  247. f.write(json.dumps({'data': shared.history['internal'], 'data_visible': shared.history['visible']}, indent=2))
  248. return Path(f'logs/{fname}')
  249. def load_history(file, name1, name2):
  250. file = file.decode('utf-8')
  251. try:
  252. j = json.loads(file)
  253. if 'data' in j:
  254. shared.history['internal'] = j['data']
  255. if 'data_visible' in j:
  256. shared.history['visible'] = j['data_visible']
  257. else:
  258. shared.history['visible'] = copy.deepcopy(shared.history['internal'])
  259. # Compatibility with Pygmalion AI's official web UI
  260. elif 'chat' in j:
  261. shared.history['internal'] = [':'.join(x.split(':')[1:]).strip() for x in j['chat']]
  262. if len(j['chat']) > 0 and j['chat'][0].startswith(f'{name2}:'):
  263. shared.history['internal'] = [['<|BEGIN-VISIBLE-CHAT|>', shared.history['internal'][0]]] + [[shared.history['internal'][i], shared.history['internal'][i + 1]] for i in range(1, len(shared.history['internal']) - 1, 2)]
  264. shared.history['visible'] = copy.deepcopy(shared.history['internal'])
  265. shared.history['visible'][0][0] = ''
  266. else:
  267. shared.history['internal'] = [[shared.history['internal'][i], shared.history['internal'][i + 1]] for i in range(0, len(shared.history['internal']) - 1, 2)]
  268. shared.history['visible'] = copy.deepcopy(shared.history['internal'])
  269. except:
  270. shared.history['internal'] = tokenize_dialogue(file, name1, name2)
  271. shared.history['visible'] = copy.deepcopy(shared.history['internal'])
  272. def replace_character_names(text, name1, name2):
  273. text = text.replace('{{user}}', name1).replace('{{char}}', name2)
  274. return text.replace('<USER>', name1).replace('<BOT>', name2)
  275. def build_pygmalion_style_context(data):
  276. context = ""
  277. if 'char_persona' in data and data['char_persona'] != '':
  278. context += f"{data['char_name']}'s Persona: {data['char_persona']}\n"
  279. if 'world_scenario' in data and data['world_scenario'] != '':
  280. context += f"Scenario: {data['world_scenario']}\n"
  281. context = f"{context.strip()}\n<START>\n"
  282. return context
  283. def generate_pfp_cache(character):
  284. cache_folder = Path("cache")
  285. if not cache_folder.exists():
  286. cache_folder.mkdir()
  287. for path in [Path(f"characters/{character}.{extension}") for extension in ['png', 'jpg', 'jpeg']]:
  288. if path.exists():
  289. img = make_thumbnail(Image.open(path))
  290. img.save(Path('cache/pfp_character.png'), format='PNG')
  291. return img
  292. return None
  293. def load_character(character, name1, name2, mode):
  294. shared.character = character
  295. shared.history['internal'] = []
  296. shared.history['visible'] = []
  297. context = greeting = end_of_turn = ""
  298. greeting_field = 'greeting'
  299. picture = None
  300. # Deleting the profile picture cache, if any
  301. if Path("cache/pfp_character.png").exists():
  302. Path("cache/pfp_character.png").unlink()
  303. if character != 'None':
  304. folder = 'characters' if not mode == 'instruct' else 'characters/instruction-following'
  305. picture = generate_pfp_cache(character)
  306. for extension in ["yml", "yaml", "json"]:
  307. filepath = Path(f'{folder}/{character}.{extension}')
  308. if filepath.exists():
  309. break
  310. file_contents = open(filepath, 'r', encoding='utf-8').read()
  311. data = json.loads(file_contents) if extension == "json" else yaml.safe_load(file_contents)
  312. if 'your_name' in data and data['your_name'] != '':
  313. name1 = data['your_name']
  314. name2 = data['name'] if 'name' in data else data['char_name']
  315. for field in ['context', 'greeting', 'example_dialogue', 'char_persona', 'char_greeting', 'world_scenario']:
  316. if field in data:
  317. data[field] = replace_character_names(data[field], name1, name2)
  318. if 'context' in data:
  319. context = f"{data['context'].strip()}\n\n"
  320. elif "char_persona" in data:
  321. context = build_pygmalion_style_context(data)
  322. greeting_field = 'char_greeting'
  323. if 'example_dialogue' in data:
  324. context += f"{data['example_dialogue'].strip()}\n"
  325. if greeting_field in data:
  326. greeting = data[greeting_field]
  327. if 'end_of_turn' in data:
  328. end_of_turn = data['end_of_turn']
  329. else:
  330. context = shared.settings['context']
  331. name2 = shared.settings['name2']
  332. greeting = shared.settings['greeting']
  333. end_of_turn = shared.settings['end_of_turn']
  334. if Path(f'logs/{shared.character}_persistent.json').exists():
  335. load_history(open(Path(f'logs/{shared.character}_persistent.json'), 'rb').read(), name1, name2)
  336. elif greeting != "":
  337. shared.history['internal'] += [['<|BEGIN-VISIBLE-CHAT|>', greeting]]
  338. shared.history['visible'] += [['', apply_extensions(greeting, "output")]]
  339. return name1, name2, picture, greeting, context, end_of_turn, chat_html_wrapper(shared.history['visible'], name1, name2, mode, reset_cache=True)
  340. def load_default_history(name1, name2):
  341. load_character("None", name1, name2, "chat")
  342. def upload_character(json_file, img, tavern=False):
  343. json_file = json_file if type(json_file) == str else json_file.decode('utf-8')
  344. data = json.loads(json_file)
  345. outfile_name = data["char_name"]
  346. i = 1
  347. while Path(f'characters/{outfile_name}.json').exists():
  348. outfile_name = f'{data["char_name"]}_{i:03d}'
  349. i += 1
  350. if tavern:
  351. outfile_name = f'TavernAI-{outfile_name}'
  352. with open(Path(f'characters/{outfile_name}.json'), 'w', encoding='utf-8') as f:
  353. f.write(json_file)
  354. if img is not None:
  355. img = Image.open(io.BytesIO(img))
  356. img.save(Path(f'characters/{outfile_name}.png'))
  357. print(f'New character saved to "characters/{outfile_name}.json".')
  358. return outfile_name
  359. def upload_tavern_character(img, name1, name2):
  360. _img = Image.open(io.BytesIO(img))
  361. _img.getexif()
  362. decoded_string = base64.b64decode(_img.info['chara'])
  363. _json = json.loads(decoded_string)
  364. _json = {"char_name": _json['name'], "char_persona": _json['description'], "char_greeting": _json["first_mes"], "example_dialogue": _json['mes_example'], "world_scenario": _json['scenario']}
  365. return upload_character(json.dumps(_json), img, tavern=True)
  366. def upload_your_profile_picture(img, name1, name2, mode):
  367. cache_folder = Path("cache")
  368. if not cache_folder.exists():
  369. cache_folder.mkdir()
  370. if img is None:
  371. if Path("cache/pfp_me.png").exists():
  372. Path("cache/pfp_me.png").unlink()
  373. else:
  374. img = make_thumbnail(img)
  375. img.save(Path('cache/pfp_me.png'))
  376. print('Profile picture saved to "cache/pfp_me.png"')
  377. return chat_html_wrapper(shared.history['visible'], name1, name2, mode, reset_cache=True)