utils.py 18.1 KB
Newer Older
sdtblck's avatar
sdtblck committed
1
import os
Leo Gao's avatar
Leo Gao committed
2
import re
Stephen Hogg's avatar
Stephen Hogg committed
3
import sys
4
5
6
7
8
9
import yaml
import inspect
import pathlib
import functools
import subprocess
import collections
lintangsutawika's avatar
lintangsutawika committed
10
import importlib.util
gakada's avatar
gakada committed
11
import fnmatch
12

Ethan Smith's avatar
Ethan Smith committed
13
from typing import Iterator, List, Literal, Union
14

15
import gc
16
import torch
haileyschoelkopf's avatar
haileyschoelkopf committed
17
import transformers
sdtblck's avatar
sdtblck committed
18

19
from jinja2 import BaseLoader, Environment, StrictUndefined
20
from itertools import islice
sdtblck's avatar
sdtblck committed
21

22
from lm_eval.logger import eval_logger
sdtblck's avatar
sdtblck committed
23
24


25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
def escaped_split(text, sep_char, maxsplit=-1):
    """Split text into a list on occurrences of the given separation
    character `sep_char`. The separation character may be escaped by a
    backslash to avoid splitting at that location.

    The separation character must be a string of size 1.

    If `maxsplit` is given, at most `maxsplit` splits are done (thus,
    the list will have at most `maxsplit + 1` elements). If `maxsplit`
    is not specified or less than 0, then there is no limit on the
    number of splits (all possible splits are made).
    """
    assert (
        len(sep_char) == 1
    ), "separation string must be a single character for escaped splitting"

    if maxsplit == 0:
        return text
    maxsplit = max(0, maxsplit)

    return re.split(r"(?<!\\)" + sep_char, text, maxsplit)


haileyschoelkopf's avatar
haileyschoelkopf committed
48
49
50
51
52
53
54
55
def handle_arg_string(arg):
    if arg.lower() == "true":
        return True
    elif arg.lower() == "false":
        return False
    return arg


Jason Phang's avatar
gpt3  
Jason Phang committed
56
57
58
59
60
61
def simple_parse_args_string(args_string):
    """
    Parses something like
        args1=val1,arg2=val2
    Into a dictionary
    """
Jason Phang's avatar
Jason Phang committed
62
    args_string = args_string.strip()
Jason Phang's avatar
gpt3  
Jason Phang committed
63
64
    if not args_string:
        return {}
65
    arg_list = [arg for arg in args_string.split(",") if arg]
haileyschoelkopf's avatar
haileyschoelkopf committed
66
67
68
    args_dict = {
        k: handle_arg_string(v) for k, v in [arg.split("=") for arg in arg_list]
    }
Jason Phang's avatar
gpt3  
Jason Phang committed
69
    return args_dict
Leo Gao's avatar
Leo Gao committed
70

Fabrizio Milo's avatar
Fabrizio Milo committed
71

Leo Gao's avatar
Leo Gao committed
72
73
def join_iters(iters):
    for iter in iters:
Leo Gao's avatar
Leo Gao committed
74
        yield from iter
Leo Gao's avatar
Leo Gao committed
75
76


Ethan Smith's avatar
Ethan Smith committed
77
def chunks(iter, n: int = 0, fn=None):
Leo Gao's avatar
Leo Gao committed
78
    arr = []
79
    for i, x in enumerate(iter):
Leo Gao's avatar
Leo Gao committed
80
        arr.append(x)
81
        if len(arr) == (fn(i, iter) if fn else n):
Leo Gao's avatar
Leo Gao committed
82
83
            yield arr
            arr = []
Fabrizio Milo's avatar
Fabrizio Milo committed
84
85
86
87

    if arr:
        yield arr

Leo Gao's avatar
Leo Gao committed
88

89
90
91
92
93
def group(arr, fn):
    res = collections.defaultdict(list)

    for ob in arr:
        res[fn(ob)].append(ob)
Fabrizio Milo's avatar
Fabrizio Milo committed
94

95
96
    return list(res.values())

Fabrizio Milo's avatar
Fabrizio Milo committed
97

gakada's avatar
gakada committed
98
class MultiChoice:
Ethan Smith's avatar
Ethan Smith committed
99
    def __init__(self, choices) -> None:
gakada's avatar
gakada committed
100
101
102
        self.choices = choices

    # Simple wildcard support (linux filename patterns)
Ethan Smith's avatar
Ethan Smith committed
103
    def __contains__(self, values) -> bool:
gakada's avatar
gakada committed
104
        for value in values.split(","):
105
106
107
108
            if len(fnmatch.filter(self.choices, value)) == 0:
                eval_logger.info(f"Available tasks to choose:")
                for choice in self.choices:
                    eval_logger.info(f"  - {choice}")
109
                raise ValueError("'{}' is not in task list".format(value))
gakada's avatar
gakada committed
110
111
        return True

Ethan Smith's avatar
Ethan Smith committed
112
    def __iter__(self) -> Iterator:
gakada's avatar
gakada committed
113
114
115
116
117
118
119
        for choice in self.choices:
            yield choice


# Returns a list containing all values of the source_list that
# match at least one of the patterns
def pattern_match(patterns, source_list):
120
121
122
    if type(patterns) == str:
        patterns = [patterns]

gakada's avatar
gakada committed
123
124
125
126
127
128
129
    task_names = set()
    for pattern in patterns:
        for matching in fnmatch.filter(source_list, pattern):
            task_names.add(matching)
    return sorted(list(task_names))


Leo Gao's avatar
Leo Gao committed
130
131
132
133
def general_detokenize(string):
    string = string.replace(" n't", "n't")
    string = string.replace(" )", ")")
    string = string.replace("( ", "(")
Fabrizio Milo's avatar
Fabrizio Milo committed
134
135
    string = string.replace('" ', '"')
    string = string.replace(' "', '"')
Leo Gao's avatar
Fix  
Leo Gao committed
136
    string = re.sub(r" (['.,])", r"\1", string)
137
138
139
    return string


Jason Phang's avatar
Jason Phang committed
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
def get_rolling_token_windows(token_list, prefix_token, max_seq_len, context_len):
    """
    - context_len allows for a rolling window context, allowing each prediction window to potentially
      condition on some context

    :param token_list: list
        List of tokens to be PREDICTED
    :param max_seq_len: int
        max_seq_len of model (or max_seq_len we want to use)
    :param context_len: int
        Amount of desired token context for prediction. Needs to be at least 1.
    :param prefix_token: token
        Dummy token like <eos> so the first token has something to condition on
    :return: generator
        Generator of tuples
            (input_tokens, pred_tokens)
        Note: Score only the last len(pred_tokens) logits of the LM
    """
    assert 1 <= context_len <= max_seq_len
    if not token_list:
        return
    # +1 offset, going from input->preds
    pred_len = max_seq_len - context_len + 1
    predicted = 0

    # Special handling for first window: predict all tokens
    first_seq_len = min(max_seq_len, len(token_list))
Fabrizio Milo's avatar
Fabrizio Milo committed
167
    yield ([prefix_token] + token_list[: first_seq_len - 1], token_list[:first_seq_len])
Jason Phang's avatar
Jason Phang committed
168
169
170
171
172
    predicted += first_seq_len

    while predicted < len(token_list):
        window_pred_len = min(len(token_list) - predicted, pred_len)
        window_end = predicted + window_pred_len
Leo Gao's avatar
Leo Gao committed
173

Jason Phang's avatar
Jason Phang committed
174
        yield (
lintangsutawika's avatar
lintangsutawika committed
175
176
            token_list[window_end - max_seq_len - 1 : window_end - 1],
            token_list[window_end - window_pred_len : window_end],
Jason Phang's avatar
Jason Phang committed
177
178
179
        )
        predicted += window_pred_len

Fabrizio Milo's avatar
Fabrizio Milo committed
180

Leo Gao's avatar
Leo Gao committed
181
def make_disjoint_window(pair):
Fabrizio Milo's avatar
Fabrizio Milo committed
182
    """Takes output from get_rolling_token_windows and makes the context not overlap with the continuation"""
Leo Gao's avatar
Leo Gao committed
183
    a, b = pair
184
    return a[: len(a) - (len(b) - 1)], b
Fabrizio Milo's avatar
Fabrizio Milo committed
185

Jason Phang's avatar
Jason Phang committed
186

187
class Reorderer:
Ethan Smith's avatar
Ethan Smith committed
188
    def __init__(self, arr, fn) -> None:
189
190
191
        self.size = len(arr)
        arr = list(enumerate(arr))
        arr = group(arr, lambda x: fn(x[1]))
192
193
194
        # arr = [([y[0] for y in x], x[0][1]) for x in arr]
        # TODO: overhaul reorderer. It currently grouped requests by content but we don't want this
        arr = [([y[0]], x[0][1]) for x in arr for y in x]
195
196
197
        arr.sort(key=lambda x: fn(x[1]))

        self.arr = arr
Fabrizio Milo's avatar
Fabrizio Milo committed
198

199
200
    def get_reordered(self):
        return [x[1] for x in self.arr]
Fabrizio Milo's avatar
Fabrizio Milo committed
201

202
203
204
205
206
    def get_original(self, newarr):
        res = [None] * self.size
        cov = [False] * self.size

        for (inds, _), v in zip(self.arr, newarr):
Fabrizio Milo's avatar
Fabrizio Milo committed
207
            for ind in inds:
208
209
                res[ind] = v
                cov[ind] = True
Fabrizio Milo's avatar
Fabrizio Milo committed
210

211
        assert all(cov)
Fabrizio Milo's avatar
Fabrizio Milo committed
212

213
214
        return res

Fabrizio Milo's avatar
Fabrizio Milo committed
215

haileyschoelkopf's avatar
haileyschoelkopf committed
216
217
218
219
220
221
222
class Grouper:
    """
    takes an array `arr` and function `fn` and returns a dictionary
    with keys fn(ob) for each ob in `arr` and with values `self.arr[key]` a list of all
    objects in `arr` satisfying `key == fn(ob)`.
    """

Ethan Smith's avatar
Ethan Smith committed
223
    def __init__(self, arr, fn) -> None:
haileyschoelkopf's avatar
haileyschoelkopf committed
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
        # self.orig_arr = arr
        self.size = len(arr)
        arr = list(enumerate(arr))

        def group_return_dict(arr, fn):
            res = collections.defaultdict(list)

            for ob in arr:
                res[fn(ob)].append(ob)
            return res

        arr = group_return_dict(arr, lambda x: fn(x[1]))

        # self.arr has format Dict[Tuple[int, <entry from orig. arr>]]
        self.arr = arr
        self._grouped = None

    def get_grouped(self):
        # return the contents but not indices for our grouped dict.
        if self._grouped:
            return self._grouped
        grouped = {}
        for key in self.arr.keys():
            # drop the index from each element of self.arr
            grouped[key] = [y[1] for y in self.arr[key]]
        self._grouped = grouped
        return grouped

    def get_original(self, grouped_dict):
        # take in a grouped dictionary with e.g. results for each key listed
        # in the same order as the instances in `self.arr`, and
        # return the results in the same (single list) order as `self.orig_arr`.
        res = [None] * self.size
        cov = [False] * self.size
        # orig = [None] * self.size

        assert grouped_dict.keys() == self.arr.keys()

        for key in grouped_dict.keys():
            for (ind, _), v in zip(self.arr[key], grouped_dict[key]):
                res[ind] = v
                cov[ind] = True
                # orig[ind] = _

        assert all(cov)
        # assert orig == self.orig_arr

        return res


Ethan Smith's avatar
Ethan Smith committed
274
def make_table(result_dict, column: str = "results"):
275
276
277
    """Generate table of results."""
    from pytablewriter import MarkdownTableWriter, LatexTableWriter

lintangsutawika's avatar
lintangsutawika committed
278
    if column == "results":
lintangsutawika's avatar
lintangsutawika committed
279
280
281
        column_name = "Tasks"
    elif column == "groups":
        column_name = "Groups"
lintangsutawika's avatar
lintangsutawika committed
282

283
284
    md_writer = MarkdownTableWriter()
    latex_writer = LatexTableWriter()
lintangsutawika's avatar
lintangsutawika committed
285
286
287
288
    md_writer.headers = [
        column_name,
        "Version",
        "Filter",
289
        "n-shot",
lintangsutawika's avatar
lintangsutawika committed
290
291
292
293
294
        "Metric",
        "Value",
        "",
        "Stderr",
    ]
lintangsutawika's avatar
lintangsutawika committed
295
    latex_writer.headers = [
lintangsutawika's avatar
lintangsutawika committed
296
        column_name,
lintangsutawika's avatar
lintangsutawika committed
297
298
        "Version",
        "Filter",
299
        "n-shot",
lintangsutawika's avatar
lintangsutawika committed
300
301
302
303
304
        "Metric",
        "Value",
        "",
        "Stderr",
    ]
305
306
307

    values = []

lintangsutawika's avatar
lintangsutawika committed
308
    for k, dic in result_dict[column].items():
309
        version = result_dict["versions"][k]
310
        n = str(result_dict["n-shot"][k])
311
312
        for (mf), v in dic.items():
            m, _, f = mf.partition(",")
313
314
315
            if m.endswith("_stderr"):
                continue

316
317
            if m + "_stderr" + "," + f in dic:
                se = dic[m + "_stderr" + "," + f]
318
                values.append([k, version, f, n, m, "%.4f" % v, "±", "%.4f" % se])
319
            else:
320
                values.append([k, version, f, n, m, "%.4f" % v, "", ""])
321
322
323
324
325
326
327
328
329
330
331
            k = ""
            version = ""
    md_writer.value_matrix = values
    latex_writer.value_matrix = values

    # todo: make latex table look good
    # print(latex_writer.dumps())

    return md_writer.dumps()


332
333
def positional_deprecated(fn):
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
334
    A decorator to nudge users into passing only keyword args (`kwargs`) to the
335
336
    wrapped function, `fn`.
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
337

338
339
    @functools.wraps(fn)
    def _wrapper(*args, **kwargs):
Fabrizio Milo's avatar
Fabrizio Milo committed
340
341
342
        if len(args) != 1 if inspect.ismethod(fn) else 0:
            print(
                f"WARNING: using {fn.__name__} with positional arguments is "
343
                "deprecated and will be disallowed in a future version of "
Fabrizio Milo's avatar
Fabrizio Milo committed
344
345
                "lm-evaluation-harness!"
            )
346
        return fn(*args, **kwargs)
Fabrizio Milo's avatar
Fabrizio Milo committed
347

348
    return _wrapper
Stephen Hogg's avatar
Stephen Hogg committed
349

Fabrizio Milo's avatar
Fabrizio Milo committed
350

Stephen Hogg's avatar
Stephen Hogg committed
351
352
353
354
355
356
357
358
359
@positional_deprecated
def find_test_root(start_path: pathlib.Path) -> pathlib.Path:
    """
    Search upward in the directory tree to a maximum of three layers
    to find and return the package root (containing the 'tests' folder)
    """
    cur_path = start_path.resolve()
    max_layers = 3
    for _ in range(max_layers):
Fabrizio Milo's avatar
Fabrizio Milo committed
360
        if (cur_path / "tests" / "test_version_stable.py").exists():
Stephen Hogg's avatar
Stephen Hogg committed
361
362
363
            return cur_path
        else:
            cur_path = cur_path.parent.resolve()
Fabrizio Milo's avatar
Fabrizio Milo committed
364
365
366
367
    raise FileNotFoundError(
        f"Unable to find package root within {max_layers} upwards" + f"of {start_path}"
    )

Stephen Hogg's avatar
Stephen Hogg committed
368
369

@positional_deprecated
370
def run_task_tests(task_list: List[str]):
Stephen Hogg's avatar
Stephen Hogg committed
371
372
373
    """
    Find the package root and run the tests for the given tasks
    """
jon-tow's avatar
jon-tow committed
374
375
    import pytest

376
    package_root = find_test_root(start_path=pathlib.Path(__file__))
Fabrizio Milo's avatar
Fabrizio Milo committed
377
378
379
380
381
382
383
    task_string = " or ".join(task_list)
    args = [
        f"{package_root}/tests/test_version_stable.py",
        f"--rootdir={package_root}",
        "-k",
        f"{task_string}",
    ]
Stephen Hogg's avatar
Stephen Hogg committed
384
385
386
    sys.path.append(str(package_root))
    pytest_return_val = pytest.main(args)
    if pytest_return_val:
Fabrizio Milo's avatar
Fabrizio Milo committed
387
388
389
        raise ValueError(
            f"Not all tests for the specified tasks ({task_list}) ran successfully! Error code: {pytest_return_val}"
        )
390
391


392
393
394
395
396
397
def get_git_commit_hash():
    """
    Gets the git commit hash of your current repo (if it exists).
    Source: https://github.com/EleutherAI/gpt-neox/blob/b608043be541602170bfcfb8ec9bf85e8a0799e0/megatron/neox_arguments/neox_args.py#L42
    """
    try:
Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
398
        git_hash = subprocess.check_output(["git", "describe", "--always"]).strip()
399
        git_hash = git_hash.decode()
400
401
    except subprocess.CalledProcessError or FileNotFoundError:
        # FileNotFoundError occurs when git not installed on system
402
403
404
405
        git_hash = None
    return git_hash


lintangsutawika's avatar
lintangsutawika committed
406
407
408
409
def import_function(loader, node):
    function_name = loader.construct_scalar(node)
    yaml_path = os.path.dirname(loader.name)

lintangsutawika's avatar
lintangsutawika committed
410
411
412
413
    *module_name, function_name = function_name.split(".")
    if type(module_name) == list:
        module_name = ".".join(module_name)
    module_path = os.path.normpath(os.path.join(yaml_path, "{}.py".format(module_name)))
lintangsutawika's avatar
lintangsutawika committed
414
415
416
417
418
419
420
421

    spec = importlib.util.spec_from_file_location(module_name, module_path)
    module = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(module)

    function = getattr(module, function_name)
    return function

lintangsutawika's avatar
lintangsutawika committed
422

lintangsutawika's avatar
lintangsutawika committed
423
# Add the import_function constructor to the YAML loader
lintangsutawika's avatar
lintangsutawika committed
424
yaml.add_constructor("!function", import_function)
lintangsutawika's avatar
lintangsutawika committed
425
426


427
428
429
430
431
def load_yaml_config(yaml_path=None, yaml_config=None, yaml_dir=None):

    if yaml_config is None:
        with open(yaml_path, "rb") as file:
            yaml_config = yaml.full_load(file)
lintangsutawika's avatar
lintangsutawika committed
432

lintangsutawika's avatar
lintangsutawika committed
433
434
    if yaml_dir is None:
        yaml_dir = os.path.dirname(yaml_path)
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465

    assert yaml_dir is not None

    if "include" in yaml_config:
        include_path = yaml_config["include"]
        del yaml_config["include"]

        if type(include_path) == str:
            include_path = [include_path]

        # Load from the last one first
        include_path.reverse()
        final_yaml_config = {}
        for path in include_path:

            # Assumes that path is a full path.
            # If not found, assume the included yaml
            # is in the same dir as the original yaml
            if not os.path.isfile(path):
                path = os.path.join(yaml_dir, path)

            try:
                included_yaml_config = load_yaml_config(path)
                final_yaml_config.update(included_yaml_config)
            except Exception as ex:
                # If failed to load, ignore
                raise ex

        final_yaml_config.update(yaml_config)
        return final_yaml_config
    return yaml_config
lintangsutawika's avatar
lintangsutawika committed
466
467


Ethan Smith's avatar
Ethan Smith committed
468
def regex_replace(string, pattern, repl, count: int = 0):
469
470
    """Implements the `re.sub` function as a custom Jinja filter."""
    return re.sub(pattern, repl, string, count=count)
lintangsutawika's avatar
lintangsutawika committed
471

lintangsutawika's avatar
lintangsutawika committed
472

473
env = Environment(loader=BaseLoader, undefined=StrictUndefined)
474
env.filters["regex_replace"] = regex_replace
475
476


baberabb's avatar
baberabb committed
477
def apply_template(template: str, doc: dict) -> str:
478
479
    rtemplate = env.from_string(template)
    return rtemplate.render(**doc)
480
481


482
483
484
485
def create_iterator(raw_iterator, rank, world_size, limit=None):
    """
    Method for creating a (potentially) sliced and limited
    iterator from a raw document iterator. Used for splitting data
486
487
488
    among ranks in multigpu setting or only pulling a sample of documents
    """
    return islice(raw_iterator, rank, limit, world_size)
489
490


haileyschoelkopf's avatar
haileyschoelkopf committed
491
492
493
494
495
def pad_and_concat(
    max_length: int,
    tensors: List[torch.Tensor],
    padding_side: Literal["right", "left"] = "right",
):
haileyschoelkopf's avatar
haileyschoelkopf committed
496
497
498
499
    """
    Method for padding a list of tensors given the maximum tensor
    length in the batch. Used for batching inputs and continuations in
    seq2seq models.
lintangsutawika's avatar
lintangsutawika committed
500
    """
haileyschoelkopf's avatar
haileyschoelkopf committed
501
502
503
    assert (
        padding_side == "left" or padding_side == "right"
    ), f"Unrecognized padding type: '{padding_side}' not 'left' or 'right'"
haileyschoelkopf's avatar
haileyschoelkopf committed
504

lintangsutawika's avatar
lintangsutawika committed
505
    for i, tensor in enumerate(tensors):
506
507
        if len(tensor.shape) == 2:
            tensor = tensor.squeeze(0)  # squeeze, in case passed [1, seq] size
lintangsutawika's avatar
lintangsutawika committed
508
509
        tensor_len = tensor.shape[0]
        if tensor_len < max_length:
haileyschoelkopf's avatar
haileyschoelkopf committed
510
511
512
            if padding_side == "right":
                # right-pad
                tensors[i] = torch.cat(
haileyschoelkopf's avatar
haileyschoelkopf committed
513
514
515
516
517
518
519
520
521
522
                    [
                        tensor,  # [seq]
                        torch.zeros(
                            max_length - tensor_len,
                            dtype=torch.long,
                            device=tensor.device,
                        ),  # [padding_length - seq]
                    ],
                    dim=0,
                ).unsqueeze(0)
haileyschoelkopf's avatar
haileyschoelkopf committed
523
524
525
526
            else:
                # left-pad
                tensors[i] = torch.cat(
                    [
527
                        torch.zeros(
haileyschoelkopf's avatar
haileyschoelkopf committed
528
                            max_length - tensor_len,
529
530
                            dtype=torch.long,
                            device=tensor.device,
haileyschoelkopf's avatar
haileyschoelkopf committed
531
                        ),  # [padding_length - seq]
haileyschoelkopf's avatar
haileyschoelkopf committed
532
                        tensor,  # [seq]
haileyschoelkopf's avatar
haileyschoelkopf committed
533
534
535
                    ],
                    dim=0,
                ).unsqueeze(0)
lintangsutawika's avatar
lintangsutawika committed
536
537
538
        else:
            tensors[i] = tensor.unsqueeze(0)

haileyschoelkopf's avatar
haileyschoelkopf committed
539
    return torch.cat(tensors, dim=0)
haileyschoelkopf's avatar
haileyschoelkopf committed
540
541


Ethan Smith's avatar
Ethan Smith committed
542
def clear_torch_cache() -> None:
543
544
    gc.collect()
    torch.cuda.empty_cache()
haileyschoelkopf's avatar
haileyschoelkopf committed
545
546


lintangsutawika's avatar
lintangsutawika committed
547
548
549
550
551
552
553
554
555
556
def get_dtype(dtype: Union[str, torch.dtype]) -> torch.dtype:
    """Converts `dtype` from `str` to torch.dtype when possible. Does not use an instantiated HF AutoConfig"""
    if isinstance(dtype, str) and dtype != "auto":
        # Convert `str` args torch dtype: `float16` -> `torch.float16`
        _torch_dtype = getattr(torch, dtype)
    else:
        _torch_dtype = dtype
    return _torch_dtype


haileyschoelkopf's avatar
haileyschoelkopf committed
557
# Multi-token stopping criteria
haileyschoelkopf's avatar
haileyschoelkopf committed
558
559
560
561
562
563
564
565
566
class MultiTokenEOSCriteria(transformers.StoppingCriteria):
    """Criteria to stop on the specified multi-token sequence."""

    def __init__(
        self,
        sequence: str,
        tokenizer: transformers.PreTrainedTokenizer,
        initial_decoder_input_length: int,
        batch_size: int,
Ethan Smith's avatar
Ethan Smith committed
567
    ) -> None:
haileyschoelkopf's avatar
haileyschoelkopf committed
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
        self.initial_decoder_input_length = initial_decoder_input_length
        self.done_tracker = [False] * batch_size
        self.sequence = sequence
        self.sequence_ids = tokenizer.encode(sequence, add_special_tokens=False)
        self.sequence_id_len = len(self.sequence_ids)
        self.tokenizer = tokenizer

    def __call__(self, input_ids, scores, **kwargs) -> bool:
        # For efficiency, we compare the last n tokens where n is the number of tokens in the stop_sequence
        lookback_ids_batch = input_ids[:, self.initial_decoder_input_length :][
            :, -self.sequence_id_len :
        ]

        lookback_tokens_batch = self.tokenizer.batch_decode(lookback_ids_batch)

        for i, done in enumerate(self.done_tracker):
            if not done:
                self.done_tracker[i] = self.sequence in lookback_tokens_batch[i]
        return False not in self.done_tracker


def stop_sequences_criteria(
    tokenizer: transformers.PreTrainedTokenizer,
    stop_sequences: List[str],
    initial_decoder_input_length: int,
    batch_size: int,
) -> transformers.StoppingCriteriaList:
    return transformers.StoppingCriteriaList(
        [
            *[
                MultiTokenEOSCriteria(
                    sequence, tokenizer, initial_decoder_input_length, batch_size
                )
                for sequence in stop_sequences
            ],
        ]
    )