Mirrors/text-generation-webui
1
0
Fork 0
mirror of https://github.com/oobabooga/text-generation-webui.git synced 2024-11-21 23:57:58 +01:00
Code Issues Packages Projects Releases Wiki Activity

Fix stopping strings for llama-3 and phi (#6043)

Browse Source
This commit is contained in:
oobabooga 2024-05-22 13:53:59 -03:00 committed by GitHub
parent 8aaa0a6f4e
commit 5499bc9bc8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 37 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

73
modules/chat.py
View File

@ -45,34 +45,35 @@ yaml.add_representer(str, str_presenter)
yaml.representer.SafeRepresenter.add_representer(str, str_presenter)
def get_generation_prompt(renderer, impersonate=False, strip_trailing_spaces=True):
def extract_message_prefix_suffix(renderer, strip_trailing_spaces=True):
'''
Given a Jinja template, reverse-engineers the prefix and the suffix for
an assistant message (if impersonate=False) or an user message
(if impersonate=True)
Given a Jinja template, extracts the prefix and suffix for
an assistant message and a user message. It assumes that they
share the same suffix.
'''
if impersonate:
messages = [
{"role": "user", "content": "<<|user-message-1|>>"},
{"role": "user", "content": "<<|user-message-2|>>"},
]
else:
messages = [
{"role": "assistant", "content": "<<|user-message-1|>>"},
{"role": "assistant", "content": "<<|user-message-2|>>"},
]
messages = [
{"role": "user", "content": "<<|user-message-1|>>"},
{"role": "assistant", "content": "<<|assistant-message-1|>>"},
{"role": "user", "content": "<<|user-message-2|>>"},
{"role": "assistant", "content": "<<|assistant-message-2|>>"},
]
prompt = renderer(messages=messages)
unwanted_suffix = renderer(messages=[])
suffix_plus_prefix = prompt.split("<<|user-message-1|>>")[1].split("<<|user-message-2|>>")[0]
suffix = prompt.split("<<|user-message-2|>>")[1]
prefix = suffix_plus_prefix[len(suffix):]
suffix = prompt.split('<<|assistant-message-2|>>')[1]
if unwanted_suffix != '':
suffix = suffix[:-len(unwanted_suffix)]
prefix_user = prompt.split('<<|assistant-message-1|>>')[1].split('<<|user-message-2|>>')[0][len(suffix):]
prefix_assistant = prompt.split('<<|user-message-1|>>')[1].split('<<|assistant-message-1|>>')[0][len(suffix):]
if strip_trailing_spaces:
prefix = prefix.rstrip(' ')
prefix_user = prefix_user.rstrip(' ')
prefix_assistant = prefix_assistant.rstrip(' ')
return prefix, suffix
return prefix_user, prefix_assistant, suffix
def generate_chat_prompt(user_input, state, **kwargs):
@ -125,7 +126,12 @@ def generate_chat_prompt(user_input, state, **kwargs):
messages.append({"role": "user", "content": user_input})
def remove_extra_bos(prompt):
for bos_token in ['<s>', '<|startoftext|>', '<BOS_TOKEN>', '<|endoftext|>']:
if hasattr(shared.tokenizer, 'bos_token_id'):
bos_tokens = [shared.tokenizer.decode(shared.tokenizer.bos_token_id)]
else:
bos_tokens = ['<s>', '<|startoftext|>', '<BOS_TOKEN>']
for bos_token in bos_tokens:
while prompt.startswith(bos_token):
prompt = prompt[len(bos_token):]
@ -137,6 +143,9 @@ def generate_chat_prompt(user_input, state, **kwargs):
else:
prompt = renderer(messages=messages)
prefix_user, prefix_assistant, suffix = extract_message_prefix_suffix(renderer, strip_trailing_spaces=not _continue)
prefix = prefix_user if impersonate else prefix_assistant
if state['mode'] == 'chat-instruct':
outer_messages = []
if state['custom_system_message'].strip() != '':
@ -148,29 +157,25 @@ def generate_chat_prompt(user_input, state, **kwargs):
command = command.replace('<|prompt|>', prompt)
command = replace_character_names(command, state['name1'], state['name2'])
if _continue:
prefix = get_generation_prompt(renderer, impersonate=impersonate, strip_trailing_spaces=False)[0]
prefix += messages[-1]["content"]
else:
prefix = get_generation_prompt(renderer, impersonate=impersonate)[0]
if not impersonate:
prefix = apply_extensions('bot_prefix', prefix, state)
elif not impersonate:
prefix = apply_extensions('bot_prefix', prefix, state)
outer_messages.append({"role": "user", "content": command})
outer_messages.append({"role": "assistant", "content": prefix})
prompt = instruction_template.render(messages=outer_messages)
suffix = get_generation_prompt(instruct_renderer, impersonate=False)[1]
if len(suffix) > 0:
prompt = prompt[:-len(suffix)]
else:
if _continue:
suffix = get_generation_prompt(renderer, impersonate=impersonate)[1]
if len(suffix) > 0:
prompt = prompt[:-len(suffix)]
else:
prefix = get_generation_prompt(renderer, impersonate=impersonate)[0]
if state['mode'] == 'chat' and not impersonate:
prefix = apply_extensions('bot_prefix', prefix, state)
@ -249,15 +254,11 @@ def get_stopping_strings(state):
renderers.append(renderer)
for renderer in renderers:
prefix_bot, suffix_bot = get_generation_prompt(renderer, impersonate=False)
prefix_user, suffix_user = get_generation_prompt(renderer, impersonate=True)
prefix_user, prefix_assistant, suffix = extract_message_prefix_suffix(renderer)
stopping_strings += [
suffix_user + prefix_bot,
suffix_user + prefix_user,
suffix_bot + prefix_bot,
suffix_bot + prefix_user,
]
for item in [suffix + prefix_assistant, suffix + prefix_user, suffix]:
stopping_strings.append(item)
stopping_strings.append(item.rstrip())
if 'stopping_strings' in state and isinstance(state['stopping_strings'], list):
stopping_strings += state.pop('stopping_strings')