huggingface.py 24.4 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
import torch
import transformers
from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES

import copy
from tqdm import tqdm

import torch.nn.functional as F

from lm_eval import utils
from lm_eval.logger import eval_logger
from lm_eval.api.model import LM
from lm_eval.api.registry import register_model

from lm_eval.utils import MultiTokenEOSCriteria, stop_sequences_criteria

from accelerate import Accelerator
haileyschoelkopf's avatar
haileyschoelkopf committed
18
from typing import List, Union
19
20


21
@register_model("hf-auto", "hf", "huggingface")
22
class HFLM(LM):
23
24
25
26
27
28
29
    """
    An abstracted Huggingface model class. Enables usage with both models of
    `transformers.AutoModelForCausalLM` and `transformers.AutoModelForSeq2SeqLM` classes.

    Supports data-parallel multi-GPU with HF Accelerate.
    """

30
    AUTO_MODEL_CLASS = None
31
    _DEFAULT_MAX_LENGTH = 2048
haileyschoelkopf's avatar
haileyschoelkopf committed
32

33
34
35
36
37
38
    def __init__(
        self,
        device="cuda",
        pretrained="gpt2",
        revision="main",
        low_cpu_mem_usage=None,
39
        max_length=None,
40
41
42
43
44
45
46
47
48
49
50
        subfolder=None,
        tokenizer=None,
        batch_size=1,
    ):
        super().__init__()

        assert isinstance(device, str)
        assert isinstance(pretrained, str)
        assert isinstance(batch_size, int)

        gpus = torch.cuda.device_count()
haileyschoelkopf's avatar
haileyschoelkopf committed
51

52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
        if gpus <= 1:
            if device:
                if device not in ["cuda", "cpu"]:
                    device = int(device)
                self._device = torch.device(device)
                eval_logger.info(f"Using device '{device}'")
            else:
                eval_logger.info("Device not specified")
                eval_logger.info(f"Cuda Available? {torch.cuda.is_available()}")
                self._device = (
                    torch.device("cuda")
                    if torch.cuda.is_available()
                    else torch.device("cpu")
                )
            self._rank = 0
            self._world_size = 1

        else:
            self._device = "cpu"

        # TODO: update this to be less of a hack once subfolder is fixed in HF
        revision = revision + ("/" + subfolder if subfolder is not None else "")

haileyschoelkopf's avatar
haileyschoelkopf committed
75
        # get config
76
77
78
79
80
81
82
83
        self._config = transformers.AutoConfig.from_pretrained(
            pretrained,
            revision=revision,
        )

        if getattr(self._config, "model_type") in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
            self.AUTO_MODEL_CLASS = transformers.AutoModelForCausalLM
        else:
haileyschoelkopf's avatar
haileyschoelkopf committed
84
            self.AUTO_MODEL_CLASS = transformers.AutoModelForSeq2SeqLM
85

haileyschoelkopf's avatar
haileyschoelkopf committed
86
87
88
89
        assert self.AUTO_MODEL_CLASS in [
            transformers.AutoModelForCausalLM,
            transformers.AutoModelForSeq2SeqLM,
        ]
90

91
        self._model = self.AUTO_MODEL_CLASS.from_pretrained(
92
93
            pretrained, revision=revision, low_cpu_mem_usage=low_cpu_mem_usage
        ).to(self.device)
94
        # forever after, access self._model through self.model property
95
96
97
98
99
100
101
102
        self.model.eval()

        self.tokenizer = transformers.AutoTokenizer.from_pretrained(
            pretrained if tokenizer is None else tokenizer,
            revision=revision,
        )

        self.vocab_size = self.tokenizer.vocab_size
haileyschoelkopf's avatar
haileyschoelkopf committed
103
        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
104

105
106
        self._max_length = max_length

107
108
109
110
111
112
113
        # multithreading and batching
        self.batch_size_per_gpu = batch_size  # todo: adaptive batch size

        # multigpu support with accelerate
        if gpus > 1:
            accelerator = Accelerator()
            if gpus > accelerator.num_processes:
114
                # TODO: make sure there's still never an edge case where we unintentionally default to CPU
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
                eval_logger.warning(
                    "WARNING: The number of total system GPUs does not match the number of spawned processes. "
                    "If you would like to use data parallelism, please launch the script "
                    "with 'accelerate launch *script*'. "
                    f"Current run will proceed with {accelerator.num_processes} devices."
                )
                self._rank = accelerator.local_process_index
                self._world_size = accelerator.num_processes
                # manually set model to use gpu, for case where many GPUs available but
                # only seek to use one
                self._device = (
                    torch.device(f"cuda:{accelerator.local_process_index}")
                    if torch.cuda.is_available()
                    else torch.device("cpu")
                )
                self.model.to(self.device)
            else:
haileyschoelkopf's avatar
haileyschoelkopf committed
132
                self._model = accelerator.prepare(self.model)
133
134
135
136
137
138
139
140
                self._device = torch.device(f"cuda:{accelerator.local_process_index}")
                self.accelerator = accelerator

                if self.accelerator.is_local_main_process:
                    eval_logger.info(f"Using {gpus} devices with data parallelism")

                self._rank = self.accelerator.local_process_index
                self._world_size = self.accelerator.num_processes
haileyschoelkopf's avatar
haileyschoelkopf committed
141

142
143
144
145
146
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

147
148
149
150
151
152
153
154
    @property
    def model(self):
        # returns the model, unwrapping it if using Accelerate
        if hasattr(self, "accelerator"):
            return self.accelerator.unwrap_model(self._model)
        else:
            return self._model

155
156
157
158
159
160
161
    @property
    def eot_token_id(self):
        # we use EOT because end of *text* is more accurate for what we're doing than end of *sentence*
        return self.tokenizer.eos_token_id

    @property
    def max_length(self):
162
163
164
165
166
167
168
169
170
171
172
        if self._max_length:  # if max length manually set, return it
            return self._max_length
        seqlen_config_attrs = ("n_positions", "max_position_embeddings", "n_ctx")
        for attr in seqlen_config_attrs:
            if hasattr(self.model.config, attr):
                return getattr(self.model.config, attr)
        if hasattr(self.tokenizer, "model_max_length"):
            if self.tokenizer.model_max_length == 1000000000000000019884624838656:
                return self._DEFAULT_MAX_LENGTH
            return self.tokenizer.model_max_length
        return self._DEFAULT_MAX_LENGTH
173

174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
    @property
    def max_gen_toks(self):
        return 256

    @property
    def batch_size(self):
        return self.batch_size_per_gpu

    @property
    def device(self):
        return self._device

    @property
    def rank(self):
        return self._rank

    @property
    def world_size(self):
        return self._world_size

    def tok_encode(self, string: str, left_truncate_len=None):
haileyschoelkopf's avatar
haileyschoelkopf committed
195
        """ """
196
197
198
199
200
201
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
            add_special_tokens = False
        elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
            add_special_tokens = True

        encoding = self.tokenizer.encode(string, add_special_tokens=add_special_tokens)
haileyschoelkopf's avatar
haileyschoelkopf committed
202

203
204
205
        # left-truncate the encoded context to be at most `left_truncate_len` tokens long
        if left_truncate_len:
            encoding = encoding[-left_truncate_len:]
haileyschoelkopf's avatar
haileyschoelkopf committed
206

207
208
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
    def tok_batch_encode(
        self, strings: List[str], padding_side="left", left_truncate_len=None
    ):
        # encode a batch of strings. converts to tensors and pads automatically, unlike tok_encode.
        old_padding_side = self.tokenizer.padding_side
        self.tokenizer.padding_side = padding_side

        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
            add_special_tokens = False
        elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
            add_special_tokens = True

        encoding = self.tokenizer(
            strings,
            padding="longest",
            return_tensors="pt",
            add_special_tokens=add_special_tokens,
        )
        if left_truncate_len:
            encoding["input_ids"] = encoding["input_ids"][:, -left_truncate_len:]
            encoding["attention_mask"] = encoding["attention_mask"][
                :, -left_truncate_len:
            ]
        self.tokenizer.padding_side = old_padding_side

        return encoding["input_ids"], encoding["attention_mask"]

236
237
238
239
240
241
242
243
    def tok_decode(self, tokens):
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
            return self.tokenizer.decode(tokens)
        elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
            return self.tokenizer.decode(tokens, skip_special_tokens=True)

    def _model_call(self, inps, attn_mask=None, labels=None):
        """
haileyschoelkopf's avatar
haileyschoelkopf committed
244
        :param inps: torch.Tensor
245
246
247
248
249
250
251
252
253
254
255
256
257
            A torch tensor of shape [batch, (sequence_ctx + sequence_cont)] or of shape
            [batch, sequence_ctx]. the size of sequence may vary from call to call
        :param attn_mask: torch.Tensor, optional
            A torch tensor of shape [batch, (sequence_ctx + sequence_cont)]. Only passed
            (and must be passed) if self.AUTO_MODEL_CLASS is transformers.AutoModelForSeq2SeqLM
        :param labels: torch.Tensor, optional
            A torch tensor of shape [batch, (sequence_ctx + sequence_cont)]. Only passed
            (and must be passed) if self.AUTO_MODEL_CLASS is transformers.AutoModelForSeq2SeqLM
        :return
            A torch tensor of shape [batch, sequence, vocab] with the
        logits returned from the model's decoder
        """
        with torch.no_grad():
258
259
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
260
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
haileyschoelkopf's avatar
haileyschoelkopf committed
261
262
263
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
264
265
266
267
268
269
270
271
272
273
274
275
276
            else:
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                return self.model(inps).logits

    def _model_generate(self, context, max_length, stop, **generation_kwargs):
        # we require users to pass do_sample=True explicitly
        # for non-greedy gen. This should be reevaluated when considering beam search.
        if "do_sample" not in generation_kwargs.keys():
            generation_kwargs["do_sample"] = False
        # build stopping criteria
        stopping_criteria = stop_sequences_criteria(
            self.tokenizer, stop, 1, context.shape[0]
        )
277
278
279
280
281
282
283
284
        return self.model.generate(
            context,
            max_length=max_length,
            stopping_criteria=stopping_criteria,
            pad_token_id=self.eot_token_id,
            use_cache=True,
            **generation_kwargs,
        )
285
286
287

    def _select_cont_toks(self, logits, contlen=None, inplen=None):
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
288
289
290
            assert (
                contlen and inplen
            ), "Must pass input len and cont. len to select scored logits for causal LM"
291
292
293
294
            # discard right-padding.
            # also discard the input/context tokens. we'll only score continuations.
            logits = logits[inplen - contlen : inplen]
        elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
295
296
297
298
            assert (
                contlen and not inplen
            ), "Selecting scored logits for Seq2SeqLM requires only cont. len"
            # only discard right-padding.
299
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
300
301
            logits = logits[:contlen]

302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
        return logits

    def loglikelihood(self, requests):
        new_reqs = []
        for context, continuation in [req.args for req in requests]:
            if context == "":
                # end of text as context
                context_enc = [self.eot_token_id]
            else:
                context_enc = self.tok_encode(context)

            continuation_enc = self.tok_encode(continuation)

            new_reqs.append(((context, continuation), context_enc, continuation_enc))

        return self._loglikelihood_tokens(new_reqs)

    def loglikelihood_rolling(self, requests):
        loglikelihoods = []
        for (string,) in tqdm([req.args for req in requests], disable=(self.rank != 0)):
            rolling_token_windows = list(
                map(
                    utils.make_disjoint_window,
                    utils.get_rolling_token_windows(
                        token_list=self.tok_encode(string),
haileyschoelkopf's avatar
haileyschoelkopf committed
327
                        prefix_token=self.eot_token_id,
328
329
330
331
332
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
333
334

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
            rolling_token_windows = [(None,) + x for x in rolling_token_windows]

            pad_amnt = 0
            if self.world_size > 1:
                # We pad out the external document-level iterator so the inner iterator doesn't hang
                mytensor = torch.tensor(len(rolling_token_windows), device=self.device)
                gathered = (
                    self.accelerator.gather(mytensor).cpu().detach().numpy().tolist()
                )

                pad_amnt = max(gathered) - gathered[self.rank]
                if pad_amnt > 0:
                    rolling_token_windows += pad_amnt * [rolling_token_windows[0]]

            string_nll = self._loglikelihood_tokens(
                rolling_token_windows, disable_tqdm=True
            )

            if (self.world_size > 1) and (pad_amnt > 0):
                string_nll = [x[0] for x in string_nll[:-pad_amnt]]
            else:
                # discard is_greedy
                string_nll = [x[0] for x in string_nll]

            string_nll = sum(string_nll)
            loglikelihoods.append(string_nll)

        return loglikelihoods

    def _loglikelihood_tokens(self, requests, disable_tqdm=False):
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

        def _collate(x):
            # the negative sign on len(toks) sorts descending - this has a few advantages:
            # - time estimates will always be over not underestimates, which is more useful for planning
            # - to know the size of a batch when going through the list, you know the first one is always the batch
            #   padded context length. this is useful to simplify the batching logic and more importantly to make
            #   automatic adaptive batches much much easier to implement
            # - any OOMs will happen right away rather than near the end

            toks = x[1] + x[2]
            return -len(toks), tuple(toks)

        # TODO: automatic (variable) batch size detection for vectorization
        re_ord = utils.Reorderer(requests, _collate)
        for chunk in utils.chunks(
            tqdm(re_ord.get_reordered(), disable=(disable_tqdm or (self.rank != 0))),
            self.batch_size,
        ):

            inps = []
            cont_toks_list = []
            inplens = []

            conts = []
            encoder_attns = []

            padding_len_inp = None
            padding_len_cont = None
            # because vectorizing is annoying, we first convert each (context, continuation) pair to padded
            # tensors, then we pack them together into a batch, call the model, and then pick it all apart
            # again because vectorizing is annoying

            for _, context_enc, continuation_enc in chunk:
                # sanity check
                assert len(context_enc) > 0
                assert len(continuation_enc) > 0
                assert len(continuation_enc) <= self.max_length

haileyschoelkopf's avatar
haileyschoelkopf committed
405
                # how this all works (illustrated on a causal decoder-only setup):
406
407
408
409
410
411
412
413
414
415
416
                #          CTX      CONT
                # inp    0 1 2 3|4 5 6 7 8 9   <- last token is deleted by inp[:, :-1]
                # model  \               \
                # logits   1 2 3|4 5 6 7 8 9   <- the ctx half gets tossed out by the
                # cont_toks      4 5 6 7 8 9      [:, -len(continuation_enc):, :self.vocab_size] slice

                # when too long to fit in context, truncate from the left
                if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
                    inp = torch.tensor(
                        (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1],
                        dtype=torch.long,
417
418
                        device=self.device,
                    )
419
420
421
422
423
                    (inplen,) = inp.shape
                elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                    inp = torch.tensor(
                        (context_enc)[-self.max_length :],
                        dtype=torch.long,
haileyschoelkopf's avatar
haileyschoelkopf committed
424
                        device=self.device,
425
                    )
426
                    (inplen,) = inp.shape
427
428
429
430

                    # build encoder attn masks
                    encoder_attns.append(torch.ones_like(inp))

431
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
432
                        (continuation_enc)[-self.max_length :],
433
434
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
435
                        dtype=torch.long,
436
437
                        device=self.device,
                    )
438
439
                    (contlen,) = cont.shape

440
441
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
442
443
444
445
446
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
447

haileyschoelkopf's avatar
haileyschoelkopf committed
448
449
450
451
452
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
453
454
455
456

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
457

458
459
460
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
461
462
463
                batched_inps = utils.pad_and_concat(
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
464
465
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                # TODO: left-pad encoder inps and mask?
haileyschoelkopf's avatar
haileyschoelkopf committed
466
467
468
469
470
471
472
473
474
475
476
477
478
                batched_inps = utils.pad_and_concat(
                    padding_len_inp, inps
                )  # [batch, padding_len_inp]
                batched_conts = utils.pad_and_concat(
                    padding_len_cont, conts
                )  # [batch, padding_len_cont]
                batched_encoder_mask = utils.pad_and_concat(
                    padding_len_inp, encoder_attns
                )  # [batch, padding_len_inp]
                call_kwargs = {
                    "attn_mask": batched_encoder_mask,
                    "labels": batched_conts,
                }
479
480
481

            multi_logits = F.log_softmax(
                self._model_call(batched_inps, **call_kwargs), dim=-1
482
            ).cpu()  # [batch, padding_length (inp or cont), vocab]
483
484
485
486
487
488
489

            for (cache_key, _, _), logits, inplen, cont_toks in zip(
                chunk, multi_logits, inplens, cont_toks_list
            ):

                # Slice to original seq length
                contlen = len(cont_toks)
haileyschoelkopf's avatar
haileyschoelkopf committed
490
                # take only logits in the continuation
491
                # (discard context toks if decoder-only ; discard right-padding)
haileyschoelkopf's avatar
haileyschoelkopf committed
492
493
494
495
496
                ctx_len = (
                    inplen
                    if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                    else None
                )
497
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
498
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517

                # Check if per-token argmax is exactly equal to continuation
                greedy_tokens = logits.argmax(dim=-1)
                cont_toks = torch.tensor(cont_toks, dtype=torch.long).unsqueeze(
                    0
                )  # [1, seq]
                max_equal = (greedy_tokens == cont_toks).all()

                # Obtain log-probs at the corresponding continuation token indices
                # last_token_slice = logits[:, -1, :].squeeze(0).tolist()
                logits = torch.gather(logits, 2, cont_toks.unsqueeze(-1)).squeeze(
                    -1
                )  # [1, seq]

                # Answer: (log prob, is-exact-match)
                answer = (float(logits.sum()), bool(max_equal))

                res.append(answer)

haileyschoelkopf's avatar
haileyschoelkopf committed
518
519
                self.cache_hook.add_partial("loglikelihood", cache_key, answer)

520
521
522
523
524
525
526
        return re_ord.get_original(res)

    def greedy_until(self, requests):
        res = []

        def _collate(x):
            toks = self.tok_encode(x[0])
haileyschoelkopf's avatar
haileyschoelkopf committed
527
            return -len(toks), x[0]
528
529
530

        re_ord = utils.Reorderer([req.args for req in requests], _collate)

haileyschoelkopf's avatar
haileyschoelkopf committed
531
532
533
534
535
536
        for chunk in utils.chunks(
            tqdm(
                re_ord.get_reordered(),
                disable=(self.rank != 0),
            ),
            self.batch_size,
haileyschoelkopf's avatar
haileyschoelkopf committed
537
        ):
haileyschoelkopf's avatar
haileyschoelkopf committed
538
539
540
541
            contexts, all_gen_kwargs = zip(*chunk)
            gen_kwargs = all_gen_kwargs[
                0
            ]  # TODO: handle case where not all gen kwargs are same
542
543
            until = None
            if isinstance(gen_kwargs, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
544
545
546
                kwargs = copy.deepcopy(gen_kwargs)  # edge case for repeats > 1
                if "until" in kwargs.keys():
                    until = kwargs.pop("until")
547
                    if isinstance(until, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
548
                        until = [kwargs]
549
550
                    elif not isinstance(until, list):
                        raise ValueError(
haileyschoelkopf's avatar
haileyschoelkopf committed
551
                            f"Expected `generation_kwargs['until']` to be of type Union[str,list] but got {until}"
552
553
554
                        )
            else:
                raise ValueError(
haileyschoelkopf's avatar
haileyschoelkopf committed
555
                    f"Expected `generation_kwargs` to be of type `dict` but got {kwargs}"
556
557
558
                )
            if not until:
                until = [self.tok_decode(self.eot_token_id)]
haileyschoelkopf's avatar
haileyschoelkopf committed
559
560
            if "max_gen_toks" in kwargs.keys():
                max_gen_toks = kwargs.pop("max_gen_toks")
561
562
563
564
565
566
567
568
569
570
571
572
573
            else:
                max_gen_toks = self.max_gen_toks
            # first stop sequence is used to halt generation upon encountering
            (primary_until) = until[0]

            # set the max length in tokens of inputs ("context_enc")
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
                # max len for inputs = max length, minus room to generate the max new tokens
                max_ctx_len = self.max_length - max_gen_toks
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                # max len for inputs = encoder's whole max_length
                max_ctx_len = self.max_length

haileyschoelkopf's avatar
haileyschoelkopf committed
574
575
            context_enc, attn_masks = self.tok_batch_encode(
                contexts, left_truncate_len=max_ctx_len
576
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
577
578
579
580
581
582
583
584
585
            context_enc = context_enc.to(self.device)
            attn_masks = attn_masks.to(self.device)
            #     [self.tok_encode(context, left_truncate_len=max_ctx_len)],
            #     device=self.device,
            # ) for context in contexts]

            # padding_len = max([context.shape[1] for context in context_enc])
            # self.tokenizer.batch_encod
            # context_enc = utils.pad_and_concat(padding_len, context_enc, padding_side="left")
586
587

            cont = self._model_generate(
haileyschoelkopf's avatar
haileyschoelkopf committed
588
                context=context_enc,
haileyschoelkopf's avatar
haileyschoelkopf committed
589
                attention_mask=attn_masks,
haileyschoelkopf's avatar
haileyschoelkopf committed
590
                max_length=context_enc.shape[1] + max_gen_toks,
591
                stop=primary_until,
haileyschoelkopf's avatar
haileyschoelkopf committed
592
                **kwargs,
593
            )
594

haileyschoelkopf's avatar
haileyschoelkopf committed
595
596
597
598
599
            cont_toks_list = cont.tolist()
            for cont_toks, context in zip(cont_toks_list, contexts):
                # discard context + left-padding toks if using causal decoder-only LM
                if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
                    cont_toks = cont_toks[context_enc.shape[1] :]
600

haileyschoelkopf's avatar
haileyschoelkopf committed
601
                s = self.tok_decode(cont_toks)
602

haileyschoelkopf's avatar
haileyschoelkopf committed
603
604
605
606
                # use secondary stop seqs to cut off should-have-been-stopped content post-hoc
                for term in until:
                    if len(term) > 0:  # ignore '' separator, for seq2seq case where
                        s = s.split(term)[0]
607

haileyschoelkopf's avatar
haileyschoelkopf committed
608
                res.append(s)
609

haileyschoelkopf's avatar
haileyschoelkopf committed
610
                self.cache_hook.add_partial("greedy_until", (context, gen_kwargs), s)
haileyschoelkopf's avatar
haileyschoelkopf committed
611

haileyschoelkopf's avatar
haileyschoelkopf committed
612
        return re_ord.get_original(res)