Spaces:
Runtime error
Runtime error
Delete tokenization_xgen.py
Browse files- tokenization_xgen.py +0 -246
tokenization_xgen.py
DELETED
@@ -1,246 +0,0 @@
|
|
1 |
-
# Copyright (c) 2023, salesforce.com, inc.
|
2 |
-
# All rights reserved.
|
3 |
-
# SPDX-License-Identifier: Apache-2.0
|
4 |
-
# For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/Apache-2.0
|
5 |
-
"""Tokenization classes for xgen."""
|
6 |
-
|
7 |
-
from typing import List, Optional
|
8 |
-
|
9 |
-
from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
|
10 |
-
from transformers.utils import logging
|
11 |
-
|
12 |
-
try:
|
13 |
-
import tiktoken
|
14 |
-
except ModuleNotFoundError as e:
|
15 |
-
raise ModuleNotFoundError("XGen requires the installation of tiktoken. Please install it via `pip install tiktoken`.") from e
|
16 |
-
|
17 |
-
|
18 |
-
logger = logging.get_logger(__name__)
|
19 |
-
|
20 |
-
MAX_MODEL_INPUT_SIZES = {
|
21 |
-
"Salesforce/xgen-7b-4k-base": 4096,
|
22 |
-
"Salesforce/xgen-7b-8k-base": 8192,
|
23 |
-
"Salesforce/xgen-7b-4k-inst": 4096,
|
24 |
-
"Salesforce/xgen-7b-8k-inst": 8192
|
25 |
-
}
|
26 |
-
|
27 |
-
|
28 |
-
def tiktoken_tokenizer(base="gpt2", pad_token=None, add_special=True):
|
29 |
-
if not add_special:
|
30 |
-
return tiktoken.get_encoding(base)
|
31 |
-
|
32 |
-
def include_whitespace(n_min=2, n_max=20):
|
33 |
-
whitespaces = [" " * n for n in reversed(range(n_min, n_max))]
|
34 |
-
return whitespaces
|
35 |
-
|
36 |
-
def include_tabs(n_min=2, n_max=20):
|
37 |
-
tabs = ["\t" * n for n in reversed(range(n_min, n_max))]
|
38 |
-
return tabs
|
39 |
-
|
40 |
-
def include_fim_tokens():
|
41 |
-
fim_tokens = [
|
42 |
-
"<fim_prefix>",
|
43 |
-
"<fim_middle>",
|
44 |
-
"<fim_suffix>",
|
45 |
-
"<fim_pad>",
|
46 |
-
"<filename>",
|
47 |
-
"<gh_stars>",
|
48 |
-
"<issue_start>",
|
49 |
-
"<issue_comment>",
|
50 |
-
"<issue_closed>",
|
51 |
-
"<jupyter_start>",
|
52 |
-
"<jupyter_text>",
|
53 |
-
"<jupyter_code>",
|
54 |
-
"<jupyter_output>",
|
55 |
-
"<empty_output>",
|
56 |
-
"<commit_before>",
|
57 |
-
"<commit_msg>",
|
58 |
-
"<commit_after>",
|
59 |
-
"<reponame>"
|
60 |
-
]
|
61 |
-
return fim_tokens
|
62 |
-
|
63 |
-
def include_additional_tokens():
|
64 |
-
tokens = []
|
65 |
-
tokens += [f"<dummy_{i}>" for i in range(4)]
|
66 |
-
tokens.append("<sep>") # 50317
|
67 |
-
tokens.append("<eom>") # 50318
|
68 |
-
tokens += [f"<mask_{i}>" for i in reversed(range(1, 51199-50318+1))]
|
69 |
-
return tokens
|
70 |
-
|
71 |
-
add_whitespaces = include_whitespace(n_min=2, n_max=32)
|
72 |
-
add_tabs = include_tabs(n_min=2, n_max=10)
|
73 |
-
fim_tokens = include_fim_tokens()
|
74 |
-
additional_tokens = include_additional_tokens()
|
75 |
-
|
76 |
-
tokenizer = tiktoken.get_encoding(base)
|
77 |
-
|
78 |
-
idx = tokenizer.n_vocab
|
79 |
-
|
80 |
-
bpe_ranks = tokenizer._mergeable_ranks
|
81 |
-
|
82 |
-
for wsp in add_whitespaces:
|
83 |
-
bpe_ranks[bytes(wsp, 'ascii')] = idx
|
84 |
-
idx += 1
|
85 |
-
for t in add_tabs:
|
86 |
-
bpe_ranks[bytes(t, 'ascii')] = idx
|
87 |
-
idx += 1
|
88 |
-
|
89 |
-
special_tokens = dict()
|
90 |
-
|
91 |
-
for sp in fim_tokens:
|
92 |
-
special_tokens[sp] = idx
|
93 |
-
idx += 1
|
94 |
-
for sp in additional_tokens:
|
95 |
-
special_tokens[sp] = idx
|
96 |
-
idx += 1
|
97 |
-
|
98 |
-
if pad_token and pad_token not in tokenizer._special_tokens and pad_token not in special_tokens:
|
99 |
-
special_tokens[pad_token] = idx
|
100 |
-
idx += 1
|
101 |
-
# In production, load the arguments directly instead of accessing private attributes
|
102 |
-
# See openai_public.py for examples of arguments for specific encodings
|
103 |
-
enc = tiktoken.Encoding(
|
104 |
-
# If you're changing the set of special tokens, make sure to use a different name
|
105 |
-
# It should be clear from the name what behaviour to expect.
|
106 |
-
name=base.replace("base", "im"),
|
107 |
-
pat_str=tokenizer._pat_str,
|
108 |
-
mergeable_ranks=bpe_ranks,
|
109 |
-
special_tokens={
|
110 |
-
**tokenizer._special_tokens,
|
111 |
-
**special_tokens
|
112 |
-
}
|
113 |
-
)
|
114 |
-
return enc
|
115 |
-
|
116 |
-
|
117 |
-
class XgenTokenizer(PreTrainedTokenizer):
|
118 |
-
"""
|
119 |
-
Construct a Xgen tokenizer. Based on byte-level Byte-Pair-Encoding.
|
120 |
-
Args:
|
121 |
-
vocab_file (`str`):
|
122 |
-
Path to the vocabulary file.
|
123 |
-
"""
|
124 |
-
max_model_input_sizes = MAX_MODEL_INPUT_SIZES
|
125 |
-
model_input_names = ["input_ids", "attention_mask"]
|
126 |
-
|
127 |
-
def __init__(
|
128 |
-
self,
|
129 |
-
pad_token=None,
|
130 |
-
eos_token="<|endoftext|>",
|
131 |
-
add_eos_token=False,
|
132 |
-
add_special_tokens=True,
|
133 |
-
**kwargs,
|
134 |
-
):
|
135 |
-
pad_token_added = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
|
136 |
-
eos_token_added = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
|
137 |
-
self.add_eos_token = add_eos_token
|
138 |
-
self.encoder = tiktoken_tokenizer(base="gpt2", pad_token=pad_token, add_special=add_special_tokens)
|
139 |
-
super().__init__(
|
140 |
-
pad_token=pad_token_added,
|
141 |
-
eos_token=eos_token_added,
|
142 |
-
add_eos_token=add_eos_token,
|
143 |
-
add_special_tokens=add_special_tokens,
|
144 |
-
**kwargs,
|
145 |
-
)
|
146 |
-
|
147 |
-
@property
|
148 |
-
def vocab_size(self):
|
149 |
-
"""Returns vocab size"""
|
150 |
-
return self.encoder.n_vocab
|
151 |
-
|
152 |
-
def get_vocab(self):
|
153 |
-
"""Returns vocab as a dict"""
|
154 |
-
vocab = {self.encoder.decode_single_token_bytes(i): i for i in range(self.vocab_size)}
|
155 |
-
return vocab
|
156 |
-
|
157 |
-
def _tokenize(self, text, **kwargs):
|
158 |
-
"""Returns a tokenized string."""
|
159 |
-
return self.encoder.encode(text, allowed_special="all")
|
160 |
-
|
161 |
-
def _convert_token_to_id(self, token):
|
162 |
-
"""Converts a token (str) in an id using the vocab."""
|
163 |
-
if isinstance(token, str):
|
164 |
-
return self.encoder.encode_single_token(token)
|
165 |
-
else:
|
166 |
-
return token
|
167 |
-
|
168 |
-
def _convert_id_to_token(self, index):
|
169 |
-
"""Converts an index (integer) in a token (str) using the vocab."""
|
170 |
-
return self.encoder.decode_single_token_bytes(index).decode("utf-8")
|
171 |
-
|
172 |
-
def _decode(self, token_ids: List[int], skip_special_tokens: bool = False, **kwargs):
|
173 |
-
if skip_special_tokens:
|
174 |
-
token_ids = [t for t in token_ids if t not in self.all_special_ids]
|
175 |
-
return self.encoder.decode(token_ids)
|
176 |
-
|
177 |
-
def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None) -> List[int]:
|
178 |
-
"""Build model inputs from a sequence by appending eos_token_id."""
|
179 |
-
eos_token_id = [self.eos_token_id] if self.add_eos_token else []
|
180 |
-
|
181 |
-
output = token_ids_0 + eos_token_id
|
182 |
-
|
183 |
-
if token_ids_1 is not None:
|
184 |
-
output = output + token_ids_1 + eos_token_id
|
185 |
-
|
186 |
-
return output
|
187 |
-
|
188 |
-
def get_special_tokens_mask(
|
189 |
-
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None,
|
190 |
-
already_has_special_tokens: bool = False
|
191 |
-
) -> List[int]:
|
192 |
-
"""
|
193 |
-
Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
|
194 |
-
special tokens using the tokenizer `prepare_for_model` method.
|
195 |
-
Args:
|
196 |
-
token_ids_0 (`List[int]`):
|
197 |
-
List of IDs.
|
198 |
-
token_ids_1 (`List[int]`, *optional*):
|
199 |
-
Optional second list of IDs for sequence pairs.
|
200 |
-
already_has_special_tokens (`bool`, *optional*, defaults to `False`):
|
201 |
-
Whether the token list is already formatted with special tokens for the model.
|
202 |
-
Returns:
|
203 |
-
`List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
|
204 |
-
"""
|
205 |
-
if already_has_special_tokens:
|
206 |
-
return super().get_special_tokens_mask(
|
207 |
-
token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
|
208 |
-
)
|
209 |
-
|
210 |
-
eos_token_id = [1] if self.add_eos_token else []
|
211 |
-
|
212 |
-
if token_ids_1 is None:
|
213 |
-
return ([0] * len(token_ids_0)) + eos_token_id
|
214 |
-
return ([0] * len(token_ids_0)) + eos_token_id + ([0] * len(token_ids_1)) + eos_token_id
|
215 |
-
|
216 |
-
def create_token_type_ids_from_sequences(
|
217 |
-
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
|
218 |
-
) -> List[int]:
|
219 |
-
"""
|
220 |
-
Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
|
221 |
-
sequence pair mask has the following format:
|
222 |
-
```
|
223 |
-
0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
|
224 |
-
| first sequence | second sequence |
|
225 |
-
```
|
226 |
-
if token_ids_1 is None, only returns the first portion of the mask (0s).
|
227 |
-
Args:
|
228 |
-
token_ids_0 (`List[int]`):
|
229 |
-
List of ids.
|
230 |
-
token_ids_1 (`List[int]`, *optional*):
|
231 |
-
Optional second list of IDs for sequence pairs.
|
232 |
-
Returns:
|
233 |
-
`List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
|
234 |
-
"""
|
235 |
-
eos_token_id = [self.eos_token_id] if self.add_eos_token else []
|
236 |
-
|
237 |
-
output = [0] * len(token_ids_0 + eos_token_id)
|
238 |
-
|
239 |
-
if token_ids_1 is not None:
|
240 |
-
output += [1] * len(token_ids_1 + eos_token_id)
|
241 |
-
|
242 |
-
return output
|
243 |
-
|
244 |
-
# has no vocab file
|
245 |
-
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None):
|
246 |
-
return ()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|