huggingface.py 62.9 KB
Newer Older
1
import copy
Lintang Sutawika's avatar
Lintang Sutawika committed
2
import logging
3
import os
Jeevan's avatar
Jeevan committed
4
from datetime import timedelta
5
from pathlib import Path
6
from typing import TYPE_CHECKING, Dict, List, Literal, Optional, Tuple, Union
7

8
import jinja2
9
import torch
10
import torch.nn.functional as F
11
import transformers
Jeevan's avatar
Jeevan committed
12
13
14
15
16
from accelerate import (
    Accelerator,
    InitProcessGroupKwargs,
    find_executable_batch_size,
)
Nathan Habib's avatar
Nathan Habib committed
17
from accelerate.utils import get_max_memory
18
from huggingface_hub import HfApi
19
20
from packaging import 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
from lm_eval.models.utils import (
    Collator,
    clear_torch_cache,
33
    configure_pad_token,
34
    get_dtype,
35
    handle_stop_sequences,
36
37
38
    pad_and_concat,
    stop_sequences_criteria,
)
39

40

41
42
43
if TYPE_CHECKING:
    from transformers.quantizers import AutoQuantizationConfig

Lintang Sutawika's avatar
Lintang Sutawika committed
44
eval_logger = logging.getLogger(__name__)
45

lintangsutawika's avatar
lintangsutawika committed
46

47
@register_model("hf-auto", "hf", "huggingface")
48
class HFLM(TemplateLM):
49
50
51
52
53
54
55
    """
    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.
    """

56
    AUTO_MODEL_CLASS = None
57
    _DEFAULT_MAX_LENGTH = 2048
haileyschoelkopf's avatar
haileyschoelkopf committed
58

59
60
    def __init__(
        self,
61
        pretrained: Union[str, transformers.PreTrainedModel],
62
        backend: Literal["default", "causal", "seq2seq"] = "default",
Baber Abbasi's avatar
Baber Abbasi committed
63
        # override whether the model should be treated as decoder-only (causal) or encoder-decoder (seq2seq)
64
        revision: Optional[str] = "main",
65
        subfolder: str = "",
66
67
68
69
70
71
72
        tokenizer: Optional[
            Union[
                str,
                transformers.PreTrainedTokenizer,
                transformers.PreTrainedTokenizerFast,
            ]
        ] = None,
lintangsutawika's avatar
lintangsutawika committed
73
        truncation: Optional[bool] = False,
Baber Abbasi's avatar
Baber Abbasi committed
74
        logits_cache: bool = True,
75
76
        max_length: Optional[int] = None,
        device: Optional[str] = "cuda",
77
        dtype: Optional[Union[str, torch.dtype]] = "auto",
78
        softmax_dtype: Optional[Union[str, torch.dtype]] = None,
79
        mixed_precision_dtype: Optional[Union[str, torch.dtype]] = None,
Benjamin Fattori's avatar
Benjamin Fattori committed
80
81
        batch_size: Optional[Union[int, str]] = 1,
        max_batch_size: Optional[int] = 64,
82
        trust_remote_code: Optional[bool] = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
83
        use_fast_tokenizer: Optional[bool] = True,
84
        add_bos_token: Optional[bool] = False,
85
        prefix_token_id: Optional[int] = None,
86
        # arguments used for splitting a model across GPUs naively.
87
88
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
89
90
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
91
        offload_folder: Optional[Union[str, os.PathLike]] = "./offload",
92
        # PEFT, delta weights and quantization options
93
        peft: Optional[str] = None,
94
        delta: Optional[str] = None,
95
        autogptq: Optional[Union[bool, str]] = False,
96
        gptqmodel: Optional[bool] = False,
97
        gguf_file: Optional[str] = None,
98
        **kwargs,
Ethan Smith's avatar
Ethan Smith committed
99
    ) -> None:
100
        super().__init__()
101
102
103
104
        # 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."
105
            )
Baber Abbasi's avatar
Baber Abbasi committed
106
107
108
            assert not parallelize, (
                "`parallelize=True` is not compatible with passing pre-initialized model to `pretrained`"
            )
109
110
111
            self._model = pretrained
            self._device = self._model.device
            self._config = self._model.config
Baber Abbasi's avatar
Baber Abbasi committed
112
            gpus = 0
113

114
        else:
115
116
117
118
119
            assert isinstance(device, str)
            assert isinstance(pretrained, str)
            assert isinstance(batch_size, (int, str))

            gpus = torch.cuda.device_count()
Jeevan's avatar
Jeevan committed
120
121
            accelerator_kwargs = InitProcessGroupKwargs(timeout=timedelta(weeks=52))
            accelerator = Accelerator(kwargs_handlers=[accelerator_kwargs])
122
123
            if accelerator.num_processes > 1:
                self.accelerator = accelerator
124

125
126
127
            if "npu" in accelerator.device.type:
                gpus = torch.npu.device_count()

Nathan Habib's avatar
Nathan Habib committed
128
            # using one process with no model parallelism
129
130
131
132
            if not (parallelize or accelerator.num_processes > 1):
                # use user-passed device
                device_list = set(
                    ["cuda", "cpu"]
133
                    + [f"cuda:{i}" for i in range(gpus)]
134
                    + ["mps", "mps:0"]
135
                    + [f"npu:{i}" for i in range(gpus)]
136
                )
137
                if device and device in device_list:
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
                    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")
                    )
Nathan Habib's avatar
Nathan Habib committed
154
            else:  # Parallelism managed by accelerate
155
156
157
158
159
                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
Nathan Habib's avatar
Nathan Habib committed
160
161
162
163
164
                self._device = (
                    self.accelerator.device
                    if hasattr(self, "accelerator")
                    else torch.device(device)
                )
165

Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
166
            revision = str(revision)  # cast to string if not already one
167

168
            self._get_config(
169
170
171
                pretrained,
                revision=revision,
                trust_remote_code=trust_remote_code,
172
                gguf_file=gguf_file,
173
                subfolder=subfolder,
174
175
            )

176
            # determine which of 'causal' and 'seq2seq' backends to use for HF models
177
178
179
        self._get_backend(
            config=self.config, backend=backend, trust_remote_code=trust_remote_code
        )
180

181
182
183
184
185
        # load tokenizer so we know tokenizer vocabulary size before loading model and PEFT
        self._create_tokenizer(
            pretrained,
            tokenizer,
            revision=revision,
186
            subfolder=subfolder,
187
188
            trust_remote_code=trust_remote_code,
            use_fast_tokenizer=use_fast_tokenizer,
189
            gguf_file=gguf_file,
190
            add_bos_token=add_bos_token,
191
192
        )

193
194
195
196
197
198
199
        if (
            quantization_config := getattr(self.config, "quantization_config", None)
        ) is not None and isinstance(quantization_config, dict):
            from transformers.quantizers import AutoQuantizationConfig

            quantization_config = AutoQuantizationConfig.from_dict(quantization_config)

200
201
202
203
204
205
206
207
        # 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,
208
                gpus=gpus,
209
210
211
212
                max_memory_per_gpu=max_memory_per_gpu,
                max_cpu_memory=max_cpu_memory,
                offload_folder=offload_folder,
                peft=peft,
213
                delta=delta,
214
                autogptq=autogptq,
215
                gptqmodel=gptqmodel,
216
                gguf_file=gguf_file,
217
                quantization_config=quantization_config,
218
                subfolder=subfolder,
219
                **kwargs,
220
221
            )

222
        # access self._model through self.model property outside this method
223
224
225
        if isinstance(self.model, torch.nn.Module):
            self.model.eval()
            self.model.tie_weights()
haileyschoelkopf's avatar
haileyschoelkopf committed
226

lintangsutawika's avatar
lintangsutawika committed
227
        self.truncation = truncation
Baber Abbasi's avatar
Baber Abbasi committed
228
        self.logits_cache = logits_cache
229
        self.vocab_size = self.tokenizer.vocab_size
230
        # select (or create) a pad token to use
231
        self.tokenizer = configure_pad_token(self.tokenizer, model_config=self.config)
232

233
        self.add_bos_token = add_bos_token
234
        if "gemma" in getattr(self.config, "model_type", ""):
235
            self.add_bos_token = True
236
            eval_logger.info(
237
                f"Model type is '{self.config.model_type}', part of the Gemma family--a BOS token will be used as Gemma underperforms without it."
238
239
            )

240
        self._max_length = max_length
241
242
243
244
        self.pretrained = pretrained
        self.delta = delta
        self.peft = peft
        self.revision = revision
Benjamin Fattori's avatar
Benjamin Fattori committed
245
246
247
        self.batch_schedule = 1
        self.batch_sizes = {}
        self.max_batch_size = max_batch_size
248
249
250
        self.softmax_dtype = (
            get_dtype(softmax_dtype) if softmax_dtype is not None else None
        )
251
252
253
254
255
        self.mixed_precision_dtype = (
            get_dtype(mixed_precision_dtype)
            if mixed_precision_dtype is not None
            else None
        )
Benjamin Fattori's avatar
Benjamin Fattori committed
256
257
258
259
260
261
262

        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)
263

264
        if isinstance(pretrained, str):
Nathan Habib's avatar
Nathan Habib committed
265
266
267
268
269
270
271
272
273
274
275
276
            if gpus >= 1 or str(self.device) == "mps":
                # TODO: can remove this whole snippet except in the mps case, perhaps?
                if not (parallelize or autogptq or hasattr(self, "accelerator")):
                    # 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:
                        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."
                        )
277
278
            # multigpu data-parallel support when launched with accelerate
            if gpus > 1:
Nathan Habib's avatar
Nathan Habib committed
279
280
281
282
                if accelerator.num_processes > 1:
                    if parallelize:
                        eval_logger.warning(
                            "You are both using a HF Accelerate `device_map` (`--model_args parallelize=True`) and launching via `accelerate launch`. This will attempt to do model and data parallelism depending on the resources available."
283
                        )
Nathan Habib's avatar
Nathan Habib committed
284
                    elif gpus > accelerator.num_processes:
285
286
287
288
289
290
                        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."
                        )
Nathan Habib's avatar
Nathan Habib committed
291
292
293
294
295
                        if self.accelerator.is_local_main_process:
                            eval_logger.info(
                                f"Using {gpus} devices with data parallelism"
                            )

296
                    self._device = torch.device(f"{accelerator.device}")
297
                    self.accelerator = accelerator
298

299
300
                    self._rank = self.accelerator.local_process_index
                    self._world_size = self.accelerator.num_processes
Nathan Habib's avatar
Nathan Habib committed
301
302
303
304
                else:
                    # if we aren't launching via accelerate, ditch
                    self._rank = 0
                    self._world_size = 1
305
306
307
308
309
310
311
        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
312

313
        self.custom_prefix_token_id = prefix_token_id
314
315
316
317
        if prefix_token_id is not None:
            eval_logger.info(
                f"Loglikelihood prefix token id used in evaluation: {self.prefix_token_id}"
            )
318

Nathan Habib's avatar
Nathan Habib committed
319
320
    def _get_accelerate_args(
        self,
321
        parallelize: Optional[bool] = None,
Nathan Habib's avatar
Nathan Habib committed
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
        device_map: 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",
        gpus: Optional[int] = None,
    ) -> dict:
        """Returns the kwargs needed to apply `accelerate` in `AutoModel.from_pretrained`."""
        num_local_processes = int(os.environ.get("LOCAL_WORLD_SIZE", 1))
        num_machines = int(os.environ.get("WORLD_SIZE", 0)) // num_local_processes
        if (
            num_machines == 0
            and hasattr(self, "accelerator")
            and self.accelerator is not None
        ):
            eval_logger.info(
                "We are not in a distributed setting for accelerate. Setting model_parallel to False."
            )
            parallelize = False

        if parallelize is None:
            # If parallelism is unset by the user, we automatically assign model parallelism
            # if enough extra GPUs are available
            max_memory_all_gpus = get_max_memory()
            # We just want gpu, not cpu, max memory
            if "cpu" in max_memory_all_gpus:
                del max_memory_all_gpus["cpu"]
            parallelize = bool(num_local_processes < len(max_memory_all_gpus))
            eval_logger.info(
                f"Setting model parallel to {parallelize} since "
                f"the number of local processes is {num_local_processes} "
                f"and the number of GPUs is {len(max_memory_all_gpus)}"
            )

        args = {}
        if parallelize:  # Model parallelism will be used
            max_memory = {}
            if max_memory_per_gpu is not None:  # Using the provided memory requirements
                max_memory_per_gpu_map = {
                    device_idx: max_memory_per_gpu for device_idx in range(gpus)
                }
            else:  # Estimating the possible memory requirements
                max_memory_all_gpus = get_max_memory()
                if "cpu" in max_memory_all_gpus:
                    del max_memory_all_gpus["cpu"]
                if not hasattr(self, "accelerator"):
                    max_memory_per_gpu_map = {
Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
368
                        k: v for k, v in max_memory_all_gpus.items()
Nathan Habib's avatar
Nathan Habib committed
369
                    }
Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
370
                else:
Nathan Habib's avatar
Nathan Habib committed
371
372
373
374
375
376
377
378
                    # use only 1 / num_processes of the GPUs if we are running under accelerate launch
                    max_memory_per_gpu_map = {
                        k: v
                        for k, v in max_memory_all_gpus.items()
                        if k % num_local_processes
                        == (self.accelerator.process_index % num_local_processes)
                    }
            args["max_memory"] = max_memory_per_gpu_map
379
            args["device_map"] = "auto" if device_map is None else device_map
Nathan Habib's avatar
Nathan Habib committed
380
            eval_logger.info(
381
                f"Model parallel was set to True, setting max memory per GPU to {max_memory_per_gpu_map} and device map to {args.get('device_map')}"
Nathan Habib's avatar
Nathan Habib committed
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
            )

            if max_cpu_memory is not None:
                max_memory["cpu"] = max_cpu_memory

            args["offload_folder"] = offload_folder
        elif (
            device_map is None
        ):  # No model parallelism, we use the default provided device for our model
            if hasattr(self, "accelerator"):
                device_map = {"": f"{self.accelerator.device}"}
            else:
                device_map = {"": str(self.device)}
            args["max_memory"] = None
            args["device_map"] = device_map
            eval_logger.info(
                f"Model parallel was set to False, max memory was not set, and device map was set to {device_map}"
            )
        else:
            args["max_memory"] = None
            args["device_map"] = None
            eval_logger.info("Model parallel was set to False.")

        return args

407
408
409
410
411
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

412
413
414
415
416
417
418
419
    @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

420
421
422
423
424
    @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

425
426
427
428
429
430
431
432
433
    @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

434
435
    @property
    def max_length(self):
436
437
438
439
440
441
442
443
444
445
446
        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
447

448
    @property
Ethan Smith's avatar
Ethan Smith committed
449
    def max_gen_toks(self) -> int:
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
        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

KonradSzafer's avatar
KonradSzafer committed
468
469
470
471
    @property
    def tokenizer_name(self) -> str:
        return self.tokenizer.name_or_path.replace("/", "__")

472
473
    def _get_backend(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
474
        config: Union[transformers.PretrainedConfig, transformers.AutoConfig],
475
        backend: Literal["default", "causal", "seq2seq"] = "default",
476
477
478
479
        trust_remote_code: Optional[bool] = False,
    ) -> None:
        """
        Helper method during initialization.
480
        Determines the backend ("causal" (decoder-only) or "seq2seq" (encoder-decoder)) model type to be used.
481
        sets `self.AUTO_MODEL_CLASS` appropriately if not already set.
482
483
484

        **If not calling HFLM.__init__() or HFLM._get_backend() within a subclass of HFLM,
        user must set `self.backend` to be either "causal" or "seq2seq" manually!**
485
        """
486

487
488
489
490
491
        assert backend in ["default", "causal", "seq2seq"]

        if backend != "default":
            # if we've settled on non-default backend, use that manually
            if backend == "causal":
492
                self.backend = backend
493
            elif backend == "seq2seq":
494
                self.backend = backend
495
            eval_logger.info(
496
                f"Overrode HF model backend type, and using type '{self.backend}'"
497
498
499
500
501
502
503
504
505
506
            )
        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.
507
                self.backend = "seq2seq"
508
                eval_logger.debug(f"Using model type '{self.backend}'")
509
510
511
            elif (
                getattr(self.config, "model_type") in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
            ):
512
                self.backend = "causal"
513
                eval_logger.debug(f"Using model type '{self.backend}'")
514
515
516
517
518
            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."
519
                        "Setting backend to causal"
520
521
                    )
                # if model type is neither in HF transformers causal or seq2seq model registries
522
523
524
                # then we default to assuming AutoModelForCausalLM
                self.backend = "causal"
                eval_logger.info(
525
                    f"Model type cannot be determined. Using default model type '{self.backend}'"
526
                )
527

528
529
530
531
532
        if self.AUTO_MODEL_CLASS is None:
            if self.backend == "causal":
                self.AUTO_MODEL_CLASS = transformers.AutoModelForCausalLM
            elif self.backend == "seq2seq":
                self.AUTO_MODEL_CLASS = transformers.AutoModelForSeq2SeqLM
533
534
535
536
537
538

    def _get_config(
        self,
        pretrained: str,
        revision: str = "main",
        trust_remote_code: bool = False,
539
        gguf_file: Optional[str] = None,
540
        subfolder: str = "",
541
    ) -> None:
542
        """Return the model config for HuggingFace models"""
543
544
545
546
        self._config = transformers.AutoConfig.from_pretrained(
            pretrained,
            revision=revision,
            trust_remote_code=trust_remote_code,
547
            gguf_file=gguf_file,
548
            subfolder=subfolder,
549
550
551
552
553
554
555
556
557
558
559
560
        )

    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,
561
        gpus: Optional[int] = None,
562
563
564
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
        offload_folder: Optional[str] = "./offload",
565
        # PEFT, delta weights and quantization options
566
        peft: Optional[str] = None,
567
        delta: Optional[str] = None,
568
        autogptq: Optional[Union[bool, str]] = False,
569
        gptqmodel: Optional[bool] = False,
570
        gguf_file: Optional[str] = None,
571
        quantization_config: Optional["AutoQuantizationConfig"] = None,
572
        subfolder: str = "",
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
        **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 {}

Nathan Habib's avatar
Nathan Habib committed
589
590
591
592
593
594
595
596
        model_kwargs.update(
            self._get_accelerate_args(
                parallelize=parallelize,
                device_map=kwargs.get("device_map", None),
                max_memory_per_gpu=max_memory_per_gpu,
                max_cpu_memory=max_cpu_memory,
                offload_folder=offload_folder,
                gpus=gpus,
597
            )
Nathan Habib's avatar
Nathan Habib committed
598
        )
599

600
        if not autogptq and not gptqmodel:
601
            if model_kwargs.get("load_in_4bit", None):
Baber Abbasi's avatar
Baber Abbasi committed
602
603
604
                assert transformers.__version__ >= "4.30.0", (
                    "load_in_4bit requires transformers >= 4.30.0"
                )
605
606
607
            if transformers.__version__ >= "4.30.0":
                if model_kwargs.get("load_in_4bit", None):
                    if model_kwargs.get("bnb_4bit_compute_dtype", None):
608
                        model_kwargs["bnb_4bit_compute_dtype"] = get_dtype(
609
610
                            model_kwargs["bnb_4bit_compute_dtype"]
                        )
Nathan Habib's avatar
Nathan Habib committed
611

612
613
614
            self._model = self.AUTO_MODEL_CLASS.from_pretrained(
                pretrained,
                revision=revision,
615
                torch_dtype=get_dtype(dtype),
616
                trust_remote_code=trust_remote_code,
617
                gguf_file=gguf_file,
618
                quantization_config=quantization_config,
619
                subfolder=subfolder,
620
621
622
                **model_kwargs,
            )
        else:
623
624
625
            if autogptq and gptqmodel:
                raise ValueError(
                    "Cannot use both 'autogptq' and 'gptqmodel' options at the same time."
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
            if autogptq:
                try:
                    from auto_gptq import AutoGPTQForCausalLM
                except ModuleNotFoundError as exception:
                    raise type(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 gptqmodel:
                try:
                    from gptqmodel import GPTQModel
                except ModuleNotFoundError as exception:
                    raise type(exception)(
                        "Tried to load gptqmodel, but gptqmodel is not installed ",
                        "please install gptqmodel via `pip install gptqmodel --no-build-isolation` or `pip install lm-eval[gptqmodel] --no-build-isolation`",
                    )

                self._model = GPTQModel.from_quantized(
                    pretrained, trust_remote_code=trust_remote_code, **model_kwargs
                )
659

660
661
662
663
664
        if peft and delta:
            raise ValueError(
                "Cannot use both 'peft' and 'delta' options at the same time."
            )

665
        if peft:
666
667
668
            from peft import PeftModel
            from peft import __version__ as PEFT_VERSION

669
            if model_kwargs.get("load_in_4bit", None):
WoosungMyung's avatar
WoosungMyung committed
670
671
                if version.parse(PEFT_VERSION) < version.parse("0.4.0"):
                    raise AssertionError("load_in_4bit requires peft >= 0.4.0")
672
673
            if self._model.config.vocab_size != len(self.tokenizer):
                # resize model for LoRAs with added tokens
674
675
676
                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..."
                )
677
                self._model.resize_token_embeddings(len(self.tokenizer))
678
679
680
            self._model = PeftModel.from_pretrained(
                self._model, peft, revision=revision
            )
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
        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
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719

        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,
720
        gguf_file: Optional[str] = None,
721
        add_bos_token: Optional[bool] = False,
722
        subfolder: Optional[str] = "",
723
724
725
726
727
728
729
    ) -> 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.
        """
730
731
732
733
734
735
        kwargs = {
            "revision": revision,
            "trust_remote_code": trust_remote_code,
        }

        # gguf format embeds tokenizer and is not compatible with hf tokenizer `use_fast` param
736
        if not tokenizer and gguf_file is not None:
737
738
739
            kwargs["gguf_file"] = gguf_file
        else:
            kwargs["use_fast"] = use_fast_tokenizer
740

741
742
743
        if add_bos_token:
            kwargs["add_bos_token"] = True

744
745
746
        if subfolder:
            kwargs["subfolder"] = subfolder

747
748
749
        if tokenizer:
            if isinstance(tokenizer, str):
                self.tokenizer = transformers.AutoTokenizer.from_pretrained(
750
                    tokenizer, **kwargs
751
752
753
754
755
756
757
758
759
760
761
762
763
764
                )
            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(
765
                model_name, **kwargs
766
767
768
            )
        return None

Ethan Smith's avatar
Ethan Smith committed
769
    def _detect_batch_size(self, requests=None, pos: int = 0):
Benjamin Fattori's avatar
Benjamin Fattori committed
770
771
772
773
774
        if requests:
            _, context_enc, continuation_enc = requests[pos]
            max_length = len(
                (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1]
            )
775
776
            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
777
778
        else:
            max_length = self.max_length
779
780
            max_context_enc = max_length
            max_cont_enc = max_length
lintangsutawika's avatar
lintangsutawika committed
781

Benjamin Fattori's avatar
Benjamin Fattori committed
782
783
784
        # 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):
785
            if self.backend == "seq2seq":
786
                length = max(max_context_enc, max_cont_enc)
lintangsutawika's avatar
lintangsutawika committed
787
788
789
                batched_conts = torch.ones(
                    (batch_size, length), device=self.device
                ).long()
790
791
                test_batch = torch.ones((batch_size, length), device=self.device).long()
                call_kwargs = {
lintangsutawika's avatar
lintangsutawika committed
792
793
794
                    "attn_mask": test_batch,
                    "labels": batched_conts,
                }
795
796
            else:
                call_kwargs = {}
lintangsutawika's avatar
lintangsutawika committed
797
798
799
                test_batch = torch.ones(
                    (batch_size, max_length), device=self.device
                ).long()
Benjamin Fattori's avatar
Benjamin Fattori committed
800
            for _ in range(5):
801
802
803
804
805
                out = F.log_softmax(  # noqa: F841
                    self._model_call(test_batch, **call_kwargs),
                    dim=-1,
                    dtype=self.softmax_dtype,
                )
lintangsutawika's avatar
lintangsutawika committed
806

Benjamin Fattori's avatar
Benjamin Fattori committed
807
808
            return batch_size

809
810
811
812
813
814
815
        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
816

817
818
819
820
821
822
823
        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)
824
            clear_torch_cache()
825
826
            return batch_size

827
        clear_torch_cache()
Benjamin Fattori's avatar
Benjamin Fattori committed
828
829
        return batch_size

baberabb's avatar
baberabb committed
830
831
832
    def tok_encode(
        self, string: str, left_truncate_len=None, add_special_tokens=None
    ) -> List[int]:
haileyschoelkopf's avatar
haileyschoelkopf committed
833
        """ """
Lintang Sutawika's avatar
Lintang Sutawika committed
834
835
836
837
838
        # 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
839
        if add_special_tokens is None:
840
            if self.backend == "causal":
Lintang Sutawika's avatar
Lintang Sutawika committed
841
842
843
844
845
846
                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}
847

Lintang Sutawika's avatar
Lintang Sutawika committed
848
        encoding = self.tokenizer.encode(string, **special_tokens_kwargs)
haileyschoelkopf's avatar
haileyschoelkopf committed
849

850
851
852
        # 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
853

854
855
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
856
    def tok_batch_encode(
lintangsutawika's avatar
lintangsutawika committed
857
858
        self,
        strings: List[str],
lintangsutawika's avatar
lintangsutawika committed
859
        padding_side: str = "left",
860
861
        left_truncate_len: int = None,
        truncation: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
862
    ) -> Tuple[torch.Tensor, torch.Tensor]:
haileyschoelkopf's avatar
haileyschoelkopf committed
863
864
865
866
        # 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
867
        add_special_tokens = {}
868
        if self.backend == "causal":
Lintang Sutawika's avatar
Lintang Sutawika committed
869
            add_special_tokens = {"add_special_tokens": False or self.add_bos_token}
haileyschoelkopf's avatar
haileyschoelkopf committed
870
871
872

        encoding = self.tokenizer(
            strings,
lintangsutawika's avatar
lintangsutawika committed
873
            truncation=truncation,
haileyschoelkopf's avatar
haileyschoelkopf committed
874
875
            padding="longest",
            return_tensors="pt",
Lintang Sutawika's avatar
Lintang Sutawika committed
876
            **add_special_tokens,
haileyschoelkopf's avatar
haileyschoelkopf committed
877
878
        )
        if left_truncate_len:
879
880
881
882
883
884
            original_lengths = encoding["input_ids"].size(1)
            if original_lengths > left_truncate_len:
                eval_logger.warn(
                    f"Left truncation applied. Original sequence length was {original_lengths}, "
                    f"truncating to last {left_truncate_len} tokens. Some content will be lost.",
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
885
886
887
888
889
890
891
892
            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
893
894
    def tok_decode(self, tokens, skip_special_tokens=True):
        return self.tokenizer.decode(tokens, skip_special_tokens=skip_special_tokens)
895
896
897

    def _model_call(self, inps, attn_mask=None, labels=None):
        """
haileyschoelkopf's avatar
haileyschoelkopf committed
898
        :param inps: torch.Tensor
899
900
901
902
903
904
905
906
907
908
909
910
911
            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():
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
            with torch.autocast(
                device_type=self.device.type,
                dtype=self.mixed_precision_dtype,
                enabled=self.mixed_precision_dtype is not None,
            ):
                if attn_mask is not None or labels is not None:
                    assert attn_mask is not None and labels is not None
                    assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
                    return self.model(
                        input_ids=inps, attention_mask=attn_mask, labels=labels
                    ).logits
                else:
                    assert self.AUTO_MODEL_CLASS in (
                        transformers.AutoModelForCausalLM,
                        transformers.AutoModelForVision2Seq,
                    )
                    return self.model(inps).logits
929
930

    def _model_generate(self, context, max_length, stop, **generation_kwargs):
Baber Abbasi's avatar
Baber Abbasi committed
931
        # temperature = 0.0 if not set
932
933
934
        # 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
935
        generation_kwargs["temperature"] = generation_kwargs.get("temperature", 0.0)
936
        do_sample = generation_kwargs.get("do_sample", None)
937
938
939
940
941

        # 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
942
943
        if do_sample is False and generation_kwargs.get("temperature") == 0.0:
            generation_kwargs.pop("temperature")
944
945
        # build stopping criteria
        stopping_criteria = stop_sequences_criteria(
946
            self.tokenizer, stop, context.shape[1], context.shape[0]
947
        )
948
949
950
951
952
953
954
955
956
957
958
959
960
        with torch.autocast(
            device_type=self.device.type,
            dtype=self.mixed_precision_dtype,
            enabled=self.mixed_precision_dtype is not None,
        ):
            return self.model.generate(
                input_ids=context,
                max_length=max_length,
                stopping_criteria=stopping_criteria,
                pad_token_id=self.tokenizer.pad_token_id,
                use_cache=True,
                **generation_kwargs,
            )
961

Baber Abbasi's avatar
Baber Abbasi committed
962
963
964
    def _select_cont_toks(
        self, logits: torch.Tensor, contlen: int = None, inplen: int = None
    ) -> torch.Tensor:
965
        if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
966
967
968
            assert contlen and inplen, (
                "Must pass input len and cont. len to select scored logits for causal LM"
            )
969
970
971
            # discard right-padding.
            # also discard the input/context tokens. we'll only score continuations.
            logits = logits[inplen - contlen : inplen]
972
        elif self.backend == "seq2seq":
Baber Abbasi's avatar
Baber Abbasi committed
973
974
975
            assert contlen and not inplen, (
                "Selecting scored logits for Seq2SeqLM requires only cont. len"
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
976
            # only discard right-padding.
977
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
978
979
            logits = logits[:contlen]

980
981
        return logits

982
983
984
    def loglikelihood_rolling(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[float]:
Benjamin Fattori's avatar
Benjamin Fattori committed
985
986
987
988
989
990
991
992
        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

993
994
995
996
997
998
999
1000
1001
        # First, collect all windows from all requests
        all_windows = []  # List of (request_idx, window) tuples
        request_window_counts = []  # Track number of windows per request

        for req_idx, (string,) in enumerate(
            tqdm(
                [req.args for req in requests],
                disable=(disable_tqdm or (self.rank != 0)),
            )
1002
        ):
1003
            rolling_token_windows: List[Tuple[List[int], List[int]]] = list(
1004
1005
1006
1007
                map(
                    utils.make_disjoint_window,
                    utils.get_rolling_token_windows(
                        token_list=self.tok_encode(string),
1008
                        prefix_token=self.prefix_token_id,
1009
1010
1011
1012
1013
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
1014
1015

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
1016
            windows = [(None,) + x for x in rolling_token_windows]
1017

1018
1019
1020
            # Store windows with their request index
            all_windows.extend((req_idx, window) for window in windows)
            request_window_counts.append(len(windows))
1021

1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
        # Handle distributed case padding
        pad_amnt = 0
        if self.world_size > 1:
            mytensor = torch.tensor(len(all_windows), device=self.device)
            gathered = self.accelerator.gather(mytensor).cpu().detach().numpy().tolist()
            pad_amnt = max(gathered) - gathered[self.rank]
            if pad_amnt > 0:
                all_windows += pad_amnt * [all_windows[0]]

        all_nlls = []
        batch_size = adaptive_batch_size or self.batch_size
        for i in range(0, len(all_windows), batch_size):
            batch = all_windows[i : i + batch_size]
            # Extract just the windows for processing, keeping track of request indices
            batch_indices, batch_windows = zip(*batch)

            batch_nlls = self._loglikelihood_tokens(
                requests=batch_windows,
                disable_tqdm=False,
                override_bs=len(batch_windows),
1042
            )
1043
1044
            # Store results with their request indices
            all_nlls.extend(zip(batch_indices, batch_nlls))
1045

1046
1047
1048
        # Remove padding if necessary
        if (self.world_size > 1) and (pad_amnt > 0):
            all_nlls = all_nlls[:-pad_amnt]
1049

1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
        # Reconstruct per-request loglikelihoods
        loglikelihoods = []
        current_idx = 0
        for window_count in request_window_counts:
            # Get all nlls for this request
            request_nlls = all_nlls[current_idx : current_idx + window_count]
            # Sum up the nlls for this request (discarding is_greedy)
            request_total = sum(nll[0] for _, nll in request_nlls)
            loglikelihoods.append(request_total)
            current_idx += window_count

            string = requests[len(loglikelihoods) - 1].args[0]
            self.cache_hook.add_partial(
                "loglikelihood_rolling", (string,), request_total
            )
1065

1066
        return loglikelihoods
Zhiwei Zhuang's avatar
Zhiwei Zhuang committed
1067

1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
    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
1081
        self.batch_sizes[sched] = self._detect_batch_size(n_reordered_requests, pos)
1082
1083
        print(f"Determined largest batch size: {self.batch_sizes[sched]}")
        return self.batch_sizes[sched]
1084

Ethan Smith's avatar
Ethan Smith committed
1085
    def _loglikelihood_tokens(
baberabb's avatar
baberabb committed
1086
1087
1088
1089
1090
        self,
        requests: List[Tuple[Tuple[str, str], List[int], List[int]]],
        disable_tqdm: bool = False,
        override_bs: int = None,
    ) -> List[Tuple[float, bool]]:
1091
1092
1093
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

Baber Abbasi's avatar
Baber Abbasi committed
1094
        def _collate(req: Tuple[Tuple[str, str], List[int], List[int]]):
Baber Abbasi's avatar
Baber Abbasi committed
1095
            """Defines the key for the sorted method"""
1096
1097
1098
1099
1100
1101
1102
            # 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
1103
            toks = req[1] + req[2]
1104
1105
            return -len(toks), tuple(toks)

Baber Abbasi's avatar
Baber Abbasi committed
1106
1107
1108
        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
1109
            # 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
1110
1111
1112
1113
1114
1115
1116
1117
            # 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"
1118
            if self.backend == "causal" and self.logits_cache
Baber Abbasi's avatar
Baber Abbasi committed
1119
1120
1121
            else None,
            group_fn=_lookup_one_token_cont,
        )
Benjamin Fattori's avatar
Benjamin Fattori committed
1122
1123
1124

        # automatic (variable) batch size detection for vectorization
        # pull longest context sample from request
Baber Abbasi's avatar
Baber Abbasi committed
1125
1126
1127
        n_reordered_requests = len(re_ord)
        batch_size = (
            self.batch_size
1128
1129
1130
            if self.batch_size != "auto"
            else override_bs
            if override_bs is not None
Baber Abbasi's avatar
Baber Abbasi committed
1131
1132
1133
1134
            else 0
        )
        batch_fn = (
            self._batch_scheduler
1135
1136
1137
            if self.batch_size == "auto"
            and n_reordered_requests > 0
            and not override_bs
Baber Abbasi's avatar
Baber Abbasi committed
1138
            else None
1139
1140
        )

Baber Abbasi's avatar
Baber Abbasi committed
1141
        chunks = re_ord.get_batched(n=batch_size, batch_fn=batch_fn)
1142
1143
1144
1145
1146
        pbar = tqdm(
            total=len(requests),
            disable=(disable_tqdm or (self.rank != 0)),
            desc="Running loglikelihood requests",
        )
haileyschoelkopf's avatar
haileyschoelkopf committed
1147
        for chunk in chunks:
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
            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
1167
                # how this all works (illustrated on a causal decoder-only setup):
1168
1169
1170
1171
1172
1173
1174
                #          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
1175
                if self.backend == "causal":
1176
1177
                    total_length = len(context_enc) + len(continuation_enc)
                    if total_length > self.max_length + 1:
1178
                        eval_logger.warning(
1179
1180
1181
1182
                            f"Combined length of context ({len(context_enc)}) and continuation ({len(continuation_enc)}) "
                            f"exceeds model's maximum length ({self.max_length}). "
                            f"Truncating {total_length - self.max_length + 1} tokens from the left."
                        )
1183
1184
1185
                    inp = torch.tensor(
                        (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1],
                        dtype=torch.long,
1186
1187
                        device=self.device,
                    )
1188
                    (inplen,) = inp.shape
1189
                elif self.backend == "seq2seq":
1190
1191
1192
                    inp = torch.tensor(
                        (context_enc)[-self.max_length :],
                        dtype=torch.long,
haileyschoelkopf's avatar
haileyschoelkopf committed
1193
                        device=self.device,
1194
                    )
1195
                    (inplen,) = inp.shape
1196
1197
1198
1199

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

1200
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
1201
                        (continuation_enc)[-self.max_length :],
1202
1203
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
1204
                        dtype=torch.long,
1205
1206
                        device=self.device,
                    )
1207
1208
                    (contlen,) = cont.shape

1209
1210
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
1211
1212
1213
1214
1215
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
1216

haileyschoelkopf's avatar
haileyschoelkopf committed
1217
1218
1219
1220
1221
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
1222
1223
1224
1225

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
1226

1227
1228
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
1229
            if self.backend == "causal":
1230
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1231
1232
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
1233
            elif self.backend == "seq2seq":
1234
                # TODO: left-pad encoder inps and mask?
1235
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1236
1237
                    padding_len_inp, inps
                )  # [batch, padding_len_inp]
1238
                batched_conts = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1239
1240
                    padding_len_cont, conts
                )  # [batch, padding_len_cont]
1241
                batched_encoder_mask = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1242
1243
1244
1245
1246
1247
                    padding_len_inp, encoder_attns
                )  # [batch, padding_len_inp]
                call_kwargs = {
                    "attn_mask": batched_encoder_mask,
                    "labels": batched_conts,
                }
1248
1249

            multi_logits = F.log_softmax(
1250
1251
1252
                self._model_call(batched_inps, **call_kwargs),
                dim=-1,
                dtype=self.softmax_dtype,
1253
            )  # [batch, padding_length (inp or cont), vocab]
1254

Baber Abbasi's avatar
Baber Abbasi committed
1255
            for (request_str, ctx_tokens, _), logits, inplen, cont_toks in zip(
1256
1257
1258
1259
                chunk, multi_logits, inplens, cont_toks_list
            ):
                # Slice to original seq length
                contlen = len(cont_toks)
haileyschoelkopf's avatar
haileyschoelkopf committed
1260
                # take only logits in the continuation
1261
                # (discard context toks if decoder-only ; discard right-padding)
1262
1263
                # 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
1264
                ctx_len = (
1265
                    inplen + (logits.shape[0] - padding_len_inp)
1266
                    if self.backend == "causal"
haileyschoelkopf's avatar
haileyschoelkopf committed
1267
1268
                    else None
                )
1269
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
1270
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
1271
1272
1273
1274

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

Baber Abbasi's avatar
Baber Abbasi committed
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
                # 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]
1289
1290
1291
1292
1293
1294
                    # Use trailing slice [-cont_toks.shape[1]:] to handle variable length cont_len (but same ctx+cont[:-1]).
                    # i.e. continuations can be sliced at diff points. Collator ensures we have sufficient greedy_tokens
                    # by choosing key with longest cont if group_by="contexts".
                    max_equal = (
                        greedy_tokens[:, -cont_toks.shape[1] :] == cont_toks
                    ).all()
Baber Abbasi's avatar
Baber Abbasi committed
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306

                    # 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)

1307
1308
1309
1310
1311
1312
1313
                    if request_str is not None:
                        # special case: loglikelihood_rolling produces a number of loglikelihood requests
                        # all with cache key None. instead do add_partial on the per-example level
                        # in the loglikelihood_rolling() function for those.
                        self.cache_hook.add_partial(
                            "loglikelihood", request_str, answer
                        )
Baber Abbasi's avatar
Baber Abbasi committed
1314
                    pbar.update(1)
haileyschoelkopf's avatar
haileyschoelkopf committed
1315
1316

        pbar.close()
haileyschoelkopf's avatar
haileyschoelkopf committed
1317

1318
1319
        return re_ord.get_original(res)

1320
1321
1322
    def generate_until(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1323
        res = []
1324

Baber Abbasi's avatar
Baber Abbasi committed
1325
        def _collate(req: Tuple[str, dict]):
Baber Abbasi's avatar
Baber Abbasi committed
1326
            """Defines the key for the sorted method"""
1327
1328
1329
1330
1331
1332
            # 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
1333
1334
            toks = self.tok_encode(req[0])
            return -len(toks), req[0]
1335

1336
1337
        pbar = tqdm(
            total=len(requests),
1338
            disable=(disable_tqdm or (self.rank != 0)),
1339
1340
            desc="Running generate_until requests",
        )
Baber Abbasi's avatar
Baber Abbasi committed
1341
        adaptive_batch_size = None
1342
1343
1344
1345
1346
1347
        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
1348
        # for each different set of kwargs, we execute all requests, by batch.
Baber Abbasi's avatar
Baber Abbasi committed
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
        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
        )
1361

Baber Abbasi's avatar
Baber Abbasi committed
1362
1363
1364
        # 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
1365
1366
1367
1368
1369
1370
1371
        # 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
1372
        chunks = re_ords.get_batched(n=batch_size, batch_fn=batch_fn)
1373
        eos = self.tok_decode(self.eot_token_id, skip_special_tokens=False)
Baber Abbasi's avatar
Baber Abbasi committed
1374
1375
1376
1377
1378
1379
1380
1381
        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.
            if isinstance(gen_kwargs, dict):
                kwargs = copy.deepcopy(gen_kwargs)  # edge case for repeats > 1
1382
1383
                # add EOS token to stop sequences
                until = handle_stop_sequences(kwargs.pop("until", None), eos=eos)
Baber Abbasi's avatar
Baber Abbasi committed
1384
1385
            else:
                raise ValueError(
Baber Abbasi's avatar
Baber Abbasi committed
1386
                    f"Expected `kwargs` to be of type `dict` but got {type(gen_kwargs)}"
1387
                )
Baber Abbasi's avatar
Baber Abbasi committed
1388
1389
1390
1391
1392
1393
            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")
1394
            if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
1395
1396
                # max len for inputs = max length, minus room to generate the max new tokens
                max_ctx_len = self.max_length - max_gen_toks
Baber Abbasi's avatar
Baber Abbasi committed
1397
1398
1399
                assert max_ctx_len > 0, (
                    f"Invalid configuration: requested max tokens to generate ({max_gen_toks}) must be less than model's maximum sequence length ({self.max_length})."
                )
1400
            elif self.backend == "seq2seq":
Baber Abbasi's avatar
Baber Abbasi committed
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
                # 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)
1412

Baber Abbasi's avatar
Baber Abbasi committed
1413
1414
            if "max_length" not in kwargs:
                kwargs["max_length"] = context_enc.shape[1] + max_gen_toks
1415

Baber Abbasi's avatar
Baber Abbasi committed
1416
1417
1418
1419
1420
1421
1422
            # perform batched generation
            cont = self._model_generate(
                context=context_enc,
                attention_mask=attn_masks,
                stop=until,
                **kwargs,
            )
1423

Baber Abbasi's avatar
Baber Abbasi committed
1424
1425
1426
            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
1427
                if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
1428
                    cont_toks = cont_toks[context_enc.shape[1] :]
1429

Baber Abbasi's avatar
Baber Abbasi committed
1430
                s = self.tok_decode(cont_toks)
1431

Baber Abbasi's avatar
Baber Abbasi committed
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
                # 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)
1445

1446
        pbar.close()
1447

Baber Abbasi's avatar
Baber Abbasi committed
1448
        return res
1449

Baber Abbasi's avatar
Baber Abbasi committed
1450
1451
1452
    def apply_chat_template(
        self, chat_history: List[Dict[str, str]], add_generation_prompt: bool = True
    ) -> str:
KonradSzafer's avatar
KonradSzafer committed
1453
1454
1455
        """
        Method to apply a chat template to a list of chat history between user and model.
        """
1456
1457
        try:
            chat_templated = self.tokenizer.apply_chat_template(
Baber Abbasi's avatar
Baber Abbasi committed
1458
1459
1460
1461
                chat_history,
                tokenize=False,
                add_generation_prompt=add_generation_prompt,
                continue_final_message=not add_generation_prompt,
1462
1463
1464
1465
1466
1467
1468
            )
        except jinja2.exceptions.TemplateError:
            eval_logger.warning(
                "Failed to apply chat template. removing the system role in chat history."
            )
            chat_history = [msg for msg in chat_history if msg["role"] != "system"]
            chat_templated = self.tokenizer.apply_chat_template(
Baber Abbasi's avatar
Baber Abbasi committed
1469
1470
1471
1472
                chat_history,
                tokenize=False,
                add_generation_prompt=add_generation_prompt,
                continue_final_message=not add_generation_prompt,
1473
1474
1475
            )

        return chat_templated
KonradSzafer's avatar
KonradSzafer committed
1476

1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
    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:
Baber Abbasi's avatar
Baber Abbasi committed
1501
                eval_logger.debug(
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
                    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