text-generation-webui/modules/text_generation.py

255 lines
11 KiB
Python
Raw Normal View History

2023-02-25 19:50:29 +01:00
import gc
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
2023-02-23 17:28:30 +01:00
from modules.models import local_rank
2023-02-23 18:41:42 +01:00
def get_max_prompt_length(tokens):
max_length = 2048-tokens
if shared.soft_prompt:
max_length -= shared.soft_prompt_tensor.shape[1]
return max_length
def encode(prompt, tokens_to_generate=0, add_special_tokens=True):
2023-03-05 05:20:31 +01:00
if shared.is_RWKV:
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:
2023-03-06 12:45:49 +01:00
input_ids = shared.tokenizer.encode(str(prompt), return_tensors='pt', truncation=True, max_length=get_max_prompt_length(tokens_to_generate), add_special_tokens=add_special_tokens)
if 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)
2023-03-18 00:56:23 +01:00
elif torch.has_mps:
device = torch.device('mps')
return input_ids.to(device)
2023-03-06 12:45:49 +01:00
else:
return input_ids.cuda()
def decode(output_ids):
2023-03-13 03:58:25 +01:00
# Open Assistant relies on special tokens like <|endoftext|>
2023-03-16 04:46:59 +01:00
if re.match('(oasst|galactica)-*', shared.model_name.lower()):
2023-03-13 03:58:25 +01:00
return shared.tokenizer.decode(output_ids, skip_special_tokens=False)
else:
reply = shared.tokenizer.decode(output_ids, skip_special_tokens=True)
reply = reply.replace(r'<|endoftext|>', '')
return reply
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
def formatted_outputs(reply, model_name):
if not (shared.args.chat or shared.args.cai_chat):
2023-03-01 23:11:26 +01:00
if model_name.lower().startswith('galactica'):
reply = fix_galactica(reply)
return reply, reply, generate_basic_html(reply)
2023-03-01 23:11:26 +01:00
elif model_name.lower().startswith(('gpt4chan', 'gpt-4chan', '4chan')):
reply = fix_gpt4chan(reply)
return reply, 'Only applicable for GALACTICA models.', generate_4chan_html(reply)
else:
return reply, 'Only applicable for GALACTICA models.', generate_basic_html(reply)
else:
return reply
def clear_torch_cache():
2023-02-25 19:50:29 +01:00
gc.collect()
if not shared.args.cpu:
torch.cuda.empty_cache()
def generate_reply(question, 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, eos_token=None, stopping_string=None):
clear_torch_cache()
t0 = time.time()
# These models are not part of Hugging Face, so we handle them
# separately and terminate the function call earlier
2023-03-05 05:20:31 +01:00
if shared.is_RWKV:
2023-03-12 06:53:08 +01:00
try:
if shared.args.no_stream:
reply = shared.model.generate(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k)
2023-03-01 23:11:26 +01:00
yield formatted_outputs(reply, shared.model_name)
2023-03-12 06:53:08 +01:00
else:
if not (shared.args.chat or shared.args.cai_chat):
yield formatted_outputs(question, shared.model_name)
2023-03-12 06:53:08 +01:00
# RWKV has proper streaming, which is very nice.
# No need to generate 8 tokens at a time.
for reply in shared.model.generate_with_streaming(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k):
yield formatted_outputs(reply, shared.model_name)
except:
traceback.print_exc()
2023-03-12 06:53:08 +01:00
finally:
t1 = time.time()
output = encode(reply)[0]
input_ids = encode(question)
print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(input_ids[0]))/(t1-t0):.2f} tokens/s, {len(output)-len(input_ids[0])} tokens)")
return
2023-02-28 03:03:35 +01:00
original_question = question
if not (shared.args.chat or shared.args.cai_chat):
question = apply_extensions(question, "input")
if shared.args.verbose:
print(f"\n\n{question}\n--------------------\n")
input_ids = encode(question, max_new_tokens)
2023-03-08 15:26:29 +01:00
original_input_ids = input_ids
output = input_ids[0]
cuda = not any((shared.args.cpu, shared.args.deepspeed, shared.args.flexgen))
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]))
2023-03-08 16:13:40 +01:00
stopping_criteria_list = transformers.StoppingCriteriaList()
if stopping_string is not None:
2023-03-08 16:13:40 +01:00
# Copied from https://github.com/PygmalionAI/gradio-ui/blob/master/src/model.py
t = encode(stopping_string, 0, add_special_tokens=False)
2023-03-08 16:13:40 +01:00
stopping_criteria_list.append(_SentinelTokenStoppingCriteria(sentinel_token_ids=t, starting_idx=len(input_ids[0])))
generate_params = {
'use_cache': not shared.args.no_cache,
}
if not shared.args.flexgen:
generate_params.update({
"max_new_tokens": max_new_tokens,
"eos_token_id": eos_token_ids,
"stopping_criteria": stopping_criteria_list,
"do_sample": do_sample,
"temperature": temperature,
"top_p": top_p,
"typical_p": typical_p,
"repetition_penalty": repetition_penalty,
"encoder_repetition_penalty": encoder_repetition_penalty,
"top_k": top_k,
"min_length": min_length if shared.args.no_stream else 0,
"no_repeat_ngram_size": no_repeat_ngram_size,
"num_beams": num_beams,
"penalty_alpha": penalty_alpha,
"length_penalty": length_penalty,
"early_stopping": early_stopping,
})
else:
generate_params.update({
"max_new_tokens": max_new_tokens if shared.args.no_stream else 8,
"do_sample": do_sample,
"temperature": temperature,
"stop": eos_token_ids[-1],
})
if shared.args.deepspeed:
generate_params.update({"synced_gpus": True})
if shared.soft_prompt:
inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids)
generate_params.update({"inputs_embeds": inputs_embeds})
generate_params.update({"inputs": filler_input_ids})
else:
generate_params.update({"inputs": input_ids})
2023-03-12 06:31:45 +01:00
try:
# Generate the entire reply at once.
if shared.args.no_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]:]))
reply = decode(output)
if not (shared.args.chat or shared.args.cai_chat):
reply = original_question + apply_extensions(reply[len(question):], "output")
2023-03-12 06:31:45 +01:00
yield formatted_outputs(reply, shared.model_name)
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.
elif not shared.args.flexgen:
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)
if not (shared.args.chat or shared.args.cai_chat):
yield formatted_outputs(original_question, shared.model_name)
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]:]))
reply = decode(output)
if not (shared.args.chat or shared.args.cai_chat):
reply = original_question + apply_extensions(reply[len(question):], "output")
if output[-1] in eos_token_ids:
2023-03-12 06:31:45 +01:00
break
yield formatted_outputs(reply, shared.model_name)
yield formatted_outputs(reply, shared.model_name)
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(max_new_tokens//8+1):
clear_torch_cache()
with torch.no_grad():
output = shared.model.generate(**generate_params)[0]
if shared.soft_prompt:
output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))
reply = decode(output)
if not (shared.args.chat or shared.args.cai_chat):
reply = original_question + apply_extensions(reply[len(question):], "output")
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 formatted_outputs(reply, shared.model_name)
input_ids = np.reshape(output, (1, output.shape[0]))
2023-03-12 06:31:45 +01:00
if shared.soft_prompt:
inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids)
yield formatted_outputs(reply, shared.model_name)
2023-02-26 04:36:04 +01:00
except:
traceback.print_exc()
2023-03-12 06:31:45 +01:00
finally:
t1 = time.time()
print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(original_input_ids[0]))/(t1-t0):.2f} tokens/s, {len(output)-len(original_input_ids[0])} tokens)")
return