Mirrors/text-generation-webui
1
0
Fork 0
mirror of https://github.com/oobabooga/text-generation-webui.git synced 2024-11-25 09:19:23 +01:00
Code Issues Packages Projects Releases Wiki Activity

Optimize HF text generation (#4814)

Browse Source
This commit is contained in:
oobabooga 2023-12-05 00:00:40 -03:00 committed by GitHub
parent ac9f154bcc
commit 9edb193def
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 23 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
modules/text_generation.py
View File

@ -93,9 +93,10 @@ def _generate_reply(question, state, stopping_strings=None, is_chat=False, escap
last_update = time.time() last_update = time.time()
yield reply yield reply
# Limit updates to 24 per second to not stress low latency networks # Limit updates to 24 or 5 per second to avoid lag
else: else:
if cur_time - last_update > 0.041666666666666664: min_update_interval = 0.2 if (shared.args.listen or shared.args.share) else 0.0417
if cur_time - last_update > min_update_interval:
last_update = cur_time last_update = cur_time
yield reply yield reply
@ -218,20 +219,6 @@ def fix_galactica(s):
return s return s
def get_reply_from_output_ids(output_ids, input_ids, original_question, state, is_chat=False):
if shared.is_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'])
# Prevent LlamaTokenizer from skipping a space
if type(shared.tokenizer) in [transformers.LlamaTokenizer, transformers.LlamaTokenizerFast] and len(output_ids) > 0:
if shared.tokenizer.convert_ids_to_tokens(int(output_ids[-new_tokens])).startswith('▁'):
reply = ' ' + reply
return reply
def set_manual_seed(seed): def set_manual_seed(seed):
seed = int(seed) seed = int(seed)
if seed == -1: if seed == -1:
@ -242,6 +229,7 @@ def set_manual_seed(seed):
torch.cuda.manual_seed_all(seed) torch.cuda.manual_seed_all(seed)
elif is_torch_xpu_available(): elif is_torch_xpu_available():
torch.xpu.manual_seed_all(seed) torch.xpu.manual_seed_all(seed)
return seed return seed
@ -274,6 +262,19 @@ def apply_stopping_strings(reply, all_stop_strings):
return reply, stop_found return reply, stop_found
def get_reply_from_output_ids(output_ids, state, starting_from=0):
if shared.is_seq2seq:
reply = decode(output_ids, state['skip_special_tokens'])
else:
reply = decode(output_ids[starting_from:], state['skip_special_tokens'])
# Prevent LlamaTokenizer from skipping a space
if type(shared.tokenizer) in [transformers.LlamaTokenizer, transformers.LlamaTokenizerFast] and len(output_ids) > 0:
if shared.tokenizer.convert_ids_to_tokens(int(output_ids[starting_from])).startswith('▁'):
reply = ' ' + reply
return reply
def generate_reply_HF(question, original_question, seed, state, stopping_strings=None, is_chat=False): def generate_reply_HF(question, original_question, seed, state, stopping_strings=None, is_chat=False):
generate_params = {} generate_params = {}
for k in ['max_new_tokens', 'do_sample', 'temperature', 'temperature_last', 'top_p', 'min_p', 'typical_p', 'repetition_penalty', 'presence_penalty', 'frequency_penalty', 'repetition_penalty_range', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping', 'tfs', 'top_a', 'mirostat_mode', 'mirostat_tau', 'mirostat_eta', 'guidance_scale']: for k in ['max_new_tokens', 'do_sample', 'temperature', 'temperature_last', 'top_p', 'min_p', 'typical_p', 'repetition_penalty', 'presence_penalty', 'frequency_penalty', 'repetition_penalty_range', 'encoder_repetition_penalty', 'top_k', 'min_length', 'no_repeat_ngram_size', 'num_beams', 'penalty_alpha', 'length_penalty', 'early_stopping', 'tfs', 'top_a', 'mirostat_mode', 'mirostat_tau', 'mirostat_eta', 'guidance_scale']:
@ -341,7 +342,7 @@ def generate_reply_HF(question, original_question, seed, state, stopping_strings
if cuda: if cuda:
output = output.cuda() output = output.cuda()
yield get_reply_from_output_ids(output, input_ids, original_question, state, is_chat=is_chat) yield get_reply_from_output_ids(output, state, starting_from=len(input_ids[0]))
# Stream the reply 1 token at a time. # Stream the reply 1 token at a time.
# This is based on the trick of using 'stopping_criteria' to create an iterator. # This is based on the trick of using 'stopping_criteria' to create an iterator.
@ -357,11 +358,15 @@ def generate_reply_HF(question, original_question, seed, state, stopping_strings
return Iteratorize(generate_with_callback, [], kwargs, callback=None) return Iteratorize(generate_with_callback, [], kwargs, callback=None)
with generate_with_streaming(**generate_params) as generator: with generate_with_streaming(**generate_params) as generator:
cumulative_reply = ''
starting_from = len(input_ids[0])
for output in generator: for output in generator:
if output[-1] in eos_token_ids: if output[-1] in eos_token_ids:
break break
yield get_reply_from_output_ids(output, input_ids, original_question, state, is_chat=is_chat) cumulative_reply += get_reply_from_output_ids(output, state, starting_from=starting_from)
starting_from = len(output)
yield cumulative_reply
except Exception: except Exception:
traceback.print_exc() traceback.print_exc()