mmarch/text-generation-webui
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Add additive_repetition_penalty sampler setting. (#3627)

Browse source
This commit is contained in:
tdrussell 2023-10-23 00:28:07 -05:00 • committed by GitHub
parent 6086768309
commit 4440f87722
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
13 changed files with 34 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

7
modules/loaders.py
View file

@ -153,6 +153,7 @@ loaders_samplers = {
'tfs',
'top_a',
'repetition_penalty',
'additive_repetition_penalty',
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
@ -186,6 +187,7 @@ loaders_samplers = {
'tfs',
'top_a',
'repetition_penalty',
'additive_repetition_penalty',
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
@ -244,6 +246,7 @@ loaders_samplers = {
'tfs',
'top_a',
'repetition_penalty',
'additive_repetition_penalty',
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
@ -273,6 +276,7 @@ loaders_samplers = {
'tfs',
'top_a',
'repetition_penalty',
'additive_repetition_penalty',
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
@ -306,6 +310,7 @@ loaders_samplers = {
'tfs',
'top_a',
'repetition_penalty',
'additive_repetition_penalty',
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
@ -353,6 +358,7 @@ loaders_samplers = {
'tfs',
'top_a',
'repetition_penalty',
'additive_repetition_penalty',
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
@ -389,6 +395,7 @@ loaders_samplers = {
'tfs',
'top_a',
'repetition_penalty',
'additive_repetition_penalty',
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',

1
modules/presets.py
View file

@ -16,6 +16,7 @@ def default_preset():
'tfs': 1,
'top_a': 0,
'repetition_penalty': 1,
'additive_repetition_penalty': 0,
'repetition_penalty_range': 0,
'encoder_repetition_penalty': 1,
'no_repeat_ngram_size': 0,

23
modules/sampler_hijack.py
View file

@ -139,11 +139,12 @@ class RepetitionPenaltyLogitsProcessorWithRange(LogitsProcessor):
Copied from the transformers library
'''
def __init__(self, penalty: float, _range: int):
if not isinstance(penalty, float) or not (penalty > 0):
raise ValueError(f"`penalty` has to be a strictly positive float, but is {penalty}")
def __init__(self, penalty: float, additive_penalty: float, _range: int):
if not (penalty > 0):
raise ValueError(f"`penalty` has to be strictly positive, but is {penalty}")
self.penalty = penalty
self.additive_penalty = additive_penalty
self._range = _range
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
@ -153,6 +154,7 @@ class RepetitionPenaltyLogitsProcessorWithRange(LogitsProcessor):
# if score < 0 then repetition penalty has to be multiplied to reduce the previous token probability
score = torch.where(score < 0, score * self.penalty, score / self.penalty)
score -= self.additive_penalty
scores.scatter_(1, input_ids, score)
return scores
@ -185,14 +187,20 @@ def get_logits_warper_patch(self, generation_config):
def get_logits_processor_patch(self, **kwargs):
result = self._get_logits_processor_old(**kwargs)
repetition_penalty_range = kwargs['generation_config'].repetition_penalty_range
repetition_penalty = kwargs['generation_config'].repetition_penalty
additive_repetition_penalty = kwargs['generation_config'].additive_repetition_penalty
repetition_penalty_range = kwargs['generation_config'].repetition_penalty_range
do_rep_pen_hijack = (repetition_penalty > 1) or (additive_repetition_penalty > 0)
if do_rep_pen_hijack:
# Make sure that a RepetitionPenaltyLogitsProcessor will be created
kwargs['generation_config'].repetition_penalty = 1.1 # must set to some value > 1
if repetition_penalty_range > 0:
result = self._get_logits_processor_old(**kwargs)
if do_rep_pen_hijack:
for i in range(len(result)):
if result[i].__class__.__name__ == 'RepetitionPenaltyLogitsProcessor':
result[i] = RepetitionPenaltyLogitsProcessorWithRange(repetition_penalty, repetition_penalty_range)
result[i] = RepetitionPenaltyLogitsProcessorWithRange(repetition_penalty, additive_repetition_penalty, repetition_penalty_range)
return result
@ -205,6 +213,7 @@ def generation_config_init_patch(self, **kwargs):
self.mirostat_eta = kwargs.pop("mirostat_eta", 0.1)
self.mirostat_tau = kwargs.pop("mirostat_tau", 5)
self.repetition_penalty_range = kwargs.pop("repetition_penalty_range", 0)
self.additive_repetition_penalty = kwargs.pop("additive_repetition_penalty", 0)
def hijack_samplers():

2
modules/text_generation.py
View file

@ -273,7 +273,7 @@ def apply_stopping_strings(reply, all_stop_strings):
def generate_reply_HF(question, original_question, seed, state, stopping_strings=None, is_chat=False):
generate_params = {}
for k in ['max_new_tokens', 'do_sample', 'temperature', 'top_p', 'typical_p', 'repetition_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', 'top_p', 'typical_p', 'repetition_penalty', 'additive_repetition_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']:
generate_params[k] = state[k]
if state['negative_prompt'] != '':

1
modules/ui.py
View file

@ -105,6 +105,7 @@ def list_interface_input_elements():
'epsilon_cutoff',
'eta_cutoff',
'repetition_penalty',
'additive_repetition_penalty',
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',

1
modules/ui_parameters.py
View file

@ -31,6 +31,7 @@ def create_ui(default_preset):
shared.gradio['top_p'] = gr.Slider(0.0, 1.0, value=generate_params['top_p'], step=0.01, label='top_p')
shared.gradio['top_k'] = gr.Slider(0, 200, value=generate_params['top_k'], step=1, label='top_k')
shared.gradio['repetition_penalty'] = gr.Slider(1.0, 1.5, value=generate_params['repetition_penalty'], step=0.01, label='repetition_penalty')
shared.gradio['additive_repetition_penalty'] = gr.Slider(0, 4, value=generate_params['additive_repetition_penalty'], step=0.05, label='additive_repetition_penalty')
shared.gradio['repetition_penalty_range'] = gr.Slider(0, 4096, step=64, value=generate_params['repetition_penalty_range'], label='repetition_penalty_range')
shared.gradio['typical_p'] = gr.Slider(0.0, 1.0, value=generate_params['typical_p'], step=0.01, label='typical_p')
shared.gradio['tfs'] = gr.Slider(0.0, 1.0, value=generate_params['tfs'], step=0.01, label='tfs')