text-generation-webui/modules/text_generation.py

466 lines
17 KiB
Python
Raw Normal View History

import ast
2023-06-25 06:38:54 +02:00
import copy
import html
import pprint
import random
import time
import traceback
2023-02-23 17:28:30 +01:00
import numpy as np
import torch
import transformers
2024-04-11 23:42:20 +02:00
from transformers import (
LogitsProcessorList,
is_torch_npu_available,
is_torch_xpu_available
)
2023-02-23 18:41:42 +01:00
import modules.shared as shared
from modules import models
from modules.cache_utils import process_llamacpp_cache
2023-06-25 06:38:54 +02:00
from modules.callbacks import (
Iteratorize,
Stream,
_StopEverythingStoppingCriteria
)
from modules.extensions import apply_extensions
from modules.grammar.grammar_utils import initialize_grammar
from modules.grammar.logits_process import GrammarConstrainedLogitsProcessor
2024-04-05 01:10:47 +02:00
from modules.html_generator import generate_basic_html
from modules.logging_colors import logger
from modules.models import clear_torch_cache, load_model
2023-02-23 18:41:42 +01:00
def generate_reply(*args, **kwargs):
if shared.args.idle_timeout > 0 and shared.model is None and shared.previous_model_name not in [None, 'None']:
shared.model, shared.tokenizer = load_model(shared.previous_model_name)
shared.generation_lock.acquire()
try:
for result in _generate_reply(*args, **kwargs):
yield result
finally:
models.last_generation_time = time.time()
shared.generation_lock.release()
def _generate_reply(question, state, stopping_strings=None, is_chat=False, escape_html=False, for_ui=False):
2023-08-07 02:49:27 +02:00
# Find the appropriate generation function
generate_func = apply_extensions('custom_generate_reply')
if generate_func is None:
if shared.model_name == 'None' or shared.model is None:
logger.error("No model is loaded! Select one in the Model tab.")
yield ''
return
2023-08-07 02:49:27 +02:00
2024-04-05 01:23:58 +02:00
if shared.model.__class__.__name__ in ['LlamaCppModel', 'Exllamav2Model']:
2023-08-07 02:49:27 +02:00
generate_func = generate_reply_custom
else:
generate_func = generate_reply_HF
if generate_func != generate_reply_HF and shared.args.verbose:
logger.info("PROMPT=")
2024-05-18 18:57:00 +02:00
print_prompt(question)
2023-08-07 02:49:27 +02:00
# Prepare the input
original_question = question
if not is_chat:
state = apply_extensions('state', state)
question = apply_extensions('input', question, state)
# Find the stopping strings
all_stop_strings = []
for st in (stopping_strings, state['custom_stopping_strings']):
if type(st) is str:
st = ast.literal_eval(f"[{st}]")
2023-08-07 02:49:27 +02:00
if type(st) is list and len(st) > 0:
all_stop_strings += st
shared.stop_everything = False
clear_torch_cache()
seed = set_manual_seed(state['seed'])
last_update = -1
reply = ''
is_stream = state['stream']
if len(all_stop_strings) > 0 and not state['stream']:
state = copy.deepcopy(state)
state['stream'] = True
min_update_interval = 0
if state.get('max_updates_second', 0) > 0:
min_update_interval = 1 / state['max_updates_second']
2023-08-07 02:49:27 +02:00
# Generate
for reply in generate_func(question, original_question, seed, state, stopping_strings, is_chat=is_chat):
reply, stop_found = apply_stopping_strings(reply, all_stop_strings)
if escape_html:
reply = html.escape(reply)
2024-03-26 20:32:20 +01:00
2023-08-07 02:49:27 +02:00
if is_stream:
cur_time = time.time()
2023-08-29 22:44:31 +02:00
2024-03-26 20:32:20 +01:00
# Limit number of tokens/second to make text readable in real time
2023-08-29 22:44:31 +02:00
if state['max_tokens_second'] > 0:
diff = 1 / state['max_tokens_second'] - (cur_time - last_update)
if diff > 0:
time.sleep(diff)
last_update = time.time()
2023-08-07 02:49:27 +02:00
yield reply
# Limit updates to avoid lag in the Gradio UI
# API updates are not limited
2023-08-29 22:44:31 +02:00
else:
if cur_time - last_update > min_update_interval:
last_update = cur_time
yield reply
2024-03-26 20:32:20 +01:00
yield reply
2023-08-29 22:44:31 +02:00
if stop_found or (state['max_tokens_second'] > 0 and shared.stop_everything):
2023-08-07 02:49:27 +02:00
break
if not is_chat:
reply = apply_extensions('output', reply, state)
yield reply
def encode(prompt, add_special_tokens=True, add_bos_token=True, truncation_length=None):
2023-09-17 16:01:34 +02:00
if shared.tokenizer is None:
raise ValueError('No tokenizer is loaded')
2024-04-05 01:23:58 +02:00
if shared.model.__class__.__name__ in ['LlamaCppModel', 'Exllamav2Model']:
2023-03-06 12:45:49 +01:00
input_ids = shared.tokenizer.encode(str(prompt))
if shared.model.__class__.__name__ not in ['Exllamav2Model']:
input_ids = np.array(input_ids).reshape(1, len(input_ids))
else:
input_ids = shared.tokenizer.encode(str(prompt), return_tensors='pt', add_special_tokens=add_special_tokens)
if hasattr(shared.tokenizer, 'bos_token_id') and shared.tokenizer.bos_token_id is not None:
if add_bos_token:
if (len(input_ids[0]) > 0 and input_ids[0][0] != shared.tokenizer.bos_token_id) or len(input_ids[0]) == 0:
# Add a missing bos token (it may not have been added due to faulty model metadata)
bos_tensor = torch.tensor([[shared.tokenizer.bos_token_id]])
input_ids = torch.cat((bos_tensor, input_ids), 1)
# Prevent double bos token due to jinja templates with <s> somewhere
while len(input_ids[0]) > 1 and input_ids[0][0] == shared.tokenizer.bos_token_id and input_ids[0][1] == shared.tokenizer.bos_token_id:
input_ids = input_ids[:, 1:]
else:
# Remove any bos token that may have been added
while len(input_ids[0]) > 0 and input_ids[0][0] == shared.tokenizer.bos_token_id:
input_ids = input_ids[:, 1:]
2023-04-10 21:44:22 +02:00
# Handling truncation
if truncation_length is not None:
input_ids = input_ids[:, -truncation_length:]
2024-04-05 01:23:58 +02:00
if shared.model.__class__.__name__ in ['LlamaCppModel', 'Exllamav2Model'] or shared.args.cpu:
return input_ids
elif shared.args.deepspeed:
2024-04-11 23:42:20 +02:00
import deepspeed
return input_ids.to(deepspeed.get_accelerator().current_device_name())
elif torch.backends.mps.is_available():
device = torch.device('mps')
return input_ids.to(device)
elif is_torch_xpu_available():
return input_ids.to("xpu:0")
2024-04-11 23:42:20 +02:00
elif is_torch_npu_available():
return input_ids.to("npu:0")
else:
return input_ids.cuda()
2023-08-07 02:49:27 +02:00
def decode(output_ids, skip_special_tokens=True):
2023-09-17 16:01:34 +02:00
if shared.tokenizer is None:
raise ValueError('No tokenizer is loaded')
return shared.tokenizer.decode(output_ids, skip_special_tokens=skip_special_tokens)
2023-08-07 02:49:27 +02:00
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])
2023-09-16 04:30:44 +02:00
def get_token_ids(prompt):
tokens = encode(prompt)[0]
2023-09-17 16:01:34 +02:00
decoded_tokens = [shared.tokenizer.decode([i]) for i in tokens]
2023-09-16 04:30:44 +02:00
output = ''
for row in list(zip(tokens, decoded_tokens)):
2023-09-17 16:01:34 +02:00
output += f"{str(int(row[0])).ljust(5)} - {repr(row[1])}\n"
2023-09-16 04:30:44 +02:00
return output
2023-08-07 02:49:27 +02:00
def get_max_prompt_length(state):
return state['truncation_length'] - state['max_new_tokens']
def generate_reply_wrapper(question, state, stopping_strings=None):
"""
Returns formatted outputs for the UI
"""
reply = question if not shared.is_seq2seq else ''
yield formatted_outputs(reply, shared.model_name)
for reply in generate_reply(question, state, stopping_strings, is_chat=False, escape_html=True, for_ui=True):
2023-08-07 02:49:27 +02:00
if not shared.is_seq2seq:
reply = question + reply
yield formatted_outputs(reply, shared.model_name)
def formatted_outputs(reply, model_name):
2024-04-05 01:10:47 +02:00
return html.unescape(reply), 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)
elif is_torch_xpu_available():
torch.xpu.manual_seed_all(seed)
2024-04-11 23:42:20 +02:00
elif is_torch_npu_available():
torch.npu.manual_seed_all(seed)
2023-12-05 04:00:40 +01:00
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
def apply_stopping_strings(reply, all_stop_strings):
stop_found = False
for string in all_stop_strings:
idx = reply.find(string)
if idx != -1:
reply = reply[:idx]
stop_found = True
break
if not stop_found:
# If something like "\nYo" is generated just before "\nYou:"
# is completed, trim it
for string in all_stop_strings:
for j in range(len(string) - 1, 0, -1):
if reply[-j:] == string[:j]:
reply = reply[:-j]
break
else:
continue
break
return reply, stop_found
def get_reply_from_output_ids(output_ids, state=None, starting_from=0):
reply = decode(output_ids[starting_from:], state['skip_special_tokens'] if state else True)
# Handle tokenizers that do not add the leading space for the first token
if (hasattr(shared.tokenizer, 'convert_ids_to_tokens') and len(output_ids) > starting_from) and not reply.startswith(' '):
first_token = shared.tokenizer.convert_ids_to_tokens(int(output_ids[starting_from]))
if isinstance(first_token, (bytes,)):
first_token = first_token.decode('utf8')
if first_token.startswith(''):
reply = ' ' + reply
2023-12-05 04:00:40 +01:00
return reply
def generate_reply_HF(question, original_question, seed, state, stopping_strings=None, is_chat=False):
generate_params = {}
for k in ['max_new_tokens', 'temperature', 'temperature_last', 'dynamic_temperature', 'dynatemp_low', 'dynatemp_high', 'dynatemp_exponent', 'smoothing_factor', 'smoothing_curve', 'top_p', 'min_p', 'top_k', 'repetition_penalty', 'presence_penalty', 'frequency_penalty', 'repetition_penalty_range', 'typical_p', 'tfs', 'top_a', 'guidance_scale', 'penalty_alpha', 'mirostat_mode', 'mirostat_tau', 'mirostat_eta', 'do_sample', 'encoder_repetition_penalty', 'no_repeat_ngram_size', 'dry_multiplier', 'dry_base', 'dry_allowed_length', 'dry_sequence_breakers']:
if k in state:
generate_params[k] = state[k]
if isinstance(state['sampler_priority'], list) and len(state['sampler_priority']) > 0:
generate_params['sampler_priority'] = state['sampler_priority']
elif isinstance(state['sampler_priority'], str) and state['sampler_priority'].strip() != '':
generate_params['sampler_priority'] = [x.strip() for x in state['sampler_priority'].replace('\n', ',').split(',') if x.strip()]
if state['negative_prompt'] != '':
generate_params['negative_prompt_ids'] = encode(state['negative_prompt'])
if state['prompt_lookup_num_tokens'] > 0:
generate_params['prompt_lookup_num_tokens'] = state['prompt_lookup_num_tokens']
for k in ['epsilon_cutoff', 'eta_cutoff']:
if state[k] > 0:
generate_params[k] = state[k] * 1e-4
if state['ban_eos_token']:
generate_params['suppress_tokens'] = [shared.tokenizer.eos_token_id]
2023-09-15 23:27:27 +02:00
if state['custom_token_bans']:
to_ban = [int(x) for x in state['custom_token_bans'].split(',')]
if len(to_ban) > 0:
if generate_params.get('suppress_tokens', None):
generate_params['suppress_tokens'] += to_ban
else:
generate_params['suppress_tokens'] = to_ban
generate_params.update({'use_cache': not shared.args.no_cache})
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))
if state['auto_max_new_tokens']:
generate_params['max_new_tokens'] = state['truncation_length'] - input_ids.shape[-1]
2023-04-25 00:24:12 +02:00
# Add the encoded tokens to generate_params
2023-06-06 12:42:23 +02:00
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:
2023-04-07 16:14:32 +02:00
generate_params.update({'inputs_embeds': inputs_embeds})
# Stopping criteria / eos token
eos_token_ids = [shared.tokenizer.eos_token_id] if shared.tokenizer.eos_token_id is not None else []
generate_params['eos_token_id'] = eos_token_ids
generate_params['stopping_criteria'] = transformers.StoppingCriteriaList()
generate_params['stopping_criteria'].append(_StopEverythingStoppingCriteria())
# Logits processor
processor = state.get('logits_processor', LogitsProcessorList([]))
2023-09-19 22:13:13 +02:00
if not isinstance(processor, LogitsProcessorList):
processor = LogitsProcessorList([processor])
# Grammar
if state['grammar_string'].strip() != '':
grammar = initialize_grammar(state['grammar_string'])
grammar_processor = GrammarConstrainedLogitsProcessor(grammar)
processor.append(grammar_processor)
apply_extensions('logits_processor', processor, input_ids)
generate_params['logits_processor'] = processor
if shared.args.verbose:
logger.info("GENERATE_PARAMS=")
filtered_params = {key: value for key, value in generate_params.items() if not isinstance(value, torch.Tensor)}
pprint.PrettyPrinter(indent=4, sort_dicts=False).pprint(filtered_params)
print()
logger.info("PROMPT=")
2024-05-18 18:57:00 +02:00
print_prompt(decode(input_ids[0], skip_special_tokens=False))
# Handle StreamingLLM for llamacpp_HF
if shared.model.__class__.__name__ == 'LlamacppHF' and shared.args.streaming_llm:
2024-03-09 06:39:02 +01:00
tmp = process_llamacpp_cache(shared.model.model, input_ids[-1].tolist(), shared.model.model._input_ids.tolist())
shared.model.past_seq = torch.tensor(tmp)
shared.model.save_cache()
t0 = time.time()
2023-03-12 06:31:45 +01:00
try:
if not is_chat and not shared.is_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()
starting_from = 0 if shared.is_seq2seq else len(input_ids[0])
yield get_reply_from_output_ids(output, state, starting_from=starting_from)
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
2023-06-17 02:44:56 +02:00
def generate_with_callback(callback=None, *args, **kwargs):
2023-03-12 06:31:45 +01:00
kwargs['stopping_criteria'].append(Stream(callback_func=callback))
clear_torch_cache()
with torch.no_grad():
shared.model.generate(**kwargs)
def generate_with_streaming(**kwargs):
2023-06-17 02:44:56 +02:00
return Iteratorize(generate_with_callback, [], kwargs, callback=None)
2023-03-12 06:31:45 +01:00
with generate_with_streaming(**generate_params) as generator:
2023-12-05 04:00:40 +01:00
cumulative_reply = ''
starting_from = 0 if shared.is_seq2seq else len(input_ids[0])
2023-03-12 06:31:45 +01:00
for output in generator:
if output[-1] in eos_token_ids:
2023-03-12 06:31:45 +01:00
break
new_content = get_reply_from_output_ids(output, state, starting_from=starting_from)
# check the partial unicode character
if chr(0xfffd) in new_content:
continue
cumulative_reply += new_content
2023-12-05 04:00:40 +01:00
starting_from = len(output)
yield cumulative_reply
except Exception:
traceback.print_exc()
finally:
t1 = time.time()
original_tokens = len(original_input_ids[0])
new_tokens = len(output) - (original_tokens if not shared.is_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
def generate_reply_custom(question, original_question, seed, state, stopping_strings=None, is_chat=False):
2023-08-07 02:49:27 +02:00
"""
For models that do not use the transformers library for sampling
"""
seed = set_manual_seed(state['seed'])
t0 = time.time()
2023-05-19 19:46:18 +02:00
reply = ''
try:
2023-05-11 20:37:04 +02:00
if not is_chat:
yield ''
if not state['stream']:
2023-06-17 01:35:38 +02:00
reply = shared.model.generate(question, state)
yield reply
else:
2023-06-17 01:35:38 +02:00
for reply in shared.model.generate_with_streaming(question, state):
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
2024-05-18 18:57:00 +02:00
def print_prompt(prompt, max_chars=2000):
DARK_YELLOW = "\033[38;5;3m"
RESET = "\033[0m"
if len(prompt) > max_chars:
half_chars = max_chars // 2
hidden_len = len(prompt[half_chars:-half_chars])
hidden_msg = f"{DARK_YELLOW}[...{hidden_len} characters hidden...]{RESET}"
print(prompt[:half_chars] + hidden_msg + prompt[-half_chars:])
else:
print(prompt)
print()