Crystalcareai
commited on
Update modeling_quiet.py
Browse files- modeling_quiet.py +80 -123
modeling_quiet.py
CHANGED
@@ -1022,9 +1022,7 @@ class QuietForCausalLM(QuietPreTrainedModel, GenerationMixin):
|
|
1022 |
seq_len += 1
|
1023 |
|
1024 |
# Update the attention mask
|
1025 |
-
if attention_mask is None:
|
1026 |
-
attention_mask = torch.ones_like(input_ids)
|
1027 |
-
else:
|
1028 |
attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)
|
1029 |
|
1030 |
# Generate the continuation
|
@@ -1059,11 +1057,12 @@ class QuietForCausalLM(QuietPreTrainedModel, GenerationMixin):
|
|
1059 |
next_token_id = torch.argmax(next_token_logits, dim=-1)
|
1060 |
|
1061 |
# Append the generated token to the input sequence
|
1062 |
-
input_ids = torch.cat([input_ids, next_token_id.unsqueeze(-1).to(input_ids.device)], dim=-1)
|
1063 |
seq_len += 1
|
1064 |
|
1065 |
# Update the attention mask
|
1066 |
-
|
|
|
1067 |
|
1068 |
# Append the end thought token to the input sequence
|
1069 |
end_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|endthought|>")
|
@@ -1071,7 +1070,8 @@ class QuietForCausalLM(QuietPreTrainedModel, GenerationMixin):
|
|
1071 |
seq_len += 1
|
1072 |
|
1073 |
# Update the attention mask
|
1074 |
-
|
|
|
1075 |
|
1076 |
# Get the hidden states before and after the thought
|
1077 |
outputs_before = self.model(
|
@@ -1090,7 +1090,7 @@ class QuietForCausalLM(QuietPreTrainedModel, GenerationMixin):
|
|
1090 |
# two new tokens: last continuation token and end thought token
|
1091 |
outputs_after = self.model(
|
1092 |
input_ids=torch.cat([next_token_id.unsqueeze(-1).to(input_ids.device), torch.tensor([[end_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1),
|
1093 |
-
attention_mask=torch.cat([attention_mask[:, -
|
1094 |
position_ids=position_ids,
|
1095 |
past_key_values=new_key_values,
|
1096 |
inputs_embeds=inputs_embeds,
|
@@ -1110,127 +1110,25 @@ class QuietForCausalLM(QuietPreTrainedModel, GenerationMixin):
|
|
1110 |
# Apply the language model head to get the final logits
|
1111 |
logits = self.lm_head(mixed_hidden_states)
|
1112 |
return logits
|
1113 |
-
|
1114 |
-
def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **kwargs):
|
1115 |
-
if attention_mask is None:
|
1116 |
-
attention_mask = torch.ones_like(input_ids)
|
1117 |
-
return {"input_ids": input_ids, "attention_mask": attention_mask}
|
1118 |
-
|
1119 |
-
def _generate_no_beam_search(
|
1120 |
-
self,
|
1121 |
-
input_ids,
|
1122 |
-
cur_len,
|
1123 |
-
max_length,
|
1124 |
-
min_length,
|
1125 |
-
do_sample,
|
1126 |
-
temperature,
|
1127 |
-
top_k,
|
1128 |
-
top_p,
|
1129 |
-
repetition_penalty,
|
1130 |
-
no_repeat_ngram_size,
|
1131 |
-
bad_words_ids,
|
1132 |
-
pad_token_id,
|
1133 |
-
eos_token_id,
|
1134 |
-
batch_size,
|
1135 |
-
attention_mask,
|
1136 |
-
use_cache,
|
1137 |
-
model_kwargs,
|
1138 |
-
):
|
1139 |
-
finished_generating = torch.zeros(batch_size, dtype=torch.bool, device=input_ids.device)
|
1140 |
-
for cur_token_idx in range(max_length):
|
1141 |
-
# Sample the next token
|
1142 |
-
new_ids = self(
|
1143 |
-
input_ids[~finished_generating],
|
1144 |
-
attention_mask=attention_mask[~finished_generating]
|
1145 |
-
)['logits']
|
1146 |
-
# Mask out the start and end thought tokens so we don't accidentally sample them
|
1147 |
-
new_ids[:, :, self.tokenizer.vocab_size:] = -float("inf")
|
1148 |
-
for list_idx, answer_idx in enumerate((~finished_generating).nonzero(as_tuple=True)[0]):
|
1149 |
-
# Find the index of the last token that is not padding
|
1150 |
-
base_answer_ids = input_ids[answer_idx]
|
1151 |
-
new_answer_ids = new_ids[list_idx]
|
1152 |
-
last_token_idx = (base_answer_ids != self.tokenizer.pad_token_id).nonzero(as_tuple=True)[0].max()
|
1153 |
-
|
1154 |
-
new_ids_sampled = torch.multinomial(
|
1155 |
-
torch.nn.functional.softmax(new_answer_ids[last_token_idx] / temperature, dim=-1), 1)
|
1156 |
-
# Assign the new id to the last token
|
1157 |
-
if last_token_idx + 1 >= len(base_answer_ids):
|
1158 |
-
# Add padding everywhere
|
1159 |
-
new_padding = torch.full((batch_size, 1), self.tokenizer.pad_token_id, dtype=torch.long,
|
1160 |
-
device=input_ids.device)
|
1161 |
-
input_ids = torch.cat([input_ids, new_padding], dim=-1)
|
1162 |
-
attention_mask = torch.cat([attention_mask, torch.zeros_like(new_padding)], dim=-1)
|
1163 |
-
attention_mask[answer_idx, last_token_idx + 1] = 1
|
1164 |
-
input_ids[answer_idx, last_token_idx + 1] = new_ids_sampled
|
1165 |
-
if new_ids_sampled == self.tokenizer.eos_token_id or new_ids_sampled == self.tokenizer.bos_token_id or new_ids_sampled == self.tokenizer.pad_token_id:
|
1166 |
-
finished_generating[answer_idx] = 1
|
1167 |
-
# Check if the end token is generated
|
1168 |
-
if new_ids_sampled == self.tokenizer.convert_tokens_to_ids("<|/assistant|>"):
|
1169 |
-
finished_generating[answer_idx] = 1
|
1170 |
-
if finished_generating.all():
|
1171 |
-
break
|
1172 |
-
return input_ids
|
1173 |
-
|
1174 |
@torch.no_grad()
|
1175 |
def generate(
|
1176 |
self,
|
1177 |
-
input_ids=
|
1178 |
-
|
1179 |
-
|
1180 |
-
|
1181 |
-
|
1182 |
-
|
1183 |
-
|
1184 |
-
|
1185 |
-
top_p=None,
|
1186 |
-
repetition_penalty=None,
|
1187 |
-
bad_words_ids=None,
|
1188 |
-
bos_token_id=None,
|
1189 |
-
pad_token_id=None,
|
1190 |
-
eos_token_id=None,
|
1191 |
-
length_penalty=None,
|
1192 |
-
no_repeat_ngram_size=None,
|
1193 |
-
num_return_sequences=None,
|
1194 |
-
attention_mask=None,
|
1195 |
-
decoder_start_token_id=None,
|
1196 |
-
use_cache=None,
|
1197 |
-
**model_kwargs,
|
1198 |
-
):
|
1199 |
-
max_length = max_length if max_length is not None else self.config.max_length
|
1200 |
-
min_length = min_length if min_length is not None else self.config.min_length
|
1201 |
-
do_sample = do_sample if do_sample is not None else self.config.do_sample
|
1202 |
-
temperature = temperature if temperature is not None else self.config.temperature
|
1203 |
-
pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
|
1204 |
-
eos_token_id = eos_token_id if eos_token_id is not None else self.config.eos_token_id
|
1205 |
-
|
1206 |
-
# if input_ids is None:
|
1207 |
-
# raise ValueError("You have to specify either input_ids")
|
1208 |
|
1209 |
-
|
1210 |
-
|
|
|
1211 |
|
1212 |
-
|
1213 |
-
|
1214 |
-
|
1215 |
-
return self._generate_no_beam_search(
|
1216 |
-
input_ids,
|
1217 |
-
cur_len=cur_len,
|
1218 |
-
max_length=max_length,
|
1219 |
-
min_length=min_length,
|
1220 |
-
do_sample=do_sample,
|
1221 |
-
temperature=temperature,
|
1222 |
-
top_k=top_k,
|
1223 |
-
top_p=top_p,
|
1224 |
-
repetition_penalty=repetition_penalty,
|
1225 |
-
no_repeat_ngram_size=no_repeat_ngram_size,
|
1226 |
-
bad_words_ids=bad_words_ids,
|
1227 |
-
pad_token_id=pad_token_id,
|
1228 |
-
eos_token_id=eos_token_id,
|
1229 |
-
batch_size=batch_size,
|
1230 |
-
attention_mask=attention_mask,
|
1231 |
-
use_cache=use_cache,
|
1232 |
-
model_kwargs=model_kwargs,
|
1233 |
-
)
|
1234 |
|
1235 |
@add_start_docstrings_to_model_forward(QUIET_INPUTS_DOCSTRING)
|
1236 |
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
@@ -1971,6 +1869,65 @@ class QuietForCausalLM(QuietPreTrainedModel, GenerationMixin):
|
|
1971 |
hidden_states=outputs.hidden_states,
|
1972 |
attentions=outputs.attentions,
|
1973 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1974 |
|
1975 |
@staticmethod
|
1976 |
def _reorder_cache(past_key_values, beam_idx):
|
|
|
1022 |
seq_len += 1
|
1023 |
|
1024 |
# Update the attention mask
|
1025 |
+
if attention_mask is not None:
|
|
|
|
|
1026 |
attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)
|
1027 |
|
1028 |
# Generate the continuation
|
|
|
1057 |
next_token_id = torch.argmax(next_token_logits, dim=-1)
|
1058 |
|
1059 |
# Append the generated token to the input sequence
|
1060 |
+
# input_ids = torch.cat([input_ids, next_token_id.unsqueeze(-1).to(input_ids.device)], dim=-1)
|
1061 |
seq_len += 1
|
1062 |
|
1063 |
# Update the attention mask
|
1064 |
+
if attention_mask is not None:
|
1065 |
+
attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)
|
1066 |
|
1067 |
# Append the end thought token to the input sequence
|
1068 |
end_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|endthought|>")
|
|
|
1070 |
seq_len += 1
|
1071 |
|
1072 |
# Update the attention mask
|
1073 |
+
if attention_mask is not None:
|
1074 |
+
attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)
|
1075 |
|
1076 |
# Get the hidden states before and after the thought
|
1077 |
outputs_before = self.model(
|
|
|
1090 |
# two new tokens: last continuation token and end thought token
|
1091 |
outputs_after = self.model(
|
1092 |
input_ids=torch.cat([next_token_id.unsqueeze(-1).to(input_ids.device), torch.tensor([[end_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1),
|
1093 |
+
attention_mask=torch.cat([attention_mask[:, -1:], torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1),
|
1094 |
position_ids=position_ids,
|
1095 |
past_key_values=new_key_values,
|
1096 |
inputs_embeds=inputs_embeds,
|
|
|
1110 |
# Apply the language model head to get the final logits
|
1111 |
logits = self.lm_head(mixed_hidden_states)
|
1112 |
return logits
|
1113 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1114 |
@torch.no_grad()
|
1115 |
def generate(
|
1116 |
self,
|
1117 |
+
input_ids: torch.LongTensor = torch.LongTensor(),
|
1118 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1119 |
+
max_new_tokens: Optional[int] = None,
|
1120 |
+
temperature: float = 1.1,
|
1121 |
+
**kwargs,
|
1122 |
+
):
|
1123 |
+
if isinstance(input_ids, str):
|
1124 |
+
input_ids = self.tokenizer(input_ids, return_tensors="pt").input_ids
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1125 |
|
1126 |
+
if attention_mask is None:
|
1127 |
+
# Create a default attention mask if not provided
|
1128 |
+
attention_mask = torch.ones_like(input_ids)
|
1129 |
|
1130 |
+
from .generate import custom_generate
|
1131 |
+
return custom_generate(self, input_ids, attention_mask=attention_mask, max_new_tokens=max_new_tokens, temperature=temperature, **kwargs)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1132 |
|
1133 |
@add_start_docstrings_to_model_forward(QUIET_INPUTS_DOCSTRING)
|
1134 |
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
|
|
1869 |
hidden_states=outputs.hidden_states,
|
1870 |
attentions=outputs.attentions,
|
1871 |
)
|
1872 |
+
|
1873 |
+
|
1874 |
+
|
1875 |
+
def prepare_inputs_for_generation(
|
1876 |
+
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
|
1877 |
+
):
|
1878 |
+
# Omit tokens covered by past_key_values
|
1879 |
+
if past_key_values is not None:
|
1880 |
+
if isinstance(past_key_values, Cache):
|
1881 |
+
cache_length = past_key_values.get_seq_length()
|
1882 |
+
past_length = past_key_values.seen_tokens
|
1883 |
+
max_cache_length = past_key_values.get_max_length()
|
1884 |
+
else:
|
1885 |
+
cache_length = past_length = past_key_values[0][0].shape[2]
|
1886 |
+
max_cache_length = None
|
1887 |
+
|
1888 |
+
# Keep only the unprocessed tokens:
|
1889 |
+
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
1890 |
+
# some of the inputs are exclusively passed as part of the cache (e.g. when passing inputs_embeds as
|
1891 |
+
# input)
|
1892 |
+
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
|
1893 |
+
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
|
1894 |
+
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
|
1895 |
+
# input_ids based on the past_length.
|
1896 |
+
elif past_length < input_ids.shape[1]:
|
1897 |
+
input_ids = input_ids[:, past_length:]
|
1898 |
+
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
|
1899 |
+
|
1900 |
+
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
|
1901 |
+
if (
|
1902 |
+
max_cache_length is not None
|
1903 |
+
and attention_mask is not None
|
1904 |
+
and cache_length + input_ids.shape[1] > max_cache_length
|
1905 |
+
):
|
1906 |
+
attention_mask = attention_mask[:, -max_cache_length:]
|
1907 |
+
|
1908 |
+
position_ids = kwargs.get("position_ids", None)
|
1909 |
+
if attention_mask is not None and position_ids is None:
|
1910 |
+
# create position_ids on the fly for batch generation
|
1911 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
1912 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
1913 |
+
if past_key_values:
|
1914 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
1915 |
+
|
1916 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
1917 |
+
if inputs_embeds is not None and past_key_values is None:
|
1918 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
1919 |
+
else:
|
1920 |
+
model_inputs = {"input_ids": input_ids}
|
1921 |
+
|
1922 |
+
model_inputs.update(
|
1923 |
+
{
|
1924 |
+
"position_ids": position_ids,
|
1925 |
+
"past_key_values": past_key_values,
|
1926 |
+
"use_cache": kwargs.get("use_cache"),
|
1927 |
+
"attention_mask": attention_mask,
|
1928 |
+
}
|
1929 |
+
)
|
1930 |
+
return model_inputs
|
1931 |
|
1932 |
@staticmethod
|
1933 |
def _reorder_cache(past_key_values, beam_idx):
|