huggingface.py 55 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
from huggingface_hub import HfApi
17
18
19
20
from packaging import version
from peft import PeftModel
from peft import __version__ as PEFT_VERSION
from tqdm import tqdm
21
22
23
24
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
)
25
26

from lm_eval import utils
baberabb's avatar
baberabb committed
27
from lm_eval.api.instance import Instance
28
from lm_eval.api.model import TemplateLM
29
from lm_eval.api.registry import register_model
30
31
32
33
34
35
36
from lm_eval.models.utils import (
    Collator,
    clear_torch_cache,
    get_dtype,
    pad_and_concat,
    stop_sequences_criteria,
)
37

38

39
eval_logger = utils.eval_logger
40

lintangsutawika's avatar
lintangsutawika committed
41

42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
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
65
66


67
@register_model("hf-auto", "hf", "huggingface")
68
class HFLM(TemplateLM):
69
70
71
72
73
74
75
    """
    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.
    """

76
    AUTO_MODEL_CLASS = None
77
    _DEFAULT_MAX_LENGTH = 2048
haileyschoelkopf's avatar
haileyschoelkopf committed
78

79
80
    def __init__(
        self,
81
        pretrained: Union[str, transformers.PreTrainedModel],
Baber Abbasi's avatar
Baber Abbasi committed
82
83
        backend: Optional[Literal["default", "causal", "seq2seq"]] = "default",
        # override whether the model should be treated as decoder-only (causal) or encoder-decoder (seq2seq)
84
85
        revision: Optional[str] = "main",
        subfolder: Optional[str] = None,
86
87
88
89
90
91
92
        tokenizer: Optional[
            Union[
                str,
                transformers.PreTrainedTokenizer,
                transformers.PreTrainedTokenizerFast,
            ]
        ] = None,
lintangsutawika's avatar
lintangsutawika committed
93
        truncation: Optional[bool] = False,
Baber Abbasi's avatar
Baber Abbasi committed
94
        logits_cache: bool = True,
95
96
        max_length: Optional[int] = None,
        device: Optional[str] = "cuda",
97
        dtype: Optional[Union[str, torch.dtype]] = "auto",
Benjamin Fattori's avatar
Benjamin Fattori committed
98
99
        batch_size: Optional[Union[int, str]] = 1,
        max_batch_size: Optional[int] = 64,
100
        trust_remote_code: Optional[bool] = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
101
        use_fast_tokenizer: Optional[bool] = True,
102
        add_bos_token: Optional[bool] = False,
103
        prefix_token_id: Optional[int] = None,
104
        # arguments used for splitting a model across GPUs naively.
105
106
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
107
108
109
        device_map_option: Optional[str] = "auto",
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
110
        offload_folder: Optional[Union[str, os.PathLike]] = "./offload",
111
        # PEFT, delta weights and quantization options
112
        peft: Optional[str] = None,
113
        delta: Optional[str] = None,
114
115
        autogptq: Optional[Union[bool, str]] = False,
        **kwargs,
Ethan Smith's avatar
Ethan Smith committed
116
    ) -> None:
117
118
        super().__init__()

119
120
121
122
        # 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."
123
            )
124
            assert not parallelize, "`parallelize=True` is not compatible with passing pre-initialized model to `pretrained`"
125
126
127
            self._model = pretrained
            self._device = self._model.device
            self._config = self._model.config
Baber Abbasi's avatar
Baber Abbasi committed
128
            gpus = 0
129
130
131
132
133
134

            if tokenizer:
                assert isinstance(
                    tokenizer, transformers.PreTrainedTokenizer
                ) or isinstance(tokenizer, transformers.PreTrainedTokenizerFast)
                self.tokenizer = tokenizer
135
            else:
136
137
138
139
140
141
142
                # 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,
143
                )
144

145
        else:
146
147
148
149
150
            assert isinstance(device, str)
            assert isinstance(pretrained, str)
            assert isinstance(batch_size, (int, str))

            gpus = torch.cuda.device_count()
Jeevan's avatar
Jeevan committed
151
152
            accelerator_kwargs = InitProcessGroupKwargs(timeout=timedelta(weeks=52))
            accelerator = Accelerator(kwargs_handlers=[accelerator_kwargs])
153
154
            if accelerator.num_processes > 1:
                self.accelerator = accelerator
155
156
157
158
159
160
161

            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"]
162
                )
163
                if device and device in device_list:
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
                    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
186
                self._device = torch.device(device)
187

188
189
            # TODO: update this to be less of a hack once subfolder is fixed in HF
            revision = revision + ("/" + subfolder if subfolder is not None else "")
190

191
            self._get_config(
192
193
194
195
196
                pretrained,
                revision=revision,
                trust_remote_code=trust_remote_code,
            )

197
198
199
200
        # determine which of 'causal' and 'seq2seq' backends to use
        self._get_backend(
            config=self.config, backend=backend, trust_remote_code=trust_remote_code
        )
201

202
203
204
205
206
207
208
209
210
        # load tokenizer so we know tokenizer vocabulary size before loading model and PEFT
        self._create_tokenizer(
            pretrained,
            tokenizer,
            revision=revision,
            trust_remote_code=trust_remote_code,
            use_fast_tokenizer=use_fast_tokenizer,
        )

211
212
213
214
215
216
217
218
219
220
221
222
223
        # 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,
224
                delta=delta,
225
226
                autogptq=autogptq,
                **kwargs,
227
228
            )

229
        # access self._model through self.model property outside this method
230
231
232
        if isinstance(self.model, torch.nn.Module):
            self.model.eval()
            self.model.tie_weights()
haileyschoelkopf's avatar
haileyschoelkopf committed
233

234
        if isinstance(pretrained, str) and (gpus >= 1 or str(self.device) == "mps"):
235
236
            # TODO: can remove this whole snippet except in the mps case, perhaps?
            if not (parallelize or autogptq or hasattr(self, "accelerator")):
237
238
239
240
241
242
                # 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:
243
244
                    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."
245
246
                    )

lintangsutawika's avatar
lintangsutawika committed
247
        self.truncation = truncation
Baber Abbasi's avatar
Baber Abbasi committed
248
        self.logits_cache = logits_cache
249
        self.vocab_size = self.tokenizer.vocab_size
250
251
252
253
254
255
256
257
        # 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:
258
            if getattr(self.config, "model_type", None) == "qwen":
259
260
                # Qwen's trust_remote_code tokenizer does not allow for adding special tokens
                self.tokenizer.pad_token = "<|endoftext|>"
261
262
263
264
265
266
267
268
269
270
            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
271
272
            else:
                self.tokenizer.add_special_tokens({"pad_token": "<|pad|>"})
273

274
275
        # TODO: override this for Gemma
        self.add_bos_token = add_bos_token
276
277
        if getattr(self.config, "model_type", None) == "gemma":
            self.add_bos_token = True
278
            eval_logger.info(
279
                f"Model type is '{self.config.model_type}', a BOS token will be used as Gemma underperforms without it."
280
281
            )

282
        self._max_length = max_length
283
284
285
286
        self.pretrained = pretrained
        self.delta = delta
        self.peft = peft
        self.revision = revision
Benjamin Fattori's avatar
Benjamin Fattori committed
287
288
289
290
291
292
293
294
295
296
        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)
297

298
299
300
301
302
303
304
305
306
307
308
309
310
311
        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
312
                else:
313
314
315
316
317
318
319
                    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."
                        )
320
321
322
323
324
325
326
                    assert (
                        accelerator.distributed_type
                        in [
                            DistributedType.FSDP,
                            DistributedType.MULTI_GPU,
                        ]
                    ), "Unsupported distributed type provided. Only DDP and FSDP are supported."
327
328
329
330
331
332
333
334
                    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}"
335
                    )
336
                    self.accelerator = accelerator
337

338
339
                    if self.accelerator.is_local_main_process:
                        eval_logger.info(f"Using {gpus} devices with data parallelism")
340

341
342
343
344
345
346
347
348
349
                    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
350

351
        self.custom_prefix_token_id = prefix_token_id
352
353
354
355
        if prefix_token_id is not None:
            eval_logger.info(
                f"Loglikelihood prefix token id used in evaluation: {self.prefix_token_id}"
            )
356

357
358
359
360
361
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

362
363
364
365
366
367
368
369
    @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

370
371
372
373
374
    @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

375
376
377
378
379
380
381
382
383
    @property
    def prefix_token_id(self):
        # it is used as prefix for loglikelihood
        if self.custom_prefix_token_id is not None:
            return self.custom_prefix_token_id
        if self.tokenizer.bos_token_id is not None:
            return self.tokenizer.bos_token_id
        return self.tokenizer.eos_token_id

384
385
    @property
    def max_length(self):
386
387
388
389
390
391
392
393
394
395
396
        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
397

398
    @property
Ethan Smith's avatar
Ethan Smith committed
399
    def max_gen_toks(self) -> int:
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
        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

418
419
    def _get_backend(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
420
        config: Union[transformers.PretrainedConfig, transformers.AutoConfig],
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
495
        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",
496
        # PEFT, delta weights and quantization options
497
        peft: Optional[str] = None,
498
        delta: Optional[str] = None,
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
        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(
519
                    device_map_option,  # TODO: phase out device_map_option?
520
521
522
523
524
                    max_memory_per_gpu,
                    max_cpu_memory,
                    offload_folder,
                )
            )
525
526
527
528
529
530
531
532
533
534
535
536
        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)}})

537
538
539
540
541
542
543
544
        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):
545
                        model_kwargs["bnb_4bit_compute_dtype"] = get_dtype(
546
547
548
549
550
                            model_kwargs["bnb_4bit_compute_dtype"]
                        )
            self._model = self.AUTO_MODEL_CLASS.from_pretrained(
                pretrained,
                revision=revision,
551
                torch_dtype=get_dtype(dtype),
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
                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,
            )

574
575
576
577
578
        if peft and delta:
            raise ValueError(
                "Cannot use both 'peft' and 'delta' options at the same time."
            )

579
580
        if peft:
            if model_kwargs.get("load_in_4bit", None):
WoosungMyung's avatar
WoosungMyung committed
581
582
                if version.parse(PEFT_VERSION) < version.parse("0.4.0"):
                    raise AssertionError("load_in_4bit requires peft >= 0.4.0")
583
584
585
586
            if self._model.config.vocab_size != len(self.tokenizer):
                # resize model for LoRAs with added tokens
                self._model.resize_token_embeddings(len(self.tokenizer))
                eval_logger.info(f"Model config indicates vocab_size='{self._model.config.vocab_size}', but found tokenizer with vocab size '{len(self.tokenizer)}'. Resizing model embedding layer...") 
587
588
589
            self._model = PeftModel.from_pretrained(
                self._model, peft, revision=revision
            )
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
        elif delta:
            if autogptq:
                eval_logger.warning(
                    "Delta weights might trigger unexpected behavior when used with AutoGPTQ."
                )
            _model_delta = self.AUTO_MODEL_CLASS.from_pretrained(
                delta,
                revision=revision,
                torch_dtype=get_dtype(dtype),
                trust_remote_code=trust_remote_code,
                **model_kwargs,
            )
            for name, param in self._model.state_dict().items():
                try:
                    param.data += _model_delta.state_dict()[name]
                except KeyError:
                    raise KeyError(f"Delta model is missing weights for layer: {name}")
                except Exception as e:
                    raise RuntimeError(
                        f"Failed to add delta weights to layer {name}. Error: {e}"
                    )

            del _model_delta
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664

        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
665
    def _detect_batch_size(self, requests=None, pos: int = 0):
Benjamin Fattori's avatar
Benjamin Fattori committed
666
667
668
669
670
        if requests:
            _, context_enc, continuation_enc = requests[pos]
            max_length = len(
                (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1]
            )
671
672
            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
673
674
        else:
            max_length = self.max_length
675
676
            max_context_enc = max_length
            max_cont_enc = max_length
lintangsutawika's avatar
lintangsutawika committed
677

Benjamin Fattori's avatar
Benjamin Fattori committed
678
679
680
        # 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):
681
682
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                length = max(max_context_enc, max_cont_enc)
lintangsutawika's avatar
lintangsutawika committed
683
684
685
                batched_conts = torch.ones(
                    (batch_size, length), device=self.device
                ).long()
686
687
                test_batch = torch.ones((batch_size, length), device=self.device).long()
                call_kwargs = {
lintangsutawika's avatar
lintangsutawika committed
688
689
690
                    "attn_mask": test_batch,
                    "labels": batched_conts,
                }
691
692
            else:
                call_kwargs = {}
lintangsutawika's avatar
lintangsutawika committed
693
694
695
                test_batch = torch.ones(
                    (batch_size, max_length), device=self.device
                ).long()
Benjamin Fattori's avatar
Benjamin Fattori committed
696
            for _ in range(5):
697
                out = F.log_softmax(self._model_call(test_batch, **call_kwargs), dim=-1)  # noqa: F841
lintangsutawika's avatar
lintangsutawika committed
698

Benjamin Fattori's avatar
Benjamin Fattori committed
699
700
            return batch_size

701
702
703
704
705
706
707
        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
708

709
710
711
712
713
714
715
        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)
716
            clear_torch_cache()
717
718
            return batch_size

719
        clear_torch_cache()
Benjamin Fattori's avatar
Benjamin Fattori committed
720
721
        return batch_size

baberabb's avatar
baberabb committed
722
723
724
    def tok_encode(
        self, string: str, left_truncate_len=None, add_special_tokens=None
    ) -> List[int]:
haileyschoelkopf's avatar
haileyschoelkopf committed
725
        """ """
Lintang Sutawika's avatar
Lintang Sutawika committed
726
727
728
729
730
        # default for None - empty dict, use predefined tokenizer param
        # used for all models except for CausalLM or predefined value
        special_tokens_kwargs = {}

        # by default for CausalLM - false or self.add_bos_token is set
731
732
        if add_special_tokens is None:
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
Lintang Sutawika's avatar
Lintang Sutawika committed
733
734
735
736
737
738
                special_tokens_kwargs = {
                    "add_special_tokens": False or self.add_bos_token
                }
        # otherwise the method explicitly defines the value
        else:
            special_tokens_kwargs = {"add_special_tokens": add_special_tokens}
739

Lintang Sutawika's avatar
Lintang Sutawika committed
740
        encoding = self.tokenizer.encode(string, **special_tokens_kwargs)
haileyschoelkopf's avatar
haileyschoelkopf committed
741

742
743
744
        # 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
745

746
747
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
748
    def tok_batch_encode(
lintangsutawika's avatar
lintangsutawika committed
749
750
        self,
        strings: List[str],
lintangsutawika's avatar
lintangsutawika committed
751
        padding_side: str = "left",
752
753
        left_truncate_len: int = None,
        truncation: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
754
    ) -> Tuple[torch.Tensor, torch.Tensor]:
haileyschoelkopf's avatar
haileyschoelkopf committed
755
756
757
758
        # 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

Lintang Sutawika's avatar
Lintang Sutawika committed
759
        add_special_tokens = {}
haileyschoelkopf's avatar
haileyschoelkopf committed
760
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
Lintang Sutawika's avatar
Lintang Sutawika committed
761
            add_special_tokens = {"add_special_tokens": False or self.add_bos_token}
haileyschoelkopf's avatar
haileyschoelkopf committed
762
763
764

        encoding = self.tokenizer(
            strings,
lintangsutawika's avatar
lintangsutawika committed
765
            truncation=truncation,
haileyschoelkopf's avatar
haileyschoelkopf committed
766
767
            padding="longest",
            return_tensors="pt",
Lintang Sutawika's avatar
Lintang Sutawika committed
768
            **add_special_tokens,
haileyschoelkopf's avatar
haileyschoelkopf committed
769
770
771
772
773
774
775
776
777
778
        )
        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"]

Lintang Sutawika's avatar
Lintang Sutawika committed
779
780
    def tok_decode(self, tokens, skip_special_tokens=True):
        return self.tokenizer.decode(tokens, skip_special_tokens=skip_special_tokens)
781
782
783

    def _model_call(self, inps, attn_mask=None, labels=None):
        """
haileyschoelkopf's avatar
haileyschoelkopf committed
784
        :param inps: torch.Tensor
785
786
787
788
789
790
791
792
793
794
795
796
797
            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():
798
799
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
800
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
haileyschoelkopf's avatar
haileyschoelkopf committed
801
802
803
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
804
805
806
807
808
            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
809
        # temperature = 0.0 if not set
810
811
812
        # 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
813
        generation_kwargs["temperature"] = generation_kwargs.get("temperature", 0.0)
814
        do_sample = generation_kwargs.get("do_sample", None)
815
816
817
818
819

        # 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
820
821
        if do_sample is False and generation_kwargs.get("temperature") == 0.0:
            generation_kwargs.pop("temperature")
822
823
        # build stopping criteria
        stopping_criteria = stop_sequences_criteria(
824
            self.tokenizer, stop, context.shape[1], context.shape[0]
825
        )
826
        return self.model.generate(
827
            input_ids=context,
828
829
            max_length=max_length,
            stopping_criteria=stopping_criteria,
830
            pad_token_id=self.tokenizer.pad_token_id,
831
832
833
            use_cache=True,
            **generation_kwargs,
        )
834

Baber Abbasi's avatar
Baber Abbasi committed
835
836
837
    def _select_cont_toks(
        self, logits: torch.Tensor, contlen: int = None, inplen: int = None
    ) -> torch.Tensor:
838
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
839
840
841
            assert (
                contlen and inplen
            ), "Must pass input len and cont. len to select scored logits for causal LM"
842
843
844
845
            # 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
846
847
848
849
            assert (
                contlen and not inplen
            ), "Selecting scored logits for Seq2SeqLM requires only cont. len"
            # only discard right-padding.
850
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
851
852
            logits = logits[:contlen]

853
854
        return logits

855
856
857
    def loglikelihood_rolling(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[float]:
858
        loglikelihoods = []
Benjamin Fattori's avatar
Benjamin Fattori committed
859
860
861
862
863
864
865
866
867

        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

868
869
870
        for (string,) in tqdm(
            [req.args for req in requests], disable=(disable_tqdm or (self.rank != 0))
        ):
871
872
873
874
875
            rolling_token_windows = list(
                map(
                    utils.make_disjoint_window,
                    utils.get_rolling_token_windows(
                        token_list=self.tok_encode(string),
876
                        prefix_token=self.prefix_token_id,
877
878
879
880
881
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
882
883

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
            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
899
                requests=rolling_token_windows,
lintangsutawika's avatar
lintangsutawika committed
900
901
                disable_tqdm=True,
                override_bs=adaptive_batch_size,
902
903
904
905
906
907
908
909
910
911
912
913
            )

            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
914

915
916
917
918
919
920
921
922
923
924
925
926
927
    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
928
        self.batch_sizes[sched] = self._detect_batch_size(n_reordered_requests, pos)
929
930
        print(f"Determined largest batch size: {self.batch_sizes[sched]}")
        return self.batch_sizes[sched]
931

Ethan Smith's avatar
Ethan Smith committed
932
    def _loglikelihood_tokens(
baberabb's avatar
baberabb committed
933
934
935
936
937
        self,
        requests: List[Tuple[Tuple[str, str], List[int], List[int]]],
        disable_tqdm: bool = False,
        override_bs: int = None,
    ) -> List[Tuple[float, bool]]:
938
939
940
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

Baber Abbasi's avatar
Baber Abbasi committed
941
        def _collate(req: Tuple[Tuple[str, str], List[int], List[int]]):
Baber Abbasi's avatar
Baber Abbasi committed
942
            """Defines the key for the sorted method"""
943
944
945
946
947
948
949
            # 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
950
            toks = req[1] + req[2]
951
952
            return -len(toks), tuple(toks)

Baber Abbasi's avatar
Baber Abbasi committed
953
954
955
        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)"
Baber Abbasi's avatar
Baber Abbasi committed
956
            # allows for the creation of a lookup, so we can reuse logits in case of one-token continuations.
Baber Abbasi's avatar
Baber Abbasi committed
957
958
959
960
961
962
963
964
965
966
967
968
969
            # 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
970
971
972

        # automatic (variable) batch size detection for vectorization
        # pull longest context sample from request
Baber Abbasi's avatar
Baber Abbasi committed
973
974
975
        n_reordered_requests = len(re_ord)
        batch_size = (
            self.batch_size
976
977
978
            if self.batch_size != "auto"
            else override_bs
            if override_bs is not None
Baber Abbasi's avatar
Baber Abbasi committed
979
980
981
982
            else 0
        )
        batch_fn = (
            self._batch_scheduler
983
984
985
            if self.batch_size == "auto"
            and n_reordered_requests > 0
            and not override_bs
Baber Abbasi's avatar
Baber Abbasi committed
986
            else None
987
988
        )

Baber Abbasi's avatar
Baber Abbasi committed
989
        chunks = re_ord.get_batched(n=batch_size, batch_fn=batch_fn)
990
991
992
993
994
        pbar = tqdm(
            total=len(requests),
            disable=(disable_tqdm or (self.rank != 0)),
            desc="Running loglikelihood requests",
        )
haileyschoelkopf's avatar
haileyschoelkopf committed
995
        for chunk in chunks:
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
            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
1015
                # how this all works (illustrated on a causal decoder-only setup):
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
                #          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,
1027
1028
                        device=self.device,
                    )
1029
1030
1031
1032
1033
                    (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
1034
                        device=self.device,
1035
                    )
1036
                    (inplen,) = inp.shape
1037
1038
1039
1040

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

1041
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
1042
                        (continuation_enc)[-self.max_length :],
1043
1044
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
1045
                        dtype=torch.long,
1046
1047
                        device=self.device,
                    )
1048
1049
                    (contlen,) = cont.shape

1050
1051
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
1052
1053
1054
1055
1056
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
1057

haileyschoelkopf's avatar
haileyschoelkopf committed
1058
1059
1060
1061
1062
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
1063
1064
1065
1066

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
1067

1068
1069
1070
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
1071
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1072
1073
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
1074
1075
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                # TODO: left-pad encoder inps and mask?
1076
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1077
1078
                    padding_len_inp, inps
                )  # [batch, padding_len_inp]
1079
                batched_conts = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1080
1081
                    padding_len_cont, conts
                )  # [batch, padding_len_cont]
1082
                batched_encoder_mask = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1083
1084
1085
1086
1087
1088
                    padding_len_inp, encoder_attns
                )  # [batch, padding_len_inp]
                call_kwargs = {
                    "attn_mask": batched_encoder_mask,
                    "labels": batched_conts,
                }
1089
1090
1091

            multi_logits = F.log_softmax(
                self._model_call(batched_inps, **call_kwargs), dim=-1
1092
            )  # [batch, padding_length (inp or cont), vocab]
1093

Baber Abbasi's avatar
Baber Abbasi committed
1094
            for (request_str, ctx_tokens, _), logits, inplen, cont_toks in zip(
1095
1096
1097
1098
                chunk, multi_logits, inplens, cont_toks_list
            ):
                # Slice to original seq length
                contlen = len(cont_toks)
haileyschoelkopf's avatar
haileyschoelkopf committed
1099
                # take only logits in the continuation
1100
                # (discard context toks if decoder-only ; discard right-padding)
1101
1102
                # 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
1103
                ctx_len = (
1104
                    inplen + (logits.shape[0] - padding_len_inp)
haileyschoelkopf's avatar
haileyschoelkopf committed
1105
1106
1107
                    if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                    else None
                )
1108
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
1109
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
1110
1111
1112
1113

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

Baber Abbasi's avatar
Baber Abbasi committed
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
                # 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
1143
1144

        pbar.close()
haileyschoelkopf's avatar
haileyschoelkopf committed
1145

1146
1147
        return re_ord.get_original(res)

1148
1149
1150
    def generate_until(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1151
        res = []
1152

Baber Abbasi's avatar
Baber Abbasi committed
1153
        def _collate(req: Tuple[str, dict]):
Baber Abbasi's avatar
Baber Abbasi committed
1154
            """Defines the key for the sorted method"""
1155
1156
1157
1158
1159
1160
            # 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
1161
1162
            toks = self.tok_encode(req[0])
            return -len(toks), req[0]
1163

1164
1165
        pbar = tqdm(
            total=len(requests),
1166
            disable=(disable_tqdm or (self.rank != 0)),
1167
1168
            desc="Running generate_until requests",
        )
Baber Abbasi's avatar
Baber Abbasi committed
1169
        adaptive_batch_size = None
1170
1171
1172
1173
1174
1175
        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
1176
        # for each different set of kwargs, we execute all requests, by batch.
Baber Abbasi's avatar
Baber Abbasi committed
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
        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
        )
1189

Baber Abbasi's avatar
Baber Abbasi committed
1190
1191
1192
        # 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
1193
1194
1195
1196
1197
1198
1199
        # 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
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
        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):
achervyakov's avatar
achervyakov committed
1213
                        until = [until]
Baber Abbasi's avatar
Baber Abbasi committed
1214
1215
1216
1217
1218
1219
                    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
1220
                    f"Expected `kwargs` to be of type `dict` but got {type(gen_kwargs)}"
1221
                )
1222
            # add EOS token to stop sequences
Lintang Sutawika's avatar
Lintang Sutawika committed
1223
            eos = self.tok_decode(self.eot_token_id, skip_special_tokens=False)
Baber Abbasi's avatar
Baber Abbasi committed
1224
            if not until:
1225
1226
1227
                until = [eos]
            else:
                until.append(eos)
Baber Abbasi's avatar
Baber Abbasi committed
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
            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)
1249

Baber Abbasi's avatar
Baber Abbasi committed
1250
1251
            if "max_length" not in kwargs:
                kwargs["max_length"] = context_enc.shape[1] + max_gen_toks
1252

Baber Abbasi's avatar
Baber Abbasi committed
1253
1254
1255
1256
1257
1258
1259
            # perform batched generation
            cont = self._model_generate(
                context=context_enc,
                attention_mask=attn_masks,
                stop=until,
                **kwargs,
            )
1260

Baber Abbasi's avatar
Baber Abbasi committed
1261
1262
1263
1264
1265
            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] :]
1266

Baber Abbasi's avatar
Baber Abbasi committed
1267
                s = self.tok_decode(cont_toks)
1268

Baber Abbasi's avatar
Baber Abbasi committed
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
                # 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)
1282

1283
        pbar.close()
1284

Baber Abbasi's avatar
Baber Abbasi committed
1285
        return res
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326

    def get_model_info(self) -> dict:
        """
        Method to get Hugging Face model information for experiment reproducibility.
        """

        def get_model_num_params(model) -> int:
            if hasattr(model, "num_parameters"):
                return model.num_parameters()
            if hasattr(model, "parameters"):
                return sum(p.numel() for p in model.parameters())
            else:
                return -1

        def get_model_dtype(model) -> str:
            if hasattr(model, "dtype"):
                return model.dtype
            else:
                return ""

        def get_model_sha(pretrained: str, revision: str) -> str:
            try:
                model_info = HfApi().model_info(repo_id=pretrained, revision=revision)
                return model_info.sha
            except Exception as e:
                eval_logger.warn(
                    f"Failed to get model SHA for {pretrained} at revision {revision}. Error: {e}"
                )
                return ""

        model_info = {
            "model_num_parameters": get_model_num_params(self._model),
            "model_dtype": get_model_dtype(self._model),
            "model_revision": self.revision,
            "model_sha": get_model_sha(self.pretrained, self.revision),
        }
        if self.peft:
            model_info["peft_sha"] = get_model_sha(self.peft, self.revision)
        if self.delta:
            model_info["delta_sha"] = get_model_sha(self.delta, self.revision)
        return model_info