text-generation-webui/modules/text_generation.py

385 lines
15 KiB
Python
Raw Normal View History

import ast
2023-05-04 02:43:17 +02:00
import logging
import random
2023-02-23 17:28:30 +01:00
import re
import time
import traceback
2023-02-23 17:28:30 +01:00
import numpy as np
import torch
import transformers
2023-02-23 18:41:42 +01:00
import modules.shared as shared
from modules.callbacks import (Iteratorize, Stream,
_SentinelTokenStoppingCriteria)
from modules.extensions import apply_extensions
2023-02-23 18:41:42 +01:00
from modules.html_generator import generate_4chan_html, generate_basic_html
2023-04-08 02:36:04 +02:00
from modules.models import clear_torch_cache, local_rank
2023-02-23 18:41:42 +01:00
def get_max_prompt_length(state):
max_length = state['truncation_length'] - state['max_new_tokens']
if shared.soft_prompt:
max_length -= shared.soft_prompt_tensor.shape[1]
return max_length
def encode(prompt, add_special_tokens=True, add_bos_token=True, truncation_length=None):
if shared.model_type in ['rwkv', 'llamacpp']:
2023-03-06 12:45:49 +01:00
input_ids = shared.tokenizer.encode(str(prompt))
input_ids = np.array(input_ids).reshape(1, len(input_ids))
return input_ids
else:
input_ids = shared.tokenizer.encode(str(prompt), return_tensors='pt', add_special_tokens=add_special_tokens)
2023-04-10 21:44:22 +02:00
# This is a hack for making replies more creative.
if not add_bos_token and input_ids[0][0] == shared.tokenizer.bos_token_id:
input_ids = input_ids[:, 1:]
# Llama adds this extra token when the first character is '\n', and this
# compromises the stopping criteria, so we just remove it
if type(shared.tokenizer) is transformers.LlamaTokenizer and input_ids[0][0] == 29871:
input_ids = input_ids[:, 1:]
# Handling truncation
if truncation_length is not None:
input_ids = input_ids[:, -truncation_length:]
if shared.model_type in ['rwkv', 'llamacpp'] or shared.args.cpu:
return input_ids
elif shared.args.flexgen:
return input_ids.numpy()
elif shared.args.deepspeed:
return input_ids.to(device=local_rank)
elif torch.has_mps:
device = torch.device('mps')
return input_ids.to(device)
else:
return input_ids.cuda()
def get_encoded_length(prompt):
length_after_extensions = apply_extensions('tokenized_length', prompt)
if length_after_extensions is not None:
return length_after_extensions
return len(encode(prompt)[0])
def decode(output_ids, skip_special_tokens=True):
return shared.tokenizer.decode(output_ids, skip_special_tokens)
def generate_softprompt_input_tensors(input_ids):
inputs_embeds = shared.model.transformer.wte(input_ids)
inputs_embeds = torch.cat((shared.soft_prompt_tensor, inputs_embeds), dim=1)
filler_input_ids = torch.zeros((1, inputs_embeds.shape[1]), dtype=input_ids.dtype).to(shared.model.device)
# filler_input_ids += shared.model.config.bos_token_id # setting dummy input_ids to bos tokens
return inputs_embeds, filler_input_ids
# Removes empty replies from gpt4chan outputs
def fix_gpt4chan(s):
for i in range(10):
s = re.sub("--- [0-9]*\n>>[0-9]*\n---", "---", s)
s = re.sub("--- [0-9]*\n *\n---", "---", s)
s = re.sub("--- [0-9]*\n\n\n---", "---", s)
return s
# Fix the LaTeX equations in galactica
def fix_galactica(s):
s = s.replace(r'\[', r'$')
s = s.replace(r'\]', r'$')
s = s.replace(r'\(', r'$')
s = s.replace(r'\)', r'$')
s = s.replace(r'$$', r'$')
s = re.sub(r'\n', r'\n\n', s)
s = re.sub(r"\n{3,}", "\n\n", s)
return s
2023-05-11 20:37:04 +02:00
def get_reply_from_output_ids(output_ids, input_ids, original_question, state, is_chat=False):
if shared.model_type == 'HF_seq2seq':
reply = decode(output_ids, state['skip_special_tokens'])
else:
new_tokens = len(output_ids) - len(input_ids[0])
reply = decode(output_ids[-new_tokens:], state['skip_special_tokens'])
2023-05-11 22:55:50 +02:00
# Prevent LlamaTokenizer from skipping a space
if type(shared.tokenizer) is transformers.LlamaTokenizer and len(output_ids) > 0:
if shared.tokenizer.convert_ids_to_tokens(int(output_ids[-new_tokens])).startswith(''):
reply = ' ' + reply
if not is_chat:
reply = apply_extensions('output', reply)
return reply
def formatted_outputs(reply, model_name):
2023-05-11 20:37:04 +02:00
if shared.model_type == 'galactica':
reply = fix_galactica(reply)
return reply, reply, generate_basic_html(reply)
elif shared.model_type == 'gpt4chan':
reply = fix_gpt4chan(reply)
return reply, 'Only applicable for GALACTICA models.', generate_4chan_html(reply)
else:
2023-05-11 20:37:04 +02:00
return reply, 'Only applicable for GALACTICA models.', generate_basic_html(reply)
2023-03-22 19:40:20 +01:00
def set_manual_seed(seed):
2023-04-10 15:53:31 +02:00
seed = int(seed)
if seed == -1:
seed = random.randint(1, 2**31)
2023-04-25 00:24:12 +02:00
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
2023-04-25 00:24:12 +02:00
2023-04-10 15:53:31 +02:00
return seed
2023-03-22 19:40:20 +01:00
2023-03-27 18:23:59 +02:00
def stop_everything_event():
shared.stop_everything = True
2023-05-11 20:37:04 +02:00
def generate_reply_wrapper(question, state, eos_token=None, stopping_strings=None):
for reply in generate_reply(question, state, eos_token, stopping_strings, is_chat=False):
if shared.model_type not in ['HF_seq2seq']:
2023-05-11 22:11:10 +02:00
reply = question + reply
2023-05-11 20:37:04 +02:00
yield formatted_outputs(reply, shared.model_name)
def generate_reply(question, state, eos_token=None, stopping_strings=None, is_chat=False):
state = apply_extensions('state', state)
generate_func = apply_extensions('custom_generate_reply')
if generate_func is None:
if shared.model_name == 'None' or shared.model is None:
logging.error("No model is loaded! Select one in the Model tab.")
2023-05-11 20:37:04 +02:00
yield question
return
if shared.model_type in ['rwkv', 'llamacpp']:
generate_func = generate_reply_custom
elif shared.args.flexgen:
generate_func = generate_reply_flexgen
else:
generate_func = generate_reply_HF
2023-04-25 00:24:12 +02:00
# Preparing the input
original_question = question
2023-05-11 20:37:04 +02:00
if not is_chat:
2023-04-24 01:32:22 +02:00
question = apply_extensions('input', question)
2023-05-04 20:56:06 +02:00
if shared.args.verbose:
print(f'\n\n{question}\n--------------------\n')
shared.stop_everything = False
clear_torch_cache()
seed = set_manual_seed(state['seed'])
2023-05-11 20:37:04 +02:00
for reply in generate_func(question, original_question, seed, state, eos_token, stopping_strings, is_chat=is_chat):
yield reply
2023-04-25 00:24:12 +02:00
2023-05-11 20:37:04 +02:00
def generate_reply_HF(question, original_question, seed, state, eos_token=None, stopping_strings=None, is_chat=False):
generate_params = {}
for k in ['max_new_tokens', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_penalty', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping']:
generate_params[k] = state[k]
if state['ban_eos_token']:
generate_params['suppress_tokens'] = [shared.tokenizer.eos_token_id]
if shared.args.no_cache:
generate_params.update({'use_cache': False})
if shared.args.deepspeed:
generate_params.update({'synced_gpus': True})
2023-02-28 03:03:35 +01:00
2023-04-25 00:24:12 +02:00
# Encode the input
input_ids = encode(question, add_bos_token=state['add_bos_token'], truncation_length=get_max_prompt_length(state))
2023-03-08 15:26:29 +01:00
output = input_ids[0]
cuda = not any((shared.args.cpu, shared.args.deepspeed))
2023-04-25 00:24:12 +02:00
# Find the eos tokens
2023-03-13 14:32:28 +01:00
eos_token_ids = [shared.tokenizer.eos_token_id] if shared.tokenizer.eos_token_id is not None else []
if eos_token is not None:
eos_token_ids.append(int(encode(eos_token)[0][-1]))
2023-04-25 00:24:12 +02:00
# Add the encoded tokens to generate_params
if shared.soft_prompt:
inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids)
2023-04-24 01:32:22 +02:00
question, filler_input_ids, inputs_embeds = apply_extensions('tokenizer', state, question, filler_input_ids, inputs_embeds)
original_input_ids = input_ids
2023-04-07 16:14:32 +02:00
generate_params.update({'inputs_embeds': inputs_embeds})
generate_params.update({'inputs': filler_input_ids})
else:
2023-04-24 01:32:22 +02:00
question, input_ids, inputs_embeds = apply_extensions('tokenizer', state, question, input_ids, None)
original_input_ids = input_ids
2023-04-07 16:14:32 +02:00
generate_params.update({'inputs': input_ids})
2023-04-24 01:32:22 +02:00
if inputs_embeds is not None:
generate_params.update({'inputs_embeds': inputs_embeds})
# Create the StoppingCriteriaList with the stopping strings (needs to be done after tokenizer extensions)
stopping_criteria_list = transformers.StoppingCriteriaList()
for st in (stopping_strings, ast.literal_eval(f"[{state['custom_stopping_strings']}]")):
if type(st) is list and len(st) > 0:
sentinel_token_ids = [encode(string, add_special_tokens=False) for string in st]
stopping_criteria_list.append(_SentinelTokenStoppingCriteria(sentinel_token_ids=sentinel_token_ids, starting_idx=len(input_ids[0])))
break
# Update generate_params with the eos token and the stopping strings
generate_params['eos_token_id'] = eos_token_ids
generate_params['stopping_criteria'] = stopping_criteria_list
t0 = time.time()
2023-03-12 06:31:45 +01:00
try:
2023-05-11 20:37:04 +02:00
if not is_chat and shared.model_type != 'HF_seq2seq':
yield ''
2023-03-12 06:31:45 +01:00
# Generate the entire reply at once.
if not state['stream']:
with torch.no_grad():
output = shared.model.generate(**generate_params)[0]
if cuda:
output = output.cuda()
if shared.soft_prompt:
output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))
2023-05-11 20:37:04 +02:00
yield get_reply_from_output_ids(output, input_ids, original_question, state, is_chat=is_chat)
2023-03-12 06:31:45 +01:00
# Stream the reply 1 token at a time.
# This is based on the trick of using 'stopping_criteria' to create an iterator.
else:
2023-03-12 06:31:45 +01:00
def generate_with_callback(callback=None, **kwargs):
kwargs['stopping_criteria'].append(Stream(callback_func=callback))
clear_torch_cache()
with torch.no_grad():
shared.model.generate(**kwargs)
def generate_with_streaming(**kwargs):
return Iteratorize(generate_with_callback, kwargs, callback=None)
with generate_with_streaming(**generate_params) as generator:
2023-03-12 06:31:45 +01:00
for output in generator:
if shared.soft_prompt:
output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))
2023-05-11 20:37:04 +02:00
yield get_reply_from_output_ids(output, input_ids, original_question, state, is_chat=is_chat)
if output[-1] in eos_token_ids:
2023-03-12 06:31:45 +01:00
break
except Exception:
traceback.print_exc()
finally:
t1 = time.time()
original_tokens = len(original_input_ids[0])
new_tokens = len(output) - (original_tokens if shared.model_type != 'HF_seq2seq' else 0)
print(f'Output generated in {(t1-t0):.2f} seconds ({new_tokens/(t1-t0):.2f} tokens/s, {new_tokens} tokens, context {original_tokens}, seed {seed})')
return
2023-05-11 20:37:04 +02:00
def generate_reply_custom(question, original_question, seed, state, eos_token=None, stopping_strings=None, is_chat=False):
seed = set_manual_seed(state['seed'])
generate_params = {'token_count': state['max_new_tokens']}
for k in ['temperature', 'top_p', 'top_k', 'repetition_penalty']:
generate_params[k] = state[k]
t0 = time.time()
try:
2023-05-11 20:37:04 +02:00
if not is_chat:
yield ''
if not state['stream']:
reply = shared.model.generate(context=question, **generate_params)
2023-05-11 20:37:04 +02:00
if not is_chat:
reply = apply_extensions('output', reply)
yield reply
else:
for reply in shared.model.generate_with_streaming(context=question, **generate_params):
2023-05-11 20:37:04 +02:00
if not is_chat:
reply = apply_extensions('output', reply)
yield reply
except Exception:
traceback.print_exc()
finally:
t1 = time.time()
original_tokens = len(encode(original_question)[0])
new_tokens = len(encode(original_question + reply)[0]) - original_tokens
print(f'Output generated in {(t1-t0):.2f} seconds ({new_tokens/(t1-t0):.2f} tokens/s, {new_tokens} tokens, context {original_tokens}, seed {seed})')
return
2023-05-11 20:37:04 +02:00
def generate_reply_flexgen(question, original_question, seed, state, eos_token=None, stopping_strings=None, is_chat=False):
generate_params = {}
for k in ['max_new_tokens', 'do_sample', 'temperature']:
generate_params[k] = state[k]
if state['stream']:
generate_params['max_new_tokens'] = 8
# Encode the input
input_ids = encode(question, add_bos_token=state['add_bos_token'], truncation_length=get_max_prompt_length(state))
output = input_ids[0]
# Find the eos tokens
eos_token_ids = [shared.tokenizer.eos_token_id] if shared.tokenizer.eos_token_id is not None else []
if eos_token is not None:
eos_token_ids.append(int(encode(eos_token)[0][-1]))
# Add the encoded tokens to generate_params
question, input_ids, inputs_embeds = apply_extensions('tokenizer', state, question, input_ids, None)
original_input_ids = input_ids
generate_params.update({'inputs': input_ids})
if inputs_embeds is not None:
generate_params.update({'inputs_embeds': inputs_embeds})
# Update generate_params with the eos token and the stopping strings
generate_params['stop'] = eos_token_ids[-1]
t0 = time.time()
try:
2023-05-11 20:37:04 +02:00
if not is_chat:
yield ''
# Generate the entire reply at once.
if not state['stream']:
with torch.no_grad():
output = shared.model.generate(**generate_params)[0]
2023-05-11 20:37:04 +02:00
yield get_reply_from_output_ids(output, input_ids, original_question, state, is_chat=is_chat)
2023-03-12 06:31:45 +01:00
# Stream the output naively for FlexGen since it doesn't support 'stopping_criteria'
else:
for i in range(state['max_new_tokens'] // 8 + 1):
if shared.stop_everything:
break
2023-03-12 06:31:45 +01:00
clear_torch_cache()
with torch.no_grad():
output = shared.model.generate(**generate_params)[0]
if np.count_nonzero(np.isin(input_ids[0], eos_token_ids)) < np.count_nonzero(np.isin(output, eos_token_ids)):
2023-02-26 04:36:04 +01:00
break
yield get_reply_from_output_ids(output, original_input_ids, original_question, state)
input_ids = np.reshape(output, (1, output.shape[0]))
generate_params.update({'inputs': input_ids})
2023-02-26 04:36:04 +01:00
except Exception:
traceback.print_exc()
2023-03-12 06:31:45 +01:00
finally:
t1 = time.time()
original_tokens = len(original_input_ids[0])
new_tokens = len(output) - (original_tokens if shared.model_type != 'HF_seq2seq' else 0)
2023-04-10 15:53:31 +02:00
print(f'Output generated in {(t1-t0):.2f} seconds ({new_tokens/(t1-t0):.2f} tokens/s, {new_tokens} tokens, context {original_tokens}, seed {seed})')
2023-03-12 06:31:45 +01:00
return