"""), dummy1, dummy1, dummy1
else:
ip = get_local_ip()
document1 = url_path.replace('file://', f'http://{ip}:{port}/')
# document1 = url
return gr.update(visible=True, value=f""""""), dummy1, dummy1, dummy1
else:
return dummy_ret
view_document_choice.select(fn=show_doc, inputs=view_document_choice,
outputs=[doc_view, doc_view2, doc_view3, doc_view4])
# Get inputs to evaluate() and make_db()
# don't deepcopy, can contain model itself
all_kwargs = kwargs.copy()
all_kwargs.update(locals())
refresh_sources1 = functools.partial(update_and_get_source_files_given_langchain_mode,
**get_kwargs(update_and_get_source_files_given_langchain_mode,
exclude_names=['db1', 'langchain_mode'],
**all_kwargs))
eventdb9 = refresh_sources_btn.click(fn=refresh_sources1, inputs=[my_db_state, langchain_mode],
outputs=sources_text,
api_name='refresh_sources' if allow_api else None)
def check_admin_pass(x):
return gr.update(visible=x == admin_pass)
def close_admin(x):
return gr.update(visible=not (x == admin_pass))
admin_btn.click(check_admin_pass, inputs=admin_pass_textbox, outputs=system_row, queue=False) \
.then(close_admin, inputs=admin_pass_textbox, outputs=admin_row, queue=False)
inputs_list, inputs_dict = get_inputs_list(all_kwargs, kwargs['model_lower'], model_id=1)
inputs_list2, inputs_dict2 = get_inputs_list(all_kwargs, kwargs['model_lower'], model_id=2)
from functools import partial
kwargs_evaluate = {k: v for k, v in all_kwargs.items() if k in inputs_kwargs_list}
# ensure present
for k in inputs_kwargs_list:
assert k in kwargs_evaluate, "Missing %s" % k
def evaluate_nochat(*args1, default_kwargs1=None, str_api=False, **kwargs1):
args_list = list(args1)
if str_api:
user_kwargs = args_list[2]
assert isinstance(user_kwargs, str)
user_kwargs = ast.literal_eval(user_kwargs)
else:
user_kwargs = {k: v for k, v in zip(eval_func_param_names, args_list[2:])}
# only used for submit_nochat_api
user_kwargs['chat'] = False
if 'stream_output' not in user_kwargs:
user_kwargs['stream_output'] = False
if 'langchain_mode' not in user_kwargs:
# if user doesn't specify, then assume disabled, not use default
user_kwargs['langchain_mode'] = 'Disabled'
if 'langchain_action' not in user_kwargs:
user_kwargs['langchain_action'] = LangChainAction.QUERY.value
set1 = set(list(default_kwargs1.keys()))
set2 = set(eval_func_param_names)
assert set1 == set2, "Set diff: %s %s: %s" % (set1, set2, set1.symmetric_difference(set2))
# correct ordering. Note some things may not be in default_kwargs, so can't be default of user_kwargs.get()
model_state1 = args_list[0]
my_db_state1 = args_list[1]
args_list = [user_kwargs[k] if k in user_kwargs and user_kwargs[k] is not None else default_kwargs1[k] for k
in eval_func_param_names]
assert len(args_list) == len(eval_func_param_names)
args_list = [model_state1, my_db_state1] + args_list
try:
for res_dict in evaluate(*tuple(args_list), **kwargs1):
if str_api:
# full return of dict
yield res_dict
elif kwargs['langchain_mode'] == 'Disabled':
yield fix_text_for_gradio(res_dict['response'])
else:
yield '
' + fix_text_for_gradio(res_dict['response'])
finally:
clear_torch_cache()
clear_embeddings(user_kwargs['langchain_mode'], my_db_state1)
fun = partial(evaluate_nochat,
default_kwargs1=default_kwargs,
str_api=False,
**kwargs_evaluate)
fun2 = partial(evaluate_nochat,
default_kwargs1=default_kwargs,
str_api=False,
**kwargs_evaluate)
fun_with_dict_str = partial(evaluate_nochat,
default_kwargs1=default_kwargs,
str_api=True,
**kwargs_evaluate
)
dark_mode_btn.click(
None,
None,
None,
_js=get_dark_js(),
api_name="dark" if allow_api else None,
queue=False,
)
def visible_toggle(x):
x = 'off' if x == 'on' else 'on'
return x, gr.Column.update(visible=True if x == 'on' else False)
side_bar_btn.click(fn=visible_toggle,
inputs=side_bar_text,
outputs=[side_bar_text, side_bar],
queue=False)
submit_buttons_btn.click(fn=visible_toggle,
inputs=submit_buttons_text,
outputs=[submit_buttons_text, submit_buttons],
queue=False)
# examples after submit or any other buttons for chat or no chat
if kwargs['examples'] is not None and kwargs['show_examples']:
gr.Examples(examples=kwargs['examples'], inputs=inputs_list)
# Score
def score_last_response(*args, nochat=False, num_model_lock=0):
try:
if num_model_lock > 0:
# then lock way
args_list = list(args).copy()
outputs = args_list[-num_model_lock:]
score_texts1 = []
for output in outputs:
# same input, put into form good for _score_last_response()
args_list[-1] = output
score_texts1.append(
_score_last_response(*tuple(args_list), nochat=nochat,
num_model_lock=num_model_lock, prefix=''))
if len(score_texts1) > 1:
return "Response Scores: %s" % ' '.join(score_texts1)
else:
return "Response Scores: %s" % score_texts1[0]
else:
return _score_last_response(*args, nochat=nochat, num_model_lock=num_model_lock)
finally:
clear_torch_cache()
def _score_last_response(*args, nochat=False, num_model_lock=0, prefix='Response Score: '):
""" Similar to user() """
args_list = list(args)
smodel = score_model_state0['model']
stokenizer = score_model_state0['tokenizer']
sdevice = score_model_state0['device']
if memory_restriction_level > 0:
max_length_tokenize = 768 - 256 if memory_restriction_level <= 2 else 512 - 256
elif hasattr(stokenizer, 'model_max_length'):
max_length_tokenize = stokenizer.model_max_length
else:
# limit to 1024, not worth OOMing on reward score
max_length_tokenize = 2048 - 1024
cutoff_len = max_length_tokenize * 4 # restrict deberta related to max for LLM
if not nochat:
history = args_list[-1]
if history is None:
history = []
if smodel is not None and \
stokenizer is not None and \
sdevice is not None and \
history is not None and len(history) > 0 and \
history[-1] is not None and \
len(history[-1]) >= 2:
os.environ['TOKENIZERS_PARALLELISM'] = 'false'
question = history[-1][0]
answer = history[-1][1]
else:
return '%sNA' % prefix
else:
answer = args_list[-1]
instruction_nochat_arg_id = eval_func_param_names.index('instruction_nochat')
question = args_list[instruction_nochat_arg_id]
if question is None:
return '%sBad Question' % prefix
if answer is None:
return '%sBad Answer' % prefix
try:
score = score_qa(smodel, stokenizer, max_length_tokenize, question, answer, cutoff_len)
finally:
clear_torch_cache()
if isinstance(score, str):
return '%sNA' % prefix
return '{}{:.1%}'.format(prefix, score)
def noop_score_last_response(*args, **kwargs):
return "Response Score: Disabled"
if kwargs['score_model']:
score_fun = score_last_response
else:
score_fun = noop_score_last_response
score_args = dict(fn=score_fun,
inputs=inputs_list + [text_output],
outputs=[score_text],
)
score_args2 = dict(fn=partial(score_fun),
inputs=inputs_list2 + [text_output2],
outputs=[score_text2],
)
score_fun_func = functools.partial(score_fun, num_model_lock=len(text_outputs))
all_score_args = dict(fn=score_fun_func,
inputs=inputs_list + text_outputs,
outputs=score_text,
)
score_args_nochat = dict(fn=partial(score_fun, nochat=True),
inputs=inputs_list + [text_output_nochat],
outputs=[score_text_nochat],
)
def update_history(*args, undo=False, retry=False, sanitize_user_prompt=False):
"""
User that fills history for bot
:param args:
:param undo:
:param retry:
:param sanitize_user_prompt:
:return:
"""
args_list = list(args)
user_message = args_list[eval_func_param_names.index('instruction')] # chat only
input1 = args_list[eval_func_param_names.index('iinput')] # chat only
prompt_type1 = args_list[eval_func_param_names.index('prompt_type')]
langchain_mode1 = args_list[eval_func_param_names.index('langchain_mode')]
langchain_action1 = args_list[eval_func_param_names.index('langchain_action')]
document_subset1 = args_list[eval_func_param_names.index('document_subset')]
document_choice1 = args_list[eval_func_param_names.index('document_choice')]
if not prompt_type1:
# shouldn't have to specify if CLI launched model
prompt_type1 = kwargs['prompt_type']
# apply back
args_list[eval_func_param_names.index('prompt_type')] = prompt_type1
if input1 and not user_message.endswith(':'):
user_message1 = user_message + ":" + input1
elif input1:
user_message1 = user_message + input1
else:
user_message1 = user_message
if sanitize_user_prompt:
from better_profanity import profanity
user_message1 = profanity.censor(user_message1)
history = args_list[-1]
if history is None:
# bad history
history = []
history = history.copy()
if undo:
if len(history) > 0:
history.pop()
return history
if retry:
if history:
history[-1][1] = None
return history
if user_message1 in ['', None, '\n']:
if langchain_action1 in LangChainAction.QUERY.value and \
DocumentChoices.All.name != document_subset1 \
or \
langchain_mode1 in [LangChainMode.CHAT_LLM.value, LangChainMode.LLM.value]:
# reject non-retry submit/enter
return history
user_message1 = fix_text_for_gradio(user_message1)
return history + [[user_message1, None]]
def user(*args, undo=False, retry=False, sanitize_user_prompt=False):
return update_history(*args, undo=undo, retry=retry, sanitize_user_prompt=sanitize_user_prompt)
def all_user(*args, undo=False, retry=False, sanitize_user_prompt=False, num_model_lock=0):
args_list = list(args)
history_list = args_list[-num_model_lock:]
assert len(history_list) > 0, "Bad history list: %s" % history_list
for hi, history in enumerate(history_list):
if num_model_lock > 0:
hargs = args_list[:-num_model_lock].copy()
else:
hargs = args_list.copy()
hargs += [history]
history_list[hi] = update_history(*hargs, undo=undo, retry=retry,
sanitize_user_prompt=sanitize_user_prompt)
if len(history_list) > 1:
return tuple(history_list)
else:
return history_list[0]
def get_model_max_length(model_state1):
if model_state1 and not isinstance(model_state1["tokenizer"], str):
tokenizer = model_state1["tokenizer"]
elif model_state0 and not isinstance(model_state0["tokenizer"], str):
tokenizer = model_state0["tokenizer"]
else:
tokenizer = None
if tokenizer is not None:
return tokenizer.model_max_length
else:
return 2000
def prep_bot(*args, retry=False, which_model=0):
"""
:param args:
:param retry:
:param which_model: identifies which model if doing model_lock
API only called for which_model=0, default for inputs_list, but rest should ignore inputs_list
:return: last element is True if should run bot, False if should just yield history
"""
# don't deepcopy, can contain model itself
args_list = list(args).copy()
model_state1 = args_list[-3]
my_db_state1 = args_list[-2]
history = args_list[-1]
prompt_type1 = args_list[eval_func_param_names.index('prompt_type')]
prompt_dict1 = args_list[eval_func_param_names.index('prompt_dict')]
if model_state1['model'] is None or model_state1['model'] == no_model_str:
return history, None, None, None
args_list = args_list[:-3] # only keep rest needed for evaluate()
langchain_mode1 = args_list[eval_func_param_names.index('langchain_mode')]
langchain_action1 = args_list[eval_func_param_names.index('langchain_action')]
document_subset1 = args_list[eval_func_param_names.index('document_subset')]
document_choice1 = args_list[eval_func_param_names.index('document_choice')]
if not history:
print("No history", flush=True)
history = []
return history, None, None, None
instruction1 = history[-1][0]
if retry and history:
# if retry, pop history and move onto bot stuff
instruction1 = history[-1][0]
history[-1][1] = None
elif not instruction1:
if langchain_action1 in LangChainAction.QUERY.value and \
DocumentChoices.All.name != document_choice1 \
or \
langchain_mode1 in [LangChainMode.CHAT_LLM.value, LangChainMode.LLM.value]:
# if not retrying, then reject empty query
return history, None, None, None
elif len(history) > 0 and history[-1][1] not in [None, '']:
# reject submit button if already filled and not retrying
# None when not filling with '' to keep client happy
return history, None, None, None
# shouldn't have to specify in API prompt_type if CLI launched model, so prefer global CLI one if have it
prompt_type1, prompt_dict1 = update_prompt(prompt_type1, prompt_dict1, model_state1,
which_model=which_model)
# apply back to args_list for evaluate()
args_list[eval_func_param_names.index('prompt_type')] = prompt_type1
args_list[eval_func_param_names.index('prompt_dict')] = prompt_dict1
chat1 = args_list[eval_func_param_names.index('chat')]
model_max_length1 = get_model_max_length(model_state1)
context1 = history_to_context(history, langchain_mode1, prompt_type1, prompt_dict1, chat1,
model_max_length1, memory_restriction_level,
kwargs['keep_sources_in_context'])
args_list[0] = instruction1 # override original instruction with history from user
args_list[2] = context1
fun1 = partial(evaluate,
model_state1,
my_db_state1,
*tuple(args_list),
**kwargs_evaluate)
return history, fun1, langchain_mode1, my_db_state1
def get_response(fun1, history):
"""
bot that consumes history for user input
instruction (from input_list) itself is not consumed by bot
:return:
"""
if not fun1:
yield history, ''
return
try:
for output_fun in fun1():
output = output_fun['response']
extra = output_fun['sources'] # FIXME: can show sources in separate text box etc.
# ensure good visually, else markdown ignores multiple \n
bot_message = fix_text_for_gradio(output)
history[-1][1] = bot_message
yield history, ''
except StopIteration:
yield history, ''
except RuntimeError as e:
if "generator raised StopIteration" in str(e):
# assume last entry was bad, undo
history.pop()
yield history, ''
else:
if history and len(history) > 0 and len(history[0]) > 1 and history[-1][1] is None:
history[-1][1] = ''
yield history, str(e)
raise
except Exception as e:
# put error into user input
ex = "Exception: %s" % str(e)
if history and len(history) > 0 and len(history[0]) > 1 and history[-1][1] is None:
history[-1][1] = ''
yield history, ex
raise
finally:
clear_torch_cache()
return
def clear_embeddings(langchain_mode1, my_db):
# clear any use of embedding that sits on GPU, else keeps accumulating GPU usage even if clear torch cache
if db_type == 'chroma' and langchain_mode1 not in ['ChatLLM', 'LLM', 'Disabled', None, '']:
from gpt_langchain import clear_embedding
db = dbs.get('langchain_mode1')
if db is not None and not isinstance(db, str):
clear_embedding(db)
if langchain_mode1 == LangChainMode.MY_DATA.value and my_db is not None:
clear_embedding(my_db[0])
def bot(*args, retry=False):
history, fun1, langchain_mode1, my_db_state1 = prep_bot(*args, retry=retry)
try:
for res in get_response(fun1, history):
yield res
finally:
clear_torch_cache()
clear_embeddings(langchain_mode1, my_db_state1)
def all_bot(*args, retry=False, model_states1=None):
args_list = list(args).copy()
chatbots = args_list[-len(model_states1):]
args_list0 = args_list[:-len(model_states1)] # same for all models
exceptions = []
stream_output1 = args_list[eval_func_param_names.index('stream_output')]
max_time1 = args_list[eval_func_param_names.index('max_time')]
langchain_mode1 = args_list[eval_func_param_names.index('langchain_mode')]
my_db_state1 = None # will be filled below by some bot
try:
gen_list = []
for chatboti, (chatbot1, model_state1) in enumerate(zip(chatbots, model_states1)):
args_list1 = args_list0.copy()
args_list1.insert(-1, model_state1) # insert at -1 so is at -2
# if at start, have None in response still, replace with '' so client etc. acts like normal
# assumes other parts of code treat '' and None as if no response yet from bot
# can't do this later in bot code as racy with threaded generators
if len(chatbot1) > 0 and len(chatbot1[-1]) == 2 and chatbot1[-1][1] is None:
chatbot1[-1][1] = ''
args_list1.append(chatbot1)
# so consistent with prep_bot()
# with model_state1 at -3, my_db_state1 at -2, and history(chatbot) at -1
# langchain_mode1 and my_db_state1 should be same for every bot
history, fun1, langchain_mode1, my_db_state1 = prep_bot(*tuple(args_list1), retry=retry,
which_model=chatboti)
gen1 = get_response(fun1, history)
if stream_output1:
gen1 = TimeoutIterator(gen1, timeout=0.01, sentinel=None, raise_on_exception=False)
# else timeout will truncate output for non-streaming case
gen_list.append(gen1)
bots_old = chatbots.copy()
exceptions_old = [''] * len(bots_old)
tgen0 = time.time()
for res1 in itertools.zip_longest(*gen_list):
if time.time() - tgen0 > max_time1:
print("Took too long: %s" % max_time1, flush=True)
break
bots = [x[0] if x is not None and not isinstance(x, BaseException) else y for x, y in
zip(res1, bots_old)]
bots_old = bots.copy()
def larger_str(x, y):
return x if len(x) > len(y) else y
exceptions = [x[1] if x is not None and not isinstance(x, BaseException) else larger_str(str(x), y)
for x, y in zip(res1, exceptions_old)]
exceptions_old = exceptions.copy()
def choose_exc(x):
# don't expose ports etc. to exceptions window
if is_public:
return "Endpoint unavailable or failed"
else:
return x
exceptions_str = '\n'.join(
['Model %s: %s' % (iix, choose_exc(x)) for iix, x in enumerate(exceptions) if
x not in [None, '', 'None']])
if len(bots) > 1:
yield tuple(bots + [exceptions_str])
else:
yield bots[0], exceptions_str
if exceptions:
exceptions = [x for x in exceptions if x not in ['', None, 'None']]
if exceptions:
print("Generate exceptions: %s" % exceptions, flush=True)
finally:
clear_torch_cache()
clear_embeddings(langchain_mode1, my_db_state1)
# NORMAL MODEL
user_args = dict(fn=functools.partial(user, sanitize_user_prompt=kwargs['sanitize_user_prompt']),
inputs=inputs_list + [text_output],
outputs=text_output,
)
bot_args = dict(fn=bot,
inputs=inputs_list + [model_state, my_db_state] + [text_output],
outputs=[text_output, chat_exception_text],
)
retry_bot_args = dict(fn=functools.partial(bot, retry=True),
inputs=inputs_list + [model_state, my_db_state] + [text_output],
outputs=[text_output, chat_exception_text],
)
retry_user_args = dict(fn=functools.partial(user, retry=True),
inputs=inputs_list + [text_output],
outputs=text_output,
)
undo_user_args = dict(fn=functools.partial(user, undo=True),
inputs=inputs_list + [text_output],
outputs=text_output,
)
# MODEL2
user_args2 = dict(fn=functools.partial(user, sanitize_user_prompt=kwargs['sanitize_user_prompt']),
inputs=inputs_list2 + [text_output2],
outputs=text_output2,
)
bot_args2 = dict(fn=bot,
inputs=inputs_list2 + [model_state2, my_db_state] + [text_output2],
outputs=[text_output2, chat_exception_text],
)
retry_bot_args2 = dict(fn=functools.partial(bot, retry=True),
inputs=inputs_list2 + [model_state2, my_db_state] + [text_output2],
outputs=[text_output2, chat_exception_text],
)
retry_user_args2 = dict(fn=functools.partial(user, retry=True),
inputs=inputs_list2 + [text_output2],
outputs=text_output2,
)
undo_user_args2 = dict(fn=functools.partial(user, undo=True),
inputs=inputs_list2 + [text_output2],
outputs=text_output2,
)
# MODEL N
all_user_args = dict(fn=functools.partial(all_user,
sanitize_user_prompt=kwargs['sanitize_user_prompt'],
num_model_lock=len(text_outputs),
),
inputs=inputs_list + text_outputs,
outputs=text_outputs,
)
all_bot_args = dict(fn=functools.partial(all_bot, model_states1=model_states),
inputs=inputs_list + [my_db_state] + text_outputs,
outputs=text_outputs + [chat_exception_text],
)
all_retry_bot_args = dict(fn=functools.partial(all_bot, model_states1=model_states, retry=True),
inputs=inputs_list + [my_db_state] + text_outputs,
outputs=text_outputs + [chat_exception_text],
)
all_retry_user_args = dict(fn=functools.partial(all_user, retry=True,
sanitize_user_prompt=kwargs['sanitize_user_prompt'],
num_model_lock=len(text_outputs),
),
inputs=inputs_list + text_outputs,
outputs=text_outputs,
)
all_undo_user_args = dict(fn=functools.partial(all_user, undo=True,
sanitize_user_prompt=kwargs['sanitize_user_prompt'],
num_model_lock=len(text_outputs),
),
inputs=inputs_list + text_outputs,
outputs=text_outputs,
)
def clear_instruct():
return gr.Textbox.update(value='')
def deselect_radio_chats():
return gr.update(value=None)
def clear_all():
return gr.Textbox.update(value=''), gr.Textbox.update(value=''), gr.update(value=None), \
gr.Textbox.update(value=''), gr.Textbox.update(value='')
if kwargs['model_states']:
submits1 = submits2 = submits3 = []
submits4 = []
fun_source = [instruction.submit, submit.click, retry_btn.click]
fun_name = ['instruction', 'submit', 'retry']
user_args = [all_user_args, all_user_args, all_retry_user_args]
bot_args = [all_bot_args, all_bot_args, all_retry_bot_args]
for userargs1, botarg1, funn1, funs1 in zip(user_args, bot_args, fun_name, fun_source):
submit_event11 = funs1(fn=dummy_fun,
inputs=instruction, outputs=instruction, queue=queue)
submit_event1a = submit_event11.then(**userargs1, queue=queue,
api_name='%s' % funn1 if allow_api else None)
# if hit enter on new instruction for submitting new query, no longer the saved chat
submit_event1b = submit_event1a.then(clear_all, inputs=None,
outputs=[instruction, iinput, radio_chats, score_text,
score_text2],
queue=queue)
submit_event1c = submit_event1b.then(**botarg1,
api_name='%s_bot' % funn1 if allow_api else None,
queue=queue)
submit_event1d = submit_event1c.then(**all_score_args,
api_name='%s_bot_score' % funn1 if allow_api else None,
queue=queue)
submits1.extend([submit_event1a, submit_event1b, submit_event1c, submit_event1d])
# if undo, no longer the saved chat
submit_event4 = undo.click(fn=dummy_fun,
inputs=instruction, outputs=instruction, queue=queue) \
.then(**all_undo_user_args, api_name='undo' if allow_api else None) \
.then(clear_all, inputs=None, outputs=[instruction, iinput, radio_chats, score_text,
score_text2], queue=queue) \
.then(**all_score_args, api_name='undo_score' if allow_api else None)
submits4 = [submit_event4]
else:
# in case 2nd model, consume instruction first, so can clear quickly
# bot doesn't consume instruction itself, just history from user, so why works
submit_event11 = instruction.submit(fn=dummy_fun,
inputs=instruction, outputs=instruction, queue=queue)
submit_event1a = submit_event11.then(**user_args, queue=queue,
api_name='instruction' if allow_api else None)
# if hit enter on new instruction for submitting new query, no longer the saved chat
submit_event1a2 = submit_event1a.then(deselect_radio_chats, inputs=None, outputs=radio_chats, queue=queue)
submit_event1b = submit_event1a2.then(**user_args2, api_name='instruction2' if allow_api else None)
submit_event1c = submit_event1b.then(clear_instruct, None, instruction) \
.then(clear_instruct, None, iinput)
submit_event1d = submit_event1c.then(**bot_args, api_name='instruction_bot' if allow_api else None,
queue=queue)
submit_event1e = submit_event1d.then(**score_args,
api_name='instruction_bot_score' if allow_api else None,
queue=queue)
submit_event1f = submit_event1e.then(**bot_args2, api_name='instruction_bot2' if allow_api else None,
queue=queue)
submit_event1g = submit_event1f.then(**score_args2,
api_name='instruction_bot_score2' if allow_api else None, queue=queue)
submits1 = [submit_event1a, submit_event1a2, submit_event1b, submit_event1c, submit_event1d,
submit_event1e,
submit_event1f, submit_event1g]
submit_event21 = submit.click(fn=dummy_fun,
inputs=instruction, outputs=instruction, queue=queue)
submit_event2a = submit_event21.then(**user_args, api_name='submit' if allow_api else None)
# if submit new query, no longer the saved chat
submit_event2a2 = submit_event2a.then(deselect_radio_chats, inputs=None, outputs=radio_chats, queue=queue)
submit_event2b = submit_event2a2.then(**user_args2, api_name='submit2' if allow_api else None)
submit_event2c = submit_event2b.then(clear_all, inputs=None,
outputs=[instruction, iinput, radio_chats, score_text, score_text2],
queue=queue)
submit_event2d = submit_event2c.then(**bot_args, api_name='submit_bot' if allow_api else None, queue=queue)
submit_event2e = submit_event2d.then(**score_args,
api_name='submit_bot_score' if allow_api else None,
queue=queue)
submit_event2f = submit_event2e.then(**bot_args2, api_name='submit_bot2' if allow_api else None,
queue=queue)
submit_event2g = submit_event2f.then(**score_args2,
api_name='submit_bot_score2' if allow_api else None,
queue=queue)
submits2 = [submit_event2a, submit_event2a2, submit_event2b, submit_event2c, submit_event2d,
submit_event2e,
submit_event2f, submit_event2g]
submit_event31 = retry_btn.click(fn=dummy_fun,
inputs=instruction, outputs=instruction, queue=queue)
submit_event3a = submit_event31.then(**user_args, api_name='retry' if allow_api else None)
# if retry, no longer the saved chat
submit_event3a2 = submit_event3a.then(deselect_radio_chats, inputs=None, outputs=radio_chats, queue=queue)
submit_event3b = submit_event3a2.then(**user_args2, api_name='retry2' if allow_api else None)
submit_event3c = submit_event3b.then(clear_instruct, None, instruction) \
.then(clear_instruct, None, iinput)
submit_event3d = submit_event3c.then(**retry_bot_args, api_name='retry_bot' if allow_api else None,
queue=queue)
submit_event3e = submit_event3d.then(**score_args,
api_name='retry_bot_score' if allow_api else None,
queue=queue)
submit_event3f = submit_event3e.then(**retry_bot_args2, api_name='retry_bot2' if allow_api else None,
queue=queue)
submit_event3g = submit_event3f.then(**score_args2,
api_name='retry_bot_score2' if allow_api else None,
queue=queue)
submits3 = [submit_event3a, submit_event3a2, submit_event3b, submit_event3c, submit_event3d,
submit_event3e,
submit_event3f, submit_event3g]
# if undo, no longer the saved chat
submit_event4 = undo.click(fn=dummy_fun,
inputs=instruction, outputs=instruction, queue=queue) \
.then(**undo_user_args, api_name='undo' if allow_api else None) \
.then(**undo_user_args2, api_name='undo2' if allow_api else None) \
.then(clear_all, inputs=None, outputs=[instruction, iinput, radio_chats, score_text,
score_text2], queue=queue) \
.then(**score_args, api_name='undo_score' if allow_api else None) \
.then(**score_args2, api_name='undo_score2' if allow_api else None)
submits4 = [submit_event4]
# MANAGE CHATS
def dedup(short_chat, short_chats):
if short_chat not in short_chats:
return short_chat
for i in range(1, 1000):
short_chat_try = short_chat + "_" + str(i)
if short_chat_try not in short_chats:
return short_chat_try
# fallback and hope for best
short_chat = short_chat + "_" + str(random.random())
return short_chat
def get_short_chat(x, short_chats, short_len=20, words=4):
if x and len(x[0]) == 2 and x[0][0] is not None:
short_chat = ' '.join(x[0][0][:short_len].split(' ')[:words]).strip()
short_chat = dedup(short_chat, short_chats)
else:
short_chat = None
return short_chat
def is_chat_same(x, y):
# etc. added in chat, try to remove some of that to help avoid dup entries when hit new conversation
is_same = True
# length of conversation has to be same
if len(x) != len(y):
return False
if len(x) != len(y):
return False
for stepx, stepy in zip(x, y):
if len(stepx) != len(stepy):
# something off with a conversation
return False
for stepxx, stepyy in zip(stepx, stepy):
if len(stepxx) != len(stepyy):
# something off with a conversation
return False
if len(stepxx) != 2:
# something off
return False
if len(stepyy) != 2:
# something off
return False
questionx = stepxx[0].replace('
', '').replace('
', '') if stepxx[0] is not None else None
answerx = stepxx[1].replace('', '').replace('
', '') if stepxx[1] is not None else None
questiony = stepyy[0].replace('', '').replace('
', '') if stepyy[0] is not None else None
answery = stepyy[1].replace('', '').replace('
', '') if stepyy[1] is not None else None
if questionx != questiony or answerx != answery:
return False
return is_same
def save_chat(*args, chat_is_list=False):
args_list = list(args)
if not chat_is_list:
# list of chatbot histories,
# can't pass in list with list of chatbot histories and state due to gradio limits
chat_list = args_list[:-1]
else:
assert len(args_list) == 2
chat_list = args_list[0]
# if old chat file with single chatbot, get into shape
if isinstance(chat_list, list) and len(chat_list) > 0 and isinstance(chat_list[0], list) and len(
chat_list[0]) == 2 and isinstance(chat_list[0][0], str) and isinstance(chat_list[0][1], str):
chat_list = [chat_list]
# remove None histories
chat_list_not_none = [x for x in chat_list if x and len(x) > 0 and len(x[0]) == 2 and x[0][1] is not None]
chat_list_none = [x for x in chat_list if x not in chat_list_not_none]
if len(chat_list_none) > 0 and len(chat_list_not_none) == 0:
raise ValueError("Invalid chat file")
# dict with keys of short chat names, values of list of list of chatbot histories
chat_state1 = args_list[-1]
short_chats = list(chat_state1.keys())
if len(chat_list_not_none) > 0:
# make short_chat key from only first history, based upon question that is same anyways
chat_first = chat_list_not_none[0]
short_chat = get_short_chat(chat_first, short_chats)
if short_chat:
old_chat_lists = list(chat_state1.values())
already_exists = any([is_chat_same(chat_list, x) for x in old_chat_lists])
if not already_exists:
chat_state1[short_chat] = chat_list.copy()
# clear chat_list so saved and then new conversation starts
# FIXME: seems less confusing to clear, since have clear button right next
# chat_list = [[]] * len(chat_list)
if not chat_is_list:
ret_list = chat_list + [chat_state1]
else:
ret_list = [chat_list] + [chat_state1]
return tuple(ret_list)
def switch_chat(chat_key, chat_state1, num_model_lock=0):
chosen_chat = chat_state1[chat_key]
# deal with possible different size of chat list vs. current list
ret_chat = [None] * (2 + num_model_lock)
for chati in range(0, 2 + num_model_lock):
ret_chat[chati % len(ret_chat)] = chosen_chat[chati % len(chosen_chat)]
return tuple(ret_chat)
def clear_texts(*args):
return tuple([gr.Textbox.update(value='')] * len(args))
def clear_scores():
return gr.Textbox.update(value=res_value), \
gr.Textbox.update(value='Response Score: NA'), \
gr.Textbox.update(value='Response Score: NA')
switch_chat_fun = functools.partial(switch_chat, num_model_lock=len(text_outputs))
radio_chats.input(switch_chat_fun,
inputs=[radio_chats, chat_state],
outputs=[text_output, text_output2] + text_outputs) \
.then(clear_scores, outputs=[score_text, score_text2, score_text_nochat])
def remove_chat(chat_key, chat_state1):
if isinstance(chat_key, str):
chat_state1.pop(chat_key, None)
return gr.update(choices=list(chat_state1.keys()), value=None), chat_state1
remove_chat_event = remove_chat_btn.click(remove_chat,
inputs=[radio_chats, chat_state], outputs=[radio_chats, chat_state],
queue=False)
def get_chats1(chat_state1):
base = 'chats'
makedirs(base, exist_ok=True)
filename = os.path.join(base, 'chats_%s.json' % str(uuid.uuid4()))
with open(filename, "wt") as f:
f.write(json.dumps(chat_state1, indent=2))
return filename
export_chat_event = export_chats_btn.click(get_chats1, inputs=chat_state, outputs=chats_file, queue=False,
api_name='export_chats' if allow_api else None)
def add_chats_from_file(file, chat_state1, radio_chats1, chat_exception_text1):
if not file:
return None, chat_state1, gr.update(choices=list(chat_state1.keys()), value=None), chat_exception_text1
if isinstance(file, str):
files = [file]
else:
files = file
if not files:
return None, chat_state1, gr.update(choices=list(chat_state1.keys()), value=None), chat_exception_text1
chat_exception_list = []
for file1 in files:
try:
if hasattr(file1, 'name'):
file1 = file1.name
with open(file1, "rt") as f:
new_chats = json.loads(f.read())
for chat1_k, chat1_v in new_chats.items():
# ignore chat1_k, regenerate and de-dup to avoid loss
_, chat_state1 = save_chat(chat1_v, chat_state1, chat_is_list=True)
except BaseException as e:
t, v, tb = sys.exc_info()
ex = ''.join(traceback.format_exception(t, v, tb))
ex_str = "File %s exception: %s" % (file1, str(e))
print(ex_str, flush=True)
chat_exception_list.append(ex_str)
chat_exception_text1 = '\n'.join(chat_exception_list)
return None, chat_state1, gr.update(choices=list(chat_state1.keys()), value=None), chat_exception_text1
# note for update_user_db_func output is ignored for db
chatup_change_event = chatsup_output.change(add_chats_from_file,
inputs=[chatsup_output, chat_state, radio_chats,
chat_exception_text],
outputs=[chatsup_output, chat_state, radio_chats,
chat_exception_text],
queue=False,
api_name='add_to_chats' if allow_api else None)
clear_chat_event = clear_chat_btn.click(fn=clear_texts,
inputs=[text_output, text_output2] + text_outputs,
outputs=[text_output, text_output2] + text_outputs,
queue=False, api_name='clear' if allow_api else None) \
.then(deselect_radio_chats, inputs=None, outputs=radio_chats, queue=False) \
.then(clear_scores, outputs=[score_text, score_text2, score_text_nochat])
def update_radio_chats(chat_state1):
# reverse so newest at top
choices = list(chat_state1.keys()).copy()
choices.reverse()
return gr.update(choices=choices, value=None)
clear_event = save_chat_btn.click(save_chat,
inputs=[text_output, text_output2] + text_outputs + [chat_state],
outputs=[text_output, text_output2] + text_outputs + [chat_state],
api_name='save_chat' if allow_api else None) \
.then(update_radio_chats, inputs=chat_state, outputs=radio_chats,
api_name='update_chats' if allow_api else None) \
.then(clear_scores, outputs=[score_text, score_text2, score_text_nochat])
# NOTE: clear of instruction/iinput for nochat has to come after score,
# because score for nochat consumes actual textbox, while chat consumes chat history filled by user()
no_chat_args = dict(fn=fun,
inputs=[model_state, my_db_state] + inputs_list,
outputs=text_output_nochat,
queue=queue,
)
submit_event_nochat = submit_nochat.click(**no_chat_args, api_name='submit_nochat' if allow_api else None) \
.then(clear_torch_cache) \
.then(**score_args_nochat, api_name='instruction_bot_score_nochat' if allow_api else None, queue=queue) \
.then(clear_instruct, None, instruction_nochat) \
.then(clear_instruct, None, iinput_nochat) \
.then(clear_torch_cache)
# copy of above with text box submission
submit_event_nochat2 = instruction_nochat.submit(**no_chat_args) \
.then(clear_torch_cache) \
.then(**score_args_nochat, queue=queue) \
.then(clear_instruct, None, instruction_nochat) \
.then(clear_instruct, None, iinput_nochat) \
.then(clear_torch_cache)
submit_event_nochat_api = submit_nochat_api.click(fun_with_dict_str,
inputs=[model_state, my_db_state, inputs_dict_str],
outputs=text_output_nochat_api,
queue=True, # required for generator
api_name='submit_nochat_api' if allow_api else None) \
.then(clear_torch_cache)
def load_model(model_name, lora_weights, server_name, model_state_old, prompt_type_old, load_8bit,
use_gpu_id, gpu_id):
# ensure no API calls reach here
if is_public:
raise RuntimeError("Illegal access for %s" % model_name)
# ensure old model removed from GPU memory
if kwargs['debug']:
print("Pre-switch pre-del GPU memory: %s" % get_torch_allocated(), flush=True)
model0 = model_state0['model']
if isinstance(model_state_old['model'], str) and model0 is not None:
# best can do, move model loaded at first to CPU
model0.cpu()
if model_state_old['model'] is not None and not isinstance(model_state_old['model'], str):
try:
model_state_old['model'].cpu()
except Exception as e:
# sometimes hit NotImplementedError: Cannot copy out of meta tensor; no data!
print("Unable to put model on CPU: %s" % str(e), flush=True)
del model_state_old['model']
model_state_old['model'] = None
if model_state_old['tokenizer'] is not None and not isinstance(model_state_old['tokenizer'], str):
del model_state_old['tokenizer']
model_state_old['tokenizer'] = None
clear_torch_cache()
if kwargs['debug']:
print("Pre-switch post-del GPU memory: %s" % get_torch_allocated(), flush=True)
if model_name is None or model_name == no_model_str:
# no-op if no model, just free memory
# no detranscribe needed for model, never go into evaluate
lora_weights = no_lora_str
server_name = no_server_str
return [None, None, None, model_name, server_name], \
model_name, lora_weights, server_name, prompt_type_old, \
gr.Slider.update(maximum=256), \
gr.Slider.update(maximum=256)
# don't deepcopy, can contain model itself
all_kwargs1 = all_kwargs.copy()
all_kwargs1['base_model'] = model_name.strip()
all_kwargs1['load_8bit'] = load_8bit
all_kwargs1['use_gpu_id'] = use_gpu_id
all_kwargs1['gpu_id'] = int(gpu_id) # detranscribe
model_lower = model_name.strip().lower()
if model_lower in inv_prompt_type_to_model_lower:
prompt_type1 = inv_prompt_type_to_model_lower[model_lower]
else:
prompt_type1 = prompt_type_old
# detranscribe
if lora_weights == no_lora_str:
lora_weights = ''
all_kwargs1['lora_weights'] = lora_weights.strip()
if server_name == no_server_str:
server_name = ''
all_kwargs1['inference_server'] = server_name.strip()
model1, tokenizer1, device1 = get_model(reward_type=False,
**get_kwargs(get_model, exclude_names=['reward_type'],
**all_kwargs1))
clear_torch_cache()
tokenizer_base_model = model_name
prompt_dict1, error0 = get_prompt(prompt_type1, '',
chat=False, context='', reduced=False, making_context=False,
return_dict=True)
model_state_new = dict(model=model1, tokenizer=tokenizer1, device=device1,
base_model=model_name, tokenizer_base_model=tokenizer_base_model,
lora_weights=lora_weights, inference_server=server_name,
prompt_type=prompt_type1, prompt_dict=prompt_dict1,
)
max_max_new_tokens1 = get_max_max_new_tokens(model_state_new, **kwargs)
if kwargs['debug']:
print("Post-switch GPU memory: %s" % get_torch_allocated(), flush=True)
return model_state_new, model_name, lora_weights, server_name, prompt_type1, \
gr.Slider.update(maximum=max_max_new_tokens1), \
gr.Slider.update(maximum=max_max_new_tokens1)
def get_prompt_str(prompt_type1, prompt_dict1, which=0):
if prompt_type1 in ['', None]:
print("Got prompt_type %s: %s" % (which, prompt_type1), flush=True)
return str({})
prompt_dict1, prompt_dict_error = get_prompt(prompt_type1, prompt_dict1, chat=False, context='',
reduced=False, making_context=False, return_dict=True)
if prompt_dict_error:
return str(prompt_dict_error)
else:
# return so user can manipulate if want and use as custom
return str(prompt_dict1)
get_prompt_str_func1 = functools.partial(get_prompt_str, which=1)
get_prompt_str_func2 = functools.partial(get_prompt_str, which=2)
prompt_type.change(fn=get_prompt_str_func1, inputs=[prompt_type, prompt_dict], outputs=prompt_dict, queue=False)
prompt_type2.change(fn=get_prompt_str_func2, inputs=[prompt_type2, prompt_dict2], outputs=prompt_dict2,
queue=False)
def dropdown_prompt_type_list(x):
return gr.Dropdown.update(value=x)
def chatbot_list(x, model_used_in):
return gr.Textbox.update(label=f'h2oGPT [Model: {model_used_in}]')
load_model_args = dict(fn=load_model,
inputs=[model_choice, lora_choice, server_choice, model_state, prompt_type,
model_load8bit_checkbox, model_use_gpu_id_checkbox, model_gpu],
outputs=[model_state, model_used, lora_used, server_used,
# if prompt_type changes, prompt_dict will change via change rule
prompt_type, max_new_tokens, min_new_tokens,
])
prompt_update_args = dict(fn=dropdown_prompt_type_list, inputs=prompt_type, outputs=prompt_type)
chatbot_update_args = dict(fn=chatbot_list, inputs=[text_output, model_used], outputs=text_output)
nochat_update_args = dict(fn=chatbot_list, inputs=[text_output_nochat, model_used], outputs=text_output_nochat)
load_model_event = load_model_button.click(**load_model_args,
api_name='load_model' if allow_api and is_public else None) \
.then(**prompt_update_args) \
.then(**chatbot_update_args) \
.then(**nochat_update_args) \
.then(clear_torch_cache)
load_model_args2 = dict(fn=load_model,
inputs=[model_choice2, lora_choice2, server_choice2, model_state2, prompt_type2,
model_load8bit_checkbox2, model_use_gpu_id_checkbox2, model_gpu2],
outputs=[model_state2, model_used2, lora_used2, server_used2,
# if prompt_type2 changes, prompt_dict2 will change via change rule
prompt_type2, max_new_tokens2, min_new_tokens2
])
prompt_update_args2 = dict(fn=dropdown_prompt_type_list, inputs=prompt_type2, outputs=prompt_type2)
chatbot_update_args2 = dict(fn=chatbot_list, inputs=[text_output2, model_used2], outputs=text_output2)
load_model_event2 = load_model_button2.click(**load_model_args2,
api_name='load_model2' if allow_api and is_public else None) \
.then(**prompt_update_args2) \
.then(**chatbot_update_args2) \
.then(clear_torch_cache)
def dropdown_model_lora_server_list(model_list0, model_x,
lora_list0, lora_x,
server_list0, server_x,
model_used1, lora_used1, server_used1,
model_used2, lora_used2, server_used2,
):
model_new_state = [model_list0[0] + [model_x]]
model_new_options = [*model_new_state[0]]
x1 = model_x if model_used1 == no_model_str else model_used1
x2 = model_x if model_used2 == no_model_str else model_used2
ret1 = [gr.Dropdown.update(value=x1, choices=model_new_options),
gr.Dropdown.update(value=x2, choices=model_new_options),
'', model_new_state]
lora_new_state = [lora_list0[0] + [lora_x]]
lora_new_options = [*lora_new_state[0]]
# don't switch drop-down to added lora if already have model loaded
x1 = lora_x if model_used1 == no_model_str else lora_used1
x2 = lora_x if model_used2 == no_model_str else lora_used2
ret2 = [gr.Dropdown.update(value=x1, choices=lora_new_options),
gr.Dropdown.update(value=x2, choices=lora_new_options),
'', lora_new_state]
server_new_state = [server_list0[0] + [server_x]]
server_new_options = [*server_new_state[0]]
# don't switch drop-down to added server if already have model loaded
x1 = server_x if model_used1 == no_model_str else server_used1
x2 = server_x if model_used2 == no_model_str else server_used2
ret3 = [gr.Dropdown.update(value=x1, choices=server_new_options),
gr.Dropdown.update(value=x2, choices=server_new_options),
'', server_new_state]
return tuple(ret1 + ret2 + ret3)
add_model_lora_server_event = \
add_model_lora_server_button.click(fn=dropdown_model_lora_server_list,
inputs=[model_options_state, new_model] +
[lora_options_state, new_lora] +
[server_options_state, new_server] +
[model_used, lora_used, server_used] +
[model_used2, lora_used2, server_used2],
outputs=[model_choice, model_choice2, new_model, model_options_state] +
[lora_choice, lora_choice2, new_lora, lora_options_state] +
[server_choice, server_choice2, new_server,
server_options_state],
queue=False)
go_event = go_btn.click(lambda: gr.update(visible=False), None, go_btn, api_name="go" if allow_api else None,
queue=False) \
.then(lambda: gr.update(visible=True), None, normal_block, queue=False) \
.then(**load_model_args, queue=False).then(**prompt_update_args, queue=False)
def compare_textbox_fun(x):
return gr.Textbox.update(visible=x)
def compare_column_fun(x):
return gr.Column.update(visible=x)
def compare_prompt_fun(x):
return gr.Dropdown.update(visible=x)
def slider_fun(x):
return gr.Slider.update(visible=x)
compare_checkbox.select(compare_textbox_fun, compare_checkbox, text_output2,
api_name="compare_checkbox" if allow_api else None) \
.then(compare_column_fun, compare_checkbox, col_model2) \
.then(compare_prompt_fun, compare_checkbox, prompt_type2) \
.then(compare_textbox_fun, compare_checkbox, score_text2) \
.then(slider_fun, compare_checkbox, max_new_tokens2) \
.then(slider_fun, compare_checkbox, min_new_tokens2)
# FIXME: add score_res2 in condition, but do better
# callback for logging flagged input/output
callback.setup(inputs_list + [text_output, text_output2] + text_outputs, "flagged_data_points")
flag_btn.click(lambda *args: callback.flag(args), inputs_list + [text_output, text_output2] + text_outputs,
None,
preprocess=False,
api_name='flag' if allow_api else None, queue=False)
flag_btn_nochat.click(lambda *args: callback.flag(args), inputs_list + [text_output_nochat], None,
preprocess=False,
api_name='flag_nochat' if allow_api else None, queue=False)
def get_system_info():
if is_public:
time.sleep(10) # delay to avoid spam since queue=False
return gr.Textbox.update(value=system_info_print())
system_event = system_btn.click(get_system_info, outputs=system_text,
api_name='system_info' if allow_api else None, queue=False)
def get_system_info_dict(system_input1, **kwargs1):
if system_input1 != os.getenv("ADMIN_PASS", ""):
return json.dumps({})
exclude_list = ['admin_pass', 'examples']
sys_dict = {k: v for k, v in kwargs1.items() if
isinstance(v, (str, int, bool, float)) and k not in exclude_list}
try:
sys_dict.update(system_info())
except Exception as e:
# protection
print("Exception: %s" % str(e), flush=True)
return json.dumps(sys_dict)
get_system_info_dict_func = functools.partial(get_system_info_dict, **all_kwargs)
system_dict_event = system_btn2.click(get_system_info_dict_func,
inputs=system_input,
outputs=system_text2,
api_name='system_info_dict' if allow_api else None,
queue=False, # queue to avoid spam
)
def get_hash():
return kwargs['git_hash']
system_event = system_btn3.click(get_hash,
outputs=system_text3,
api_name='system_hash' if allow_api else None,
queue=False,
)
def count_chat_tokens(model_state1, chat1, prompt_type1, prompt_dict1,
memory_restriction_level1=0,
keep_sources_in_context1=False,
):
if model_state1 and not isinstance(model_state1['tokenizer'], str):
tokenizer = model_state1['tokenizer']
elif model_state0 and not isinstance(model_state0['tokenizer'], str):
tokenizer = model_state0['tokenizer']
else:
tokenizer = None
if tokenizer is not None:
langchain_mode1 = 'ChatLLM'
# fake user message to mimic bot()
chat1 = copy.deepcopy(chat1)
chat1 = chat1 + [['user_message1', None]]
model_max_length1 = tokenizer.model_max_length
context1 = history_to_context(chat1, langchain_mode1, prompt_type1, prompt_dict1, chat1,
model_max_length1,
memory_restriction_level1, keep_sources_in_context1)
return str(tokenizer(context1, return_tensors="pt")['input_ids'].shape[1])
else:
return "N/A"
count_chat_tokens_func = functools.partial(count_chat_tokens,
memory_restriction_level1=memory_restriction_level,
keep_sources_in_context1=kwargs['keep_sources_in_context'])
count_tokens_event = count_chat_tokens_btn.click(fn=count_chat_tokens,
inputs=[model_state, text_output, prompt_type, prompt_dict],
outputs=chat_token_count,
api_name='count_tokens' if allow_api else None)
# don't pass text_output, don't want to clear output, just stop it
# cancel only stops outer generation, not inner generation or non-generation
stop_btn.click(lambda: None, None, None,
cancels=submits1 + submits2 + submits3 + submits4 +
[submit_event_nochat, submit_event_nochat2] +
[eventdb1, eventdb2, eventdb3] +
[eventdb7, eventdb8, eventdb9, eventdb12] +
db_events +
[clear_event] +
[submit_event_nochat_api, submit_event_nochat] +
[load_model_event, load_model_event2] +
[count_tokens_event]
,
queue=False, api_name='stop' if allow_api else None).then(clear_torch_cache, queue=False)
demo.load(None, None, None, _js=get_dark_js() if kwargs['h2ocolors'] and False else None) # light best
demo.queue(concurrency_count=kwargs['concurrency_count'], api_open=kwargs['api_open'])
favicon_path = "h2o-logo.svg"
if not os.path.isfile(favicon_path):
print("favicon_path=%s not found" % favicon_path, flush=True)
favicon_path = None
scheduler = BackgroundScheduler()
scheduler.add_job(func=clear_torch_cache, trigger="interval", seconds=20)
if is_public and \
kwargs['base_model'] not in non_hf_types:
# FIXME: disable for gptj, langchain or gpt4all modify print itself
# FIXME: and any multi-threaded/async print will enter model output!
scheduler.add_job(func=ping, trigger="interval", seconds=60)
scheduler.add_job(func=ping_gpu, trigger="interval", seconds=60 * 10)
scheduler.start()
# import control
if kwargs['langchain_mode'] == 'Disabled' and \
os.environ.get("TEST_LANGCHAIN_IMPORT") and \
kwargs['base_model'] not in non_hf_types:
assert 'gpt_langchain' not in sys.modules, "Dev bug, import of langchain when should not have"
assert 'langchain' not in sys.modules, "Dev bug, import of langchain when should not have"
demo.launch(share=kwargs['share'], server_name="0.0.0.0", show_error=True,
favicon_path=favicon_path, prevent_thread_lock=True,
auth=kwargs['auth'])
if kwargs['verbose']:
print("Started GUI", flush=True)
if kwargs['block_gradio_exit']:
demo.block_thread()
input_args_list = ['model_state', 'my_db_state']
def get_inputs_list(inputs_dict, model_lower, model_id=1):
"""
map gradio objects in locals() to inputs for evaluate().
:param inputs_dict:
:param model_lower:
:param model_id: Which model (1 or 2) of 2
:return:
"""
inputs_list_names = list(inspect.signature(evaluate).parameters)
inputs_list = []
inputs_dict_out = {}
for k in inputs_list_names:
if k == 'kwargs':
continue
if k in input_args_list + inputs_kwargs_list:
# these are added at use time for args or partial for kwargs, not taken as input
continue
if 'mbart-' not in model_lower and k in ['src_lang', 'tgt_lang']:
continue
if model_id == 2:
if k == 'prompt_type':
k = 'prompt_type2'
if k == 'prompt_used':
k = 'prompt_used2'
if k == 'max_new_tokens':
k = 'max_new_tokens2'
if k == 'min_new_tokens':
k = 'min_new_tokens2'
inputs_list.append(inputs_dict[k])
inputs_dict_out[k] = inputs_dict[k]
return inputs_list, inputs_dict_out
def get_sources(db1, langchain_mode, dbs=None, docs_state0=None):
set_userid(db1)
if langchain_mode in ['ChatLLM', 'LLM']:
source_files_added = "NA"
source_list = []
elif langchain_mode in ['wiki_full']:
source_files_added = "Not showing wiki_full, takes about 20 seconds and makes 4MB file." \
" Ask jon.mckinney@h2o.ai for file if required."
source_list = []
elif langchain_mode == 'MyData' and len(db1) > 0 and db1[0] is not None:
from gpt_langchain import get_metadatas
metadatas = get_metadatas(db1[0])
source_list = sorted(set([x['source'] for x in metadatas]))
source_files_added = '\n'.join(source_list)
elif langchain_mode in dbs and dbs[langchain_mode] is not None:
from gpt_langchain import get_metadatas
db1 = dbs[langchain_mode]
metadatas = get_metadatas(db1)
source_list = sorted(set([x['source'] for x in metadatas]))
source_files_added = '\n'.join(source_list)
else:
source_list = []
source_files_added = "None"
sources_dir = "sources_dir"
makedirs(sources_dir)
sources_file = os.path.join(sources_dir, 'sources_%s_%s' % (langchain_mode, str(uuid.uuid4())))
with open(sources_file, "wt") as f:
f.write(source_files_added)
source_list = docs_state0 + source_list
return sources_file, source_list
def set_userid(db1):
# can only call this after function called so for specific userr, not in gr.State() that occurs during app init
assert db1 is not None and len(db1) == 2
if db1[1] is None:
# uuid in db is used as user ID
db1[1] = str(uuid.uuid4())
def update_user_db(file, db1, chunk, chunk_size, langchain_mode, dbs=None, **kwargs):
set_userid(db1)
if file is None:
raise RuntimeError("Don't use change, use input")
try:
return _update_user_db(file, db1=db1, chunk=chunk, chunk_size=chunk_size,
langchain_mode=langchain_mode, dbs=dbs,
**kwargs)
except BaseException as e:
print(traceback.format_exc(), flush=True)
# gradio has issues if except, so fail semi-gracefully, else would hang forever in processing textbox
ex_str = "Exception: %s" % str(e)
source_files_added = """\
Sources:
{0}
""".format(ex_str)
doc_exception_text = str(e)
return None, langchain_mode, source_files_added, doc_exception_text
finally:
clear_torch_cache()
def get_lock_file(db1, langchain_mode):
set_userid(db1)
assert len(db1) == 2 and db1[1] is not None and isinstance(db1[1], str)
user_id = db1[1]
base_path = 'locks'
makedirs(base_path)
lock_file = "db_%s_%s.lock" % (langchain_mode.replace(' ', '_'), user_id)
return lock_file
def _update_user_db(file,
db1=None,
chunk=None, chunk_size=None,
dbs=None, db_type=None, langchain_mode='UserData',
user_path=None,
use_openai_embedding=None,
hf_embedding_model=None,
caption_loader=None,
enable_captions=None,
captions_model=None,
enable_ocr=None,
verbose=None,
is_url=None, is_txt=None,
n_jobs=-1):
assert db1 is not None
assert chunk is not None
assert chunk_size is not None
assert use_openai_embedding is not None
assert hf_embedding_model is not None
assert caption_loader is not None
assert enable_captions is not None
assert captions_model is not None
assert enable_ocr is not None
assert verbose is not None
set_userid(db1)
if dbs is None:
dbs = {}
assert isinstance(dbs, dict), "Wrong type for dbs: %s" % str(type(dbs))
# assert db_type in ['faiss', 'chroma'], "db_type %s not supported" % db_type
from gpt_langchain import add_to_db, get_db, path_to_docs
# handle case of list of temp buffer
if isinstance(file, list) and len(file) > 0 and hasattr(file[0], 'name'):
file = [x.name for x in file]
# handle single file of temp buffer
if hasattr(file, 'name'):
file = file.name
if not isinstance(file, (list, tuple, typing.Generator)) and isinstance(file, str):
file = [file]
if langchain_mode == LangChainMode.DISABLED.value:
return None, langchain_mode, get_source_files(), ""
if langchain_mode in [LangChainMode.CHAT_LLM.value, LangChainMode.CHAT_LLM.value]:
# then switch to MyData, so langchain_mode also becomes way to select where upload goes
# but default to mydata if nothing chosen, since safest
langchain_mode = LangChainMode.MY_DATA.value
if langchain_mode == 'UserData' and user_path is not None:
# move temp files from gradio upload to stable location
for fili, fil in enumerate(file):
if isinstance(fil, str):
if fil.startswith('/tmp/gradio/'):
new_fil = os.path.join(user_path, os.path.basename(fil))
if os.path.isfile(new_fil):
remove(new_fil)
try:
shutil.move(fil, new_fil)
except FileExistsError:
pass
file[fili] = new_fil
if verbose:
print("Adding %s" % file, flush=True)
sources = path_to_docs(file if not is_url and not is_txt else None,
verbose=verbose,
n_jobs=n_jobs,
chunk=chunk, chunk_size=chunk_size,
url=file if is_url else None,
text=file if is_txt else None,
enable_captions=enable_captions,
captions_model=captions_model,
enable_ocr=enable_ocr,
caption_loader=caption_loader,
)
exceptions = [x for x in sources if x.metadata.get('exception')]
exceptions_strs = [x.metadata['exception'] for x in exceptions]
sources = [x for x in sources if 'exception' not in x.metadata]
lock_file = get_lock_file(db1, langchain_mode)
with filelock.FileLock(lock_file):
if langchain_mode == 'MyData':
if db1[0] is not None:
# then add
db, num_new_sources, new_sources_metadata = add_to_db(db1[0], sources, db_type=db_type,
use_openai_embedding=use_openai_embedding,
hf_embedding_model=hf_embedding_model)
else:
# in testing expect:
# assert len(db1) == 2 and db1[1] is None, "Bad MyData db: %s" % db1
# for production hit, when user gets clicky:
assert len(db1) == 2, "Bad MyData db: %s" % db1
# then create
# if added has to original state and didn't change, then would be shared db for all users
persist_directory = os.path.join(scratch_base_dir, 'db_dir_%s_%s' % (langchain_mode, db1[1]))
db = get_db(sources, use_openai_embedding=use_openai_embedding,
db_type=db_type,
persist_directory=persist_directory,
langchain_mode=langchain_mode,
hf_embedding_model=hf_embedding_model)
if db is not None:
db1[0] = db
source_files_added = get_source_files(db=db1[0], exceptions=exceptions)
return None, langchain_mode, source_files_added, '\n'.join(exceptions_strs)
else:
from gpt_langchain import get_persist_directory
persist_directory = get_persist_directory(langchain_mode)
if langchain_mode in dbs and dbs[langchain_mode] is not None:
# then add
db, num_new_sources, new_sources_metadata = add_to_db(dbs[langchain_mode], sources, db_type=db_type,
use_openai_embedding=use_openai_embedding,
hf_embedding_model=hf_embedding_model)
else:
# then create
db = get_db(sources, use_openai_embedding=use_openai_embedding,
db_type=db_type,
persist_directory=persist_directory,
langchain_mode=langchain_mode,
hf_embedding_model=hf_embedding_model)
dbs[langchain_mode] = db
# NOTE we do not return db, because function call always same code path
# return dbs[langchain_mode]
# db in this code path is updated in place
source_files_added = get_source_files(db=dbs[langchain_mode], exceptions=exceptions)
return None, langchain_mode, source_files_added, '\n'.join(exceptions_strs)
def get_db(db1, langchain_mode, dbs=None):
lock_file = get_lock_file(db1, langchain_mode)
with filelock.FileLock(lock_file):
if langchain_mode in ['wiki_full']:
# NOTE: avoid showing full wiki. Takes about 30 seconds over about 90k entries, but not useful for now
db = None
elif langchain_mode == 'MyData' and len(db1) > 0 and db1[0] is not None:
db = db1[0]
elif dbs is not None and langchain_mode in dbs and dbs[langchain_mode] is not None:
db = dbs[langchain_mode]
else:
db = None
return db
def get_source_files_given_langchain_mode(db1, langchain_mode='UserData', dbs=None):
db = get_db(db1, langchain_mode, dbs=dbs)
if langchain_mode in ['ChatLLM', 'LLM'] or db is None:
return "Sources: N/A"
return get_source_files(db=db, exceptions=None)
def get_source_files(db=None, exceptions=None, metadatas=None):
if exceptions is None:
exceptions = []
# only should be one source, not confused
# assert db is not None or metadatas is not None
# clicky user
if db is None and metadatas is None:
return "No Sources at all"
if metadatas is None:
source_label = "Sources:"
if db is not None:
from gpt_langchain import get_metadatas
metadatas = get_metadatas(db)
else:
metadatas = []
adding_new = False
else:
source_label = "New Sources:"
adding_new = True
# below automatically de-dups
from gpt_langchain import get_url
small_dict = {get_url(x['source'], from_str=True, short_name=True): get_short_name(x.get('head')) for x in
metadatas}
# if small_dict is empty dict, that's ok
df = pd.DataFrame(small_dict.items(), columns=['source', 'head'])
df.index = df.index + 1
df.index.name = 'index'
source_files_added = tabulate.tabulate(df, headers='keys', tablefmt='unsafehtml')
if exceptions:
exception_metadatas = [x.metadata for x in exceptions]
small_dict = {get_url(x['source'], from_str=True, short_name=True): get_short_name(x.get('exception')) for x in
exception_metadatas}
# if small_dict is empty dict, that's ok
df = pd.DataFrame(small_dict.items(), columns=['source', 'exception'])
df.index = df.index + 1
df.index.name = 'index'
exceptions_html = tabulate.tabulate(df, headers='keys', tablefmt='unsafehtml')
else:
exceptions_html = ''
if metadatas and exceptions:
source_files_added = """\
{0}
{1}
{2}
""".format(source_label, source_files_added, exceptions_html)
elif metadatas:
source_files_added = """\
{0}
{1}
""".format(source_label, source_files_added)
elif exceptions_html:
source_files_added = """\
Exceptions:
{0}
""".format(exceptions_html)
else:
if adding_new:
source_files_added = "No New Sources"
else:
source_files_added = "No Sources"
return source_files_added
def update_and_get_source_files_given_langchain_mode(db1, langchain_mode, dbs=None, first_para=None,
text_limit=None, chunk=None, chunk_size=None,
user_path=None, db_type=None, load_db_if_exists=None,
n_jobs=None, verbose=None):
db = get_db(db1, langchain_mode, dbs=dbs)
from gpt_langchain import make_db
db, num_new_sources, new_sources_metadata = make_db(use_openai_embedding=False,
hf_embedding_model="sentence-transformers/all-MiniLM-L6-v2",
first_para=first_para, text_limit=text_limit,
chunk=chunk,
chunk_size=chunk_size,
langchain_mode=langchain_mode,
user_path=user_path,
db_type=db_type,
load_db_if_exists=load_db_if_exists,
db=db,
n_jobs=n_jobs,
verbose=verbose)
# return only new sources with text saying such
return get_source_files(db=None, exceptions=None, metadatas=new_sources_metadata)