huggingface.py 50.6 KB
Newer Older
1
import copy
2
import os
Jeevan's avatar
Jeevan committed
3
from datetime import timedelta
4
5
6
from pathlib import Path
from typing import List, Literal, Optional, Tuple, Union

7
import torch
8
import torch.nn.functional as F
9
import transformers
Jeevan's avatar
Jeevan committed
10
11
12
13
14
15
from accelerate import (
    Accelerator,
    DistributedType,
    InitProcessGroupKwargs,
    find_executable_batch_size,
)
16
17
18
19
from packaging import version
from peft import PeftModel
from peft import __version__ as PEFT_VERSION
from tqdm import tqdm
20
21
22
23
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
)
24
25

from lm_eval import utils
baberabb's avatar
baberabb committed
26
from lm_eval.api.instance import Instance
27
from lm_eval.api.model import TemplateLM
28
from lm_eval.api.registry import register_model
29
30
31
32
33
34
35
from lm_eval.models.utils import (
    Collator,
    clear_torch_cache,
    get_dtype,
    pad_and_concat,
    stop_sequences_criteria,
)
36

37

38
eval_logger = utils.eval_logger
39

lintangsutawika's avatar
lintangsutawika committed
40

41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
def _get_accelerate_args(
    device_map_option: Optional[str] = "auto",
    max_memory_per_gpu: Optional[Union[int, str]] = None,
    max_cpu_memory: Optional[Union[int, str]] = None,
    offload_folder: Optional[str] = "./offload",
) -> dict:
    """Returns the kwargs needed to apply `accelerate` in `AutoModel.from_pretrained`."""
    max_memory = {}
    if max_memory_per_gpu is not None:
        max_memory_per_gpu_map = {
            device_idx: max_memory_per_gpu
            for device_idx in range(torch.cuda.device_count())
        }
        max_memory.update(max_memory_per_gpu_map)
    if max_cpu_memory is not None:
        max_memory["cpu"] = max_cpu_memory

    args = {}
    if max_memory:
        args["max_memory"] = max_memory
    args["device_map"] = device_map_option
    args["offload_folder"] = offload_folder
    return args
64
65


66
@register_model("hf-auto", "hf", "huggingface")
67
class HFLM(TemplateLM):
68
69
70
71
72
73
74
    """
    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.
    """

75
    AUTO_MODEL_CLASS = None
76
    _DEFAULT_MAX_LENGTH = 2048
haileyschoelkopf's avatar
haileyschoelkopf committed
77

78
79
    def __init__(
        self,
80
        pretrained: Optional[Union[str, transformers.PreTrainedModel]] = "gpt2",
Baber Abbasi's avatar
Baber Abbasi committed
81
82
        backend: Optional[Literal["default", "causal", "seq2seq"]] = "default",
        # override whether the model should be treated as decoder-only (causal) or encoder-decoder (seq2seq)
83
84
        revision: Optional[str] = "main",
        subfolder: Optional[str] = None,
85
86
87
88
89
90
91
        tokenizer: Optional[
            Union[
                str,
                transformers.PreTrainedTokenizer,
                transformers.PreTrainedTokenizerFast,
            ]
        ] = None,
lintangsutawika's avatar
lintangsutawika committed
92
        truncation: Optional[bool] = False,
Baber Abbasi's avatar
Baber Abbasi committed
93
        logits_cache: bool = True,
94
95
        max_length: Optional[int] = None,
        device: Optional[str] = "cuda",
96
        dtype: Optional[Union[str, torch.dtype]] = "auto",
Benjamin Fattori's avatar
Benjamin Fattori committed
97
98
        batch_size: Optional[Union[int, str]] = 1,
        max_batch_size: Optional[int] = 64,
99
        trust_remote_code: Optional[bool] = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
100
        use_fast_tokenizer: Optional[bool] = True,
101
        add_bos_token: Optional[bool] = False,
102
        # arguments used for splitting a model across GPUs naively.
103
104
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
105
106
107
        device_map_option: Optional[str] = "auto",
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
108
        offload_folder: Optional[Union[str, os.PathLike]] = "./offload",
109
110
        # PEFT and quantization options
        peft: Optional[str] = None,
111
112
        autogptq: Optional[Union[bool, str]] = False,
        **kwargs,
Ethan Smith's avatar
Ethan Smith committed
113
    ) -> None:
114
115
        super().__init__()

116
117
118
119
        # optionally: take in an already-initialized transformers.PreTrainedModel
        if not isinstance(pretrained, str):
            eval_logger.warning(
                "`pretrained` model kwarg is not of type `str`. Many other model arguments may be ignored. Please do not launch via accelerate or use `parallelize=True` if passing an existing model this way."
120
            )
121
            assert not parallelize, "`parallelize=True` is not compatible with passing pre-initialized model to `pretrained`"
122
123
124
            self._model = pretrained
            self._device = self._model.device
            self._config = self._model.config
Baber Abbasi's avatar
Baber Abbasi committed
125
            gpus = 0
126
127
128
129
130
131

            if tokenizer:
                assert isinstance(
                    tokenizer, transformers.PreTrainedTokenizer
                ) or isinstance(tokenizer, transformers.PreTrainedTokenizerFast)
                self.tokenizer = tokenizer
132
            else:
133
134
135
136
137
138
139
                # Get tokenizer
                model_name = self._model.name_or_path
                self.tokenizer = transformers.AutoTokenizer.from_pretrained(
                    model_name,
                    revision=revision,
                    trust_remote_code=trust_remote_code,
                    use_fast=use_fast_tokenizer,
140
                )
141

142
        else:
143
144
145
146
147
            assert isinstance(device, str)
            assert isinstance(pretrained, str)
            assert isinstance(batch_size, (int, str))

            gpus = torch.cuda.device_count()
Jeevan's avatar
Jeevan committed
148
149
            accelerator_kwargs = InitProcessGroupKwargs(timeout=timedelta(weeks=52))
            accelerator = Accelerator(kwargs_handlers=[accelerator_kwargs])
150
151
            if accelerator.num_processes > 1:
                self.accelerator = accelerator
152
153
154
155
156
157
158

            if not (parallelize or accelerator.num_processes > 1):
                # use user-passed device
                device_list = set(
                    ["cuda", "cpu"]
                    + [f"cuda:{i}" for i in range(torch.cuda.device_count())]
                    + ["mps", "mps:0"]
159
                )
160
                if device and device in device_list:
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
                    self._device = torch.device(device)
                    eval_logger.info(f"Using device '{device}'")
                    if device in ("mps", "mps:0") and version.parse(
                        torch.__version__
                    ) < version.parse("2.1"):
                        raise RuntimeError(
                            f"mps requires torch >= 2.1. You have {torch.__version__}"
                        )
                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")
                    )
            else:
                if device != "cuda":
                    eval_logger.info(
                        f"Using `accelerate launch` or `parallelize=True`, device '{device}' will be overridden when placing model."
                    )
                # TODO: include in warning that `load_in_8bit` etc. affect this too
183
                self._device = torch.device(device)
184

185
186
            # TODO: update this to be less of a hack once subfolder is fixed in HF
            revision = revision + ("/" + subfolder if subfolder is not None else "")
187

188
            self._get_config(
189
190
191
192
193
                pretrained,
                revision=revision,
                trust_remote_code=trust_remote_code,
            )

194
195
196
197
        # determine which of 'causal' and 'seq2seq' backends to use
        self._get_backend(
            config=self.config, backend=backend, trust_remote_code=trust_remote_code
        )
198

199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
        # if we passed `pretrained` as a string, initialize our model now
        if isinstance(pretrained, str):
            self._create_model(
                pretrained=pretrained,
                revision=revision,
                dtype=dtype,
                trust_remote_code=trust_remote_code,
                parallelize=parallelize,
                device_map_option=device_map_option,
                max_memory_per_gpu=max_memory_per_gpu,
                max_cpu_memory=max_cpu_memory,
                offload_folder=offload_folder,
                peft=peft,
                autogptq=autogptq,
                **kwargs,
214
215
            )

216
        # access self._model through self.model property outside this method
217
218
219
        if isinstance(self.model, torch.nn.Module):
            self.model.eval()
            self.model.tie_weights()
haileyschoelkopf's avatar
haileyschoelkopf committed
220

221
        if isinstance(pretrained, str) and (gpus >= 1 or str(self.device) == "mps"):
222
223
            # TODO: can remove this whole snippet except in the mps case, perhaps?
            if not (parallelize or autogptq or hasattr(self, "accelerator")):
224
225
226
227
228
229
                # place model onto device requested manually,
                # if not using HF Accelerate or device_map
                # or any other option that preloads model onto device
                try:
                    self.model.to(self.device)
                except ValueError:
230
231
                    eval_logger.debug(
                        "Failed to place model onto specified device. This may be because the model is quantized via `bitsandbytes` or `device_map` is provided. If the desired GPU is being used, this message is safe to ignore."
232
233
234
235
236
                    )

        self._create_tokenizer(
            pretrained,
            tokenizer,
237
            revision=revision,
238
            trust_remote_code=trust_remote_code,
239
            use_fast_tokenizer=use_fast_tokenizer,
240
241
        )

lintangsutawika's avatar
lintangsutawika committed
242
        self.truncation = truncation
Baber Abbasi's avatar
Baber Abbasi committed
243
        self.logits_cache = logits_cache
244
        self.vocab_size = self.tokenizer.vocab_size
245
246
247
248
249
250
251
252
        # select (or create) a pad token to use
        if self.tokenizer.pad_token:
            pass
        elif self.tokenizer.unk_token:
            self.tokenizer.pad_token_id = self.tokenizer.unk_token_id
        elif self.tokenizer.eos_token:
            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
        else:
253
            if getattr(self.config, "model_type", None) == "qwen":
254
255
                # Qwen's trust_remote_code tokenizer does not allow for adding special tokens
                self.tokenizer.pad_token = "<|endoftext|>"
256
257
258
259
260
261
262
263
264
265
            elif (
                self.tokenizer.__class__.__name__ == "RWKVWorldTokenizer"
                or self.tokenizer.__class__.__name__ == "Rwkv5Tokenizer"
            ):
                # The RWKV world tokenizer, does not allow for adding special tokens / setting the pad token (which is set as 0)
                # The additional tokenizer name check is needed, as there exists rwkv4 models with neox tokenizer
                # ---
                # Note that the world tokenizer class name, might change in the future for the final huggingface merge
                # https://github.com/huggingface/transformers/pull/26963
                assert self.tokenizer.pad_token_id == 0
266
267
            else:
                self.tokenizer.add_special_tokens({"pad_token": "<|pad|>"})
268

269
270
        # TODO: override this for Gemma
        self.add_bos_token = add_bos_token
271
272
        if getattr(self.config, "model_type", None) == "gemma":
            self.add_bos_token = True
273
            eval_logger.info(
274
                f"Model type is '{self.config.model_type}', a BOS token will be used as Gemma underperforms without it."
275
276
            )

277
278
        self._max_length = max_length

Benjamin Fattori's avatar
Benjamin Fattori committed
279
280
281
282
283
284
285
286
287
288
        self.batch_schedule = 1
        self.batch_sizes = {}
        self.max_batch_size = max_batch_size

        if str(batch_size).startswith("auto"):
            batch_size = batch_size.split(":")
            self.batch_size_per_gpu = batch_size[0]
            self.batch_schedule = float(batch_size[1]) if len(batch_size) > 1 else 1
        else:
            self.batch_size_per_gpu = int(batch_size)
289

290
291
292
293
294
295
296
297
298
299
300
301
302
303
        if isinstance(pretrained, str):
            # multigpu data-parallel support when launched with accelerate
            if gpus > 1:
                if parallelize:
                    if accelerator.num_processes > 1:
                        raise RuntimeError(
                            "Attempted to use both a HF Accelerate `device_map` and to launch via `accelerate launch`. If this is the case, please either remove `parallelize=True` from --model_args or launch outside of the Accelerate launcher."
                        )
                    else:
                        pass
                elif accelerator.num_processes == 1:
                    # if we aren't launching via accelerate, ditch
                    self._rank = 0
                    self._world_size = 1
304
                else:
305
306
307
308
309
310
311
                    if gpus > accelerator.num_processes:
                        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."
                        )
312
313
314
315
316
317
318
                    assert (
                        accelerator.distributed_type
                        in [
                            DistributedType.FSDP,
                            DistributedType.MULTI_GPU,
                        ]
                    ), "Unsupported distributed type provided. Only DDP and FSDP are supported."
319
320
321
322
323
324
325
326
                    if accelerator.distributed_type == DistributedType.FSDP:
                        self._model = accelerator.prepare(self.model)
                    else:
                        self._model = accelerator.prepare_model(
                            self.model, evaluation_mode=True
                        )
                    self._device = torch.device(
                        f"cuda:{accelerator.local_process_index}"
327
                    )
328
                    self.accelerator = accelerator
329

330
331
                    if self.accelerator.is_local_main_process:
                        eval_logger.info(f"Using {gpus} devices with data parallelism")
332

333
334
335
336
337
338
339
340
341
                    self._rank = self.accelerator.local_process_index
                    self._world_size = self.accelerator.num_processes
        else:
            # if a PreTrainedModel was passed into HFLM, we forgo distributed setup.
            eval_logger.warning(
                "Passed an already-initialized model through `pretrained`, assuming single-process call to evaluate() or custom distributed integration"
            )
            self._rank = 0
            self._world_size = 1
haileyschoelkopf's avatar
haileyschoelkopf committed
342

343
344
345
346
347
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

348
349
350
351
352
353
354
355
    @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

356
357
358
359
360
361
362
    @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):
363
364
365
366
367
368
369
370
371
372
373
        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
374

375
    @property
Ethan Smith's avatar
Ethan Smith committed
376
    def max_gen_toks(self) -> int:
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
        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

395
396
    def _get_backend(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
397
        config: Union[transformers.PretrainedConfig, transformers.AutoConfig],
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
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
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
        backend: Optional[Literal["default", "causal", "seq2seq"]] = "default",
        trust_remote_code: Optional[bool] = False,
    ) -> None:
        """
        Helper method during initialization.
        Determines the backend ("causal" (decoder-only) or "seq2seq" (encoder-decoder))
        model type to be used.
        """
        assert backend in ["default", "causal", "seq2seq"]

        if backend != "default":
            # if we've settled on non-default backend, use that manually
            if backend == "causal":
                self.AUTO_MODEL_CLASS = transformers.AutoModelForCausalLM
            elif backend == "seq2seq":
                self.AUTO_MODEL_CLASS = transformers.AutoModelForSeq2SeqLM
            eval_logger.info(
                f"Overrode HF model backend type, and using type '{backend}'"
            )
        else:
            # determine and use the default HF backend for this model, based on its config + metadata.
            if (
                getattr(config, "model_type")
                in MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES
            ):
                # first check if model type is listed under seq2seq models, since some
                # models like MBart are listed in both seq2seq and causal mistakenly in HF transformers.
                # these special cases should be treated as seq2seq models.
                self.AUTO_MODEL_CLASS = transformers.AutoModelForSeq2SeqLM
            elif (
                getattr(self.config, "model_type") in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
            ):
                self.AUTO_MODEL_CLASS = transformers.AutoModelForCausalLM
            else:
                if not trust_remote_code:
                    eval_logger.warning(
                        "HF model type is neither marked as CausalLM or Seq2SeqLM. \
                    This is expected if your model requires `trust_remote_code=True` but may be an error otherwise."
                    )
                # if model type is neither in HF transformers causal or seq2seq model registries
                # then we default to AutoModelForCausalLM
                self.AUTO_MODEL_CLASS = transformers.AutoModelForCausalLM

        assert self.AUTO_MODEL_CLASS in [
            transformers.AutoModelForCausalLM,
            transformers.AutoModelForSeq2SeqLM,
        ]
        return None

    def _get_config(
        self,
        pretrained: str,
        revision: str = "main",
        trust_remote_code: bool = False,
    ) -> None:
        self._config = transformers.AutoConfig.from_pretrained(
            pretrained,
            revision=revision,
            trust_remote_code=trust_remote_code,
        )

    def _create_model(
        self,
        pretrained: str,
        revision: Optional[str] = "main",
        dtype: Optional[Union[str, torch.dtype]] = "auto",
        trust_remote_code: Optional[bool] = False,
        # arguments used for splitting a model across GPUs naively.
        # only used if `parallelize=True`.
        # (accelerate naive PP (device_map) options)
        parallelize: Optional[bool] = False,
        device_map_option: Optional[str] = "auto",
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
        offload_folder: Optional[str] = "./offload",
        # PEFT and quantization options
        peft: Optional[str] = None,
        autogptq: Optional[Union[bool, str]] = False,
        **kwargs,
    ) -> None:
        """
        Initializes an HF or HF-compatible PreTrainedModel from scratch
        inside HFLM, using the kwargs passed into self.__init__().

        Also handles functionality such as AutoGPTQ usage and PEFT wrapping.

        For future similar extensions to AutoGPTQ that are not core to HF's ecosystem,
        (such as PyTorch models that are nearly, but not quite, fully mirroring
        HF's public interface relied on in this HFLM class)
        please consider subclassing HFLM and overriding this and other methods as needed.
        """

        model_kwargs = kwargs if kwargs else {}

        if parallelize:
            model_kwargs.update(
                _get_accelerate_args(
495
                    device_map_option,  # TODO: phase out device_map_option?
496
497
498
499
500
                    max_memory_per_gpu,
                    max_cpu_memory,
                    offload_folder,
                )
            )
501
502
503
504
505
506
507
508
509
510
511
512
        elif "device_map" not in model_kwargs:
            # set a device_map to initialize model on the right GPU.
            # this is needed because it seems that the default behavior
            # for quantized models now seems to be device_map="auto"
            # which breaks data-parallel mode.
            if hasattr(self, "accelerator"):
                model_kwargs.update(
                    {"device_map": {"": f"cuda:{self.accelerator.local_process_index}"}}
                )
            else:
                model_kwargs.update({"device_map": {"": str(self.device)}})

513
514
515
516
517
518
519
520
        if not autogptq:
            if model_kwargs.get("load_in_4bit", None):
                assert (
                    transformers.__version__ >= "4.30.0"
                ), "load_in_4bit requires transformers >= 4.30.0"
            if transformers.__version__ >= "4.30.0":
                if model_kwargs.get("load_in_4bit", None):
                    if model_kwargs.get("bnb_4bit_compute_dtype", None):
521
                        model_kwargs["bnb_4bit_compute_dtype"] = get_dtype(
522
523
524
525
526
                            model_kwargs["bnb_4bit_compute_dtype"]
                        )
            self._model = self.AUTO_MODEL_CLASS.from_pretrained(
                pretrained,
                revision=revision,
527
                torch_dtype=get_dtype(dtype),
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
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
605
606
607
                trust_remote_code=trust_remote_code,
                **model_kwargs,
            )
        else:
            try:
                from auto_gptq import AutoGPTQForCausalLM
            except ModuleNotFoundError:
                raise Exception(
                    "Tried to load auto_gptq, but auto-gptq is not installed ",
                    "please install auto-gptq via pip install lm-eval[gptq] or pip install -e .[gptq]",
                )

            self._model = AutoGPTQForCausalLM.from_quantized(
                pretrained,
                trust_remote_code=trust_remote_code,
                model_basename=None if autogptq is True else Path(autogptq).stem,
                use_safetensors=True
                if autogptq is True
                else autogptq.endswith(".safetensors"),
                **model_kwargs,
            )

        if peft:
            if model_kwargs.get("load_in_4bit", None):
                assert PEFT_VERSION >= "0.4.0", "load_in_4bit requires peft >= 0.4.0"
            self._model = PeftModel.from_pretrained(
                self._model, peft, revision=revision
            )

        return None

    def _create_tokenizer(
        self,
        pretrained: Union[str, transformers.PreTrainedModel],
        tokenizer: Optional[
            Union[
                str,
                transformers.PreTrainedTokenizer,
                transformers.PreTrainedTokenizerFast,
            ]
        ],
        revision: Optional[str] = "main",
        trust_remote_code: Optional[bool] = False,
        use_fast_tokenizer: Optional[bool] = True,
    ) -> None:
        """
        Helper method during initialization.

        Create a tokenizer object corresponding to the correct
        tokenizer for value of `pretrained`, or use the pre-initialized tokenizer passed.
        """

        if tokenizer:
            if isinstance(tokenizer, str):
                self.tokenizer = transformers.AutoTokenizer.from_pretrained(
                    tokenizer,
                    revision=revision,
                    trust_remote_code=trust_remote_code,
                    use_fast=use_fast_tokenizer,
                )
            else:
                assert isinstance(
                    tokenizer, transformers.PreTrainedTokenizer
                ) or isinstance(tokenizer, transformers.PreTrainedTokenizerFast)
                self.tokenizer = tokenizer
        else:
            # Get tokenizer based on 'pretrained'
            if isinstance(pretrained, str):
                model_name = pretrained
            else:
                # get the HF hub name via accessor on model
                model_name = self.model.name_or_path
            self.tokenizer = transformers.AutoTokenizer.from_pretrained(
                model_name,
                revision=revision,
                trust_remote_code=trust_remote_code,
                use_fast=use_fast_tokenizer,
            )
        return None

Ethan Smith's avatar
Ethan Smith committed
608
    def _detect_batch_size(self, requests=None, pos: int = 0):
Benjamin Fattori's avatar
Benjamin Fattori committed
609
610
611
612
613
        if requests:
            _, context_enc, continuation_enc = requests[pos]
            max_length = len(
                (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1]
            )
614
615
            max_context_enc = len(context_enc[-(self.max_length + 1) :])
            max_cont_enc = len(continuation_enc[-(self.max_length + 1) :])
Benjamin Fattori's avatar
Benjamin Fattori committed
616
617
        else:
            max_length = self.max_length
lintangsutawika's avatar
lintangsutawika committed
618

Benjamin Fattori's avatar
Benjamin Fattori committed
619
620
621
        # if OOM, then halves batch_size and tries again
        @find_executable_batch_size(starting_batch_size=self.max_batch_size)
        def forward_batch(batch_size):
622
623
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                length = max(max_context_enc, max_cont_enc)
lintangsutawika's avatar
lintangsutawika committed
624
625
626
                batched_conts = torch.ones(
                    (batch_size, length), device=self.device
                ).long()
627
628
                test_batch = torch.ones((batch_size, length), device=self.device).long()
                call_kwargs = {
lintangsutawika's avatar
lintangsutawika committed
629
630
631
                    "attn_mask": test_batch,
                    "labels": batched_conts,
                }
632
633
            else:
                call_kwargs = {}
lintangsutawika's avatar
lintangsutawika committed
634
635
636
                test_batch = torch.ones(
                    (batch_size, max_length), device=self.device
                ).long()
Benjamin Fattori's avatar
Benjamin Fattori committed
637
            for _ in range(5):
638
                out = F.log_softmax(self._model_call(test_batch, **call_kwargs), dim=-1)  # noqa: F841
lintangsutawika's avatar
lintangsutawika committed
639

Benjamin Fattori's avatar
Benjamin Fattori committed
640
641
            return batch_size

642
643
644
645
646
647
648
        try:
            batch_size = forward_batch()
        except RuntimeError as e:
            if "No executable batch size found" in str(e):
                batch_size = 1
            else:
                raise
Benjamin Fattori's avatar
Benjamin Fattori committed
649

650
651
652
653
654
655
656
        if self.world_size > 1:
            # if multi-GPU, always take minimum over all selected batch sizes
            max_rnk_bs = torch.tensor([batch_size], device=self.device)
            gathered = (
                self.accelerator.gather(max_rnk_bs).cpu().detach().numpy().tolist()
            )
            batch_size = min(gathered)
657
            clear_torch_cache()
658
659
            return batch_size

660
        clear_torch_cache()
Benjamin Fattori's avatar
Benjamin Fattori committed
661
662
        return batch_size

baberabb's avatar
baberabb committed
663
664
665
    def tok_encode(
        self, string: str, left_truncate_len=None, add_special_tokens=None
    ) -> List[int]:
haileyschoelkopf's avatar
haileyschoelkopf committed
666
        """ """
667
668
        if add_special_tokens is None:
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
669
                add_special_tokens = False or self.add_bos_token
670
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
671
                # TODO: investigate best practices for enc-dec models + special tokens
672
                add_special_tokens = True
673
674

        encoding = self.tokenizer.encode(string, add_special_tokens=add_special_tokens)
haileyschoelkopf's avatar
haileyschoelkopf committed
675

676
677
678
        # 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
679

680
681
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
682
    def tok_batch_encode(
lintangsutawika's avatar
lintangsutawika committed
683
684
        self,
        strings: List[str],
lintangsutawika's avatar
lintangsutawika committed
685
        padding_side: str = "left",
686
687
        left_truncate_len: int = None,
        truncation: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
688
    ) -> Tuple[torch.Tensor, torch.Tensor]:
haileyschoelkopf's avatar
haileyschoelkopf committed
689
690
691
692
693
        # 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:
694
            add_special_tokens = False or self.add_bos_token
haileyschoelkopf's avatar
haileyschoelkopf committed
695
696
697
698
699
        elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
            add_special_tokens = True

        encoding = self.tokenizer(
            strings,
lintangsutawika's avatar
lintangsutawika committed
700
            truncation=truncation,
haileyschoelkopf's avatar
haileyschoelkopf committed
701
702
703
704
705
706
707
708
709
710
711
712
713
            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"]

714
715
716
717
718
719
720
721
    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
722
        :param inps: torch.Tensor
723
724
725
726
727
728
729
730
731
732
733
734
735
            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():
736
737
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
738
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
haileyschoelkopf's avatar
haileyschoelkopf committed
739
740
741
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
742
743
744
745
746
            else:
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                return self.model(inps).logits

    def _model_generate(self, context, max_length, stop, **generation_kwargs):
Baber Abbasi's avatar
Baber Abbasi committed
747
        # temperature = 0.0 if not set
748
749
750
        # if do_sample is false and temp==0.0:
        # remove temperature, as do_sample=False takes care of this
        # and we don't want a warning from HF
Baber Abbasi's avatar
Baber Abbasi committed
751
        generation_kwargs["temperature"] = generation_kwargs.get("temperature", 0.0)
752
        do_sample = generation_kwargs.get("do_sample", None)
753
754
755
756
757

        # The temperature has to be a strictly positive float -- if it is 0.0, use greedy decoding strategies
        if generation_kwargs.get("temperature") == 0.0 and do_sample is None:
            generation_kwargs["do_sample"] = do_sample = False

Baber Abbasi's avatar
Baber Abbasi committed
758
759
        if do_sample is False and generation_kwargs.get("temperature") == 0.0:
            generation_kwargs.pop("temperature")
760
761
        # build stopping criteria
        stopping_criteria = stop_sequences_criteria(
762
            self.tokenizer, stop, context.shape[1], context.shape[0]
763
        )
764
        return self.model.generate(
765
            input_ids=context,
766
767
            max_length=max_length,
            stopping_criteria=stopping_criteria,
768
            pad_token_id=self.tokenizer.pad_token_id,
769
770
771
            use_cache=True,
            **generation_kwargs,
        )
772

Baber Abbasi's avatar
Baber Abbasi committed
773
774
775
    def _select_cont_toks(
        self, logits: torch.Tensor, contlen: int = None, inplen: int = None
    ) -> torch.Tensor:
776
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
777
778
779
            assert (
                contlen and inplen
            ), "Must pass input len and cont. len to select scored logits for causal LM"
780
781
782
783
            # 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
784
785
786
787
            assert (
                contlen and not inplen
            ), "Selecting scored logits for Seq2SeqLM requires only cont. len"
            # only discard right-padding.
788
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
789
790
            logits = logits[:contlen]

791
792
        return logits

baberabb's avatar
baberabb committed
793
    def loglikelihood_rolling(self, requests: List[Instance]) -> List[float]:
794
        loglikelihoods = []
Benjamin Fattori's avatar
Benjamin Fattori committed
795
796
797
798
799
800
801
802
803

        adaptive_batch_size = None
        if self.batch_size == "auto":
            # using rolling window with maximum context
            print("Passed argument batch_size = auto. Detecting largest batch size")
            batch_size = self._detect_batch_size()
            print(f"Determined Largest batch size: {batch_size}")
            adaptive_batch_size = batch_size

804
805
806
807
808
809
        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
810
                        prefix_token=self.eot_token_id,
811
812
813
814
815
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
816
817

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
            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(
Baber Abbasi's avatar
Baber Abbasi committed
833
                requests=rolling_token_windows,
lintangsutawika's avatar
lintangsutawika committed
834
835
                disable_tqdm=True,
                override_bs=adaptive_batch_size,
836
837
838
839
840
841
842
843
844
845
846
847
            )

            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
Zhiwei Zhuang's avatar
Zhiwei Zhuang committed
848

849
850
851
852
853
854
855
856
857
858
859
860
861
    def _batch_scheduler(self, pos, n_reordered_requests):
        sched = pos // int(len(n_reordered_requests) / self.batch_schedule)
        if sched in self.batch_sizes:
            return self.batch_sizes[sched]
        if (len(self.batch_sizes) > 1) and (
            self.batch_sizes[sched - 1] == self.max_batch_size
        ):
            # if previous batch size is already maximal, skip recomputation
            self.batch_sizes[sched] = self.max_batch_size
            return self.batch_sizes[sched]
        print(
            f"Passed argument batch_size = auto:{self.batch_schedule}. Detecting largest batch size"
        )
Zhiwei Zhuang's avatar
Zhiwei Zhuang committed
862
        self.batch_sizes[sched] = self._detect_batch_size(n_reordered_requests, pos)
863
864
        print(f"Determined largest batch size: {self.batch_sizes[sched]}")
        return self.batch_sizes[sched]
865

Ethan Smith's avatar
Ethan Smith committed
866
    def _loglikelihood_tokens(
baberabb's avatar
baberabb committed
867
868
869
870
871
        self,
        requests: List[Tuple[Tuple[str, str], List[int], List[int]]],
        disable_tqdm: bool = False,
        override_bs: int = None,
    ) -> List[Tuple[float, bool]]:
872
873
874
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

Baber Abbasi's avatar
Baber Abbasi committed
875
        def _collate(req: Tuple[Tuple[str, str], List[int], List[int]]):
Baber Abbasi's avatar
Baber Abbasi committed
876
            """Defines the key for the sorted method"""
877
878
879
880
881
882
883
            # 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

Baber Abbasi's avatar
Baber Abbasi committed
884
            toks = req[1] + req[2]
885
886
            return -len(toks), tuple(toks)

Baber Abbasi's avatar
Baber Abbasi committed
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
        def _lookup_one_token_cont(req: Tuple[Tuple[str, str], List[int], List[int]]):
            """Defines the key to group and lookup one-token continuations"""
            # Use with group_by="contexts" (optional)"
            # allows for the creation of a lookup, so we can re-use logits in case of one-token continuations.
            # speeds up some multiple-choice tasks proportionally to the number of choices.
            # groups requests by context+continuation[:-1] and infer on one request/group.
            return req[-2] + req[-1][:-1]

        re_ord = Collator(
            requests,
            sort_fn=_collate,
            group_by="contexts"
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
            and self.logits_cache
            else None,
            group_fn=_lookup_one_token_cont,
        )
Benjamin Fattori's avatar
Benjamin Fattori committed
904
905
906

        # automatic (variable) batch size detection for vectorization
        # pull longest context sample from request
Baber Abbasi's avatar
Baber Abbasi committed
907
908
909
        n_reordered_requests = len(re_ord)
        batch_size = (
            self.batch_size
910
911
912
            if self.batch_size != "auto"
            else override_bs
            if override_bs is not None
Baber Abbasi's avatar
Baber Abbasi committed
913
914
915
916
            else 0
        )
        batch_fn = (
            self._batch_scheduler
917
918
919
            if self.batch_size == "auto"
            and n_reordered_requests > 0
            and not override_bs
Baber Abbasi's avatar
Baber Abbasi committed
920
            else None
921
922
        )

Baber Abbasi's avatar
Baber Abbasi committed
923
        chunks = re_ord.get_batched(n=batch_size, batch_fn=batch_fn)
924
925
926
927
928
        pbar = tqdm(
            total=len(requests),
            disable=(disable_tqdm or (self.rank != 0)),
            desc="Running loglikelihood requests",
        )
haileyschoelkopf's avatar
haileyschoelkopf committed
929
        for chunk in chunks:
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
            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
949
                # how this all works (illustrated on a causal decoder-only setup):
950
951
952
953
954
955
956
957
958
959
960
                #          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,
961
962
                        device=self.device,
                    )
963
964
965
966
967
                    (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
968
                        device=self.device,
969
                    )
970
                    (inplen,) = inp.shape
971
972
973
974

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

975
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
976
                        (continuation_enc)[-self.max_length :],
977
978
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
979
                        dtype=torch.long,
980
981
                        device=self.device,
                    )
982
983
                    (contlen,) = cont.shape

984
985
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
986
987
988
989
990
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
991

haileyschoelkopf's avatar
haileyschoelkopf committed
992
993
994
995
996
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
997
998
999
1000

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
1001

1002
1003
1004
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
1005
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1006
1007
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
1008
1009
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                # TODO: left-pad encoder inps and mask?
1010
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1011
1012
                    padding_len_inp, inps
                )  # [batch, padding_len_inp]
1013
                batched_conts = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1014
1015
                    padding_len_cont, conts
                )  # [batch, padding_len_cont]
1016
                batched_encoder_mask = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1017
1018
1019
1020
1021
1022
                    padding_len_inp, encoder_attns
                )  # [batch, padding_len_inp]
                call_kwargs = {
                    "attn_mask": batched_encoder_mask,
                    "labels": batched_conts,
                }
1023
1024
1025

            multi_logits = F.log_softmax(
                self._model_call(batched_inps, **call_kwargs), dim=-1
1026
            )  # [batch, padding_length (inp or cont), vocab]
1027

Baber Abbasi's avatar
Baber Abbasi committed
1028
            for (request_str, ctx_tokens, _), logits, inplen, cont_toks in zip(
1029
1030
1031
1032
                chunk, multi_logits, inplens, cont_toks_list
            ):
                # Slice to original seq length
                contlen = len(cont_toks)
haileyschoelkopf's avatar
haileyschoelkopf committed
1033
                # take only logits in the continuation
1034
                # (discard context toks if decoder-only ; discard right-padding)
1035
1036
                # also discards + checks for "virtual tokens" in the causal LM's input window
                # from prompt/prefix tuning tokens, if applicable
haileyschoelkopf's avatar
haileyschoelkopf committed
1037
                ctx_len = (
1038
                    inplen + (logits.shape[0] - padding_len_inp)
haileyschoelkopf's avatar
haileyschoelkopf committed
1039
1040
1041
                    if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                    else None
                )
1042
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
1043
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
1044
1045
1046
1047

                # Check if per-token argmax is exactly equal to continuation
                greedy_tokens = logits.argmax(dim=-1)

Baber Abbasi's avatar
Baber Abbasi committed
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
                # check for one-token continuation cache hits.
                # noop in case group_by != "contexts" or no cache hit and returns the
                # original args. Otherwise, expands the logits batch dimension and yields each
                # batch along with matching continuation tokens and prompt strings.
                # logits -> [1, seq, vocab]
                for request_str, cont_toks, logits in re_ord.get_cache(
                    req_str=request_str,
                    cxt_toks=ctx_tokens,
                    cont_toks=cont_toks,
                    logits=logits,
                ):
                    cont_toks = torch.tensor(
                        cont_toks, dtype=torch.long, device=self.device
                    ).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)

                    self.cache_hook.add_partial("loglikelihood", request_str, answer)
                    pbar.update(1)
haileyschoelkopf's avatar
haileyschoelkopf committed
1077
1078

        pbar.close()
haileyschoelkopf's avatar
haileyschoelkopf committed
1079

1080
1081
        return re_ord.get_original(res)

baberabb's avatar
baberabb committed
1082
    def generate_until(self, requests: List[Instance]) -> List[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1083
        res = []
1084

Baber Abbasi's avatar
Baber Abbasi committed
1085
        def _collate(req: Tuple[str, dict]):
Baber Abbasi's avatar
Baber Abbasi committed
1086
            """Defines the key for the sorted method"""
1087
1088
1089
1090
1091
1092
            # 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
Baber Abbasi's avatar
Baber Abbasi committed
1093
1094
            toks = self.tok_encode(req[0])
            return -len(toks), req[0]
1095

1096
1097
1098
1099
1100
        pbar = tqdm(
            total=len(requests),
            disable=(self.rank != 0),
            desc="Running generate_until requests",
        )
Baber Abbasi's avatar
Baber Abbasi committed
1101
        adaptive_batch_size = None
1102
1103
1104
1105
1106
1107
        if self.batch_size == "auto":
            # using rolling window with maximum context
            print("Passed argument batch_size = auto. Detecting largest batch size")
            batch_size = self._detect_batch_size()
            print(f"Determined Largest batch size: {batch_size}")
            adaptive_batch_size = batch_size
1108
        # for each different set of kwargs, we execute all requests, by batch.
Baber Abbasi's avatar
Baber Abbasi committed
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
        batch_size = (
            self.batch_size
            if self.batch_size != "auto"
            else adaptive_batch_size
            if adaptive_batch_size is not None
            else 0
        )
        batch_fn = (
            self._batch_scheduler
            if self.batch_size == "auto" and not adaptive_batch_size
            else None
        )
1121

Baber Abbasi's avatar
Baber Abbasi committed
1122
1123
1124
        # we group requests by their generation_kwargs,
        # so that we don't try to execute e.g. greedy sampling and temp=0.8 sampling
        # in the same batch.
Baber Abbasi's avatar
Baber Abbasi committed
1125
1126
1127
1128
1129
1130
1131
        # group_fn=lambda x: x[1] -> x=(context, gen_kwargs)
        re_ords = Collator(
            [reg.args for reg in requests],
            sort_fn=_collate,
            group_by="gen_kwargs",
            group_fn=lambda x: x[1],
        )
Baber Abbasi's avatar
Baber Abbasi committed
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
        chunks = re_ords.get_batched(n=batch_size, batch_fn=batch_fn)
        for chunk in chunks:
            contexts, all_gen_kwargs = zip(*chunk)
            # we assume all gen kwargs in the batch are the same
            # this is safe to assume because the `grouper` object ensures it.
            gen_kwargs = all_gen_kwargs[0]
            # unpack our keyword arguments.
            until = None
            if isinstance(gen_kwargs, dict):
                kwargs = copy.deepcopy(gen_kwargs)  # edge case for repeats > 1
                if "until" in kwargs.keys():
                    until = kwargs.pop("until")
                    if isinstance(until, str):
                        until = [kwargs]
                    elif not isinstance(until, list):
                        raise ValueError(
                            f"Expected `kwargs['until']` to be of type Union[str,list] but got {until}"
                        )
            else:
                raise ValueError(
Baber Abbasi's avatar
Baber Abbasi committed
1152
                    f"Expected `kwargs` to be of type `dict` but got {type(gen_kwargs)}"
1153
                )
Baber Abbasi's avatar
Baber Abbasi committed
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
            if not until:
                until = [self.tok_decode(self.eot_token_id)]
            if "max_gen_toks" in kwargs.keys():
                max_gen_toks = kwargs.pop("max_gen_toks")
            else:
                max_gen_toks = self.max_gen_toks

            # 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

            # encode, pad, and truncate contexts for this batch
            context_enc, attn_masks = self.tok_batch_encode(
                contexts,
                left_truncate_len=max_ctx_len,
                truncation=self.truncation,
            )
            context_enc = context_enc.to(self.device)
            attn_masks = attn_masks.to(self.device)
1177

Baber Abbasi's avatar
Baber Abbasi committed
1178
1179
            if "max_length" not in kwargs:
                kwargs["max_length"] = context_enc.shape[1] + max_gen_toks
1180

Baber Abbasi's avatar
Baber Abbasi committed
1181
1182
1183
1184
1185
1186
1187
            # perform batched generation
            cont = self._model_generate(
                context=context_enc,
                attention_mask=attn_masks,
                stop=until,
                **kwargs,
            )
1188

Baber Abbasi's avatar
Baber Abbasi committed
1189
1190
1191
1192
1193
            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] :]
1194

Baber Abbasi's avatar
Baber Abbasi committed
1195
                s = self.tok_decode(cont_toks)
1196

Baber Abbasi's avatar
Baber Abbasi committed
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
                # 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 self.tok_decode(self.eot_token_id) = ''
                        s = s.split(term)[0]

                res.append(s)

                self.cache_hook.add_partial("generate_until", (context, gen_kwargs), s)
                pbar.update(1)
        # reorder this group of results back to original unsorted form
        res = re_ords.get_original(res)
1210

1211
        pbar.close()
1212

Baber Abbasi's avatar
Baber Abbasi committed
1213
        return res