huggingface.py 60.4 KB
Newer Older
1
import copy
2
import os
Jeevan's avatar
Jeevan committed
3
from datetime import timedelta
4
from pathlib import Path
KonradSzafer's avatar
KonradSzafer committed
5
from typing import Dict, List, Literal, Optional, Tuple, Union
6

7
import jinja2
8
import torch
9
import torch.nn.functional as F
10
import transformers
Jeevan's avatar
Jeevan committed
11
12
13
14
15
from accelerate import (
    Accelerator,
    InitProcessGroupKwargs,
    find_executable_batch_size,
)
Nathan Habib's avatar
Nathan Habib committed
16
from accelerate.utils import get_max_memory
17
from huggingface_hub import HfApi
18
19
20
21
from packaging import version
from peft import PeftModel
from peft import __version__ as PEFT_VERSION
from tqdm import tqdm
22
23
24
25
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
)
26
27

from lm_eval import utils
baberabb's avatar
baberabb committed
28
from lm_eval.api.instance import Instance
29
from lm_eval.api.model import TemplateLM
30
from lm_eval.api.registry import register_model
31
32
33
from lm_eval.models.utils import (
    Collator,
    clear_torch_cache,
34
    configure_pad_token,
35
    get_dtype,
36
    handle_stop_sequences,
37
38
39
    pad_and_concat,
    stop_sequences_criteria,
)
40

41

42
eval_logger = utils.eval_logger
43

lintangsutawika's avatar
lintangsutawika committed
44

45
@register_model("hf-auto", "hf", "huggingface")
46
class HFLM(TemplateLM):
47
48
49
50
51
52
53
    """
    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.
    """

54
    AUTO_MODEL_CLASS = None
55
    _DEFAULT_MAX_LENGTH = 2048
haileyschoelkopf's avatar
haileyschoelkopf committed
56

57
58
    def __init__(
        self,
59
        pretrained: Union[str, transformers.PreTrainedModel],
60
        backend: Literal["default", "causal", "seq2seq"] = "default",
Baber Abbasi's avatar
Baber Abbasi committed
61
        # override whether the model should be treated as decoder-only (causal) or encoder-decoder (seq2seq)
62
63
        revision: Optional[str] = "main",
        subfolder: Optional[str] = None,
64
65
66
67
68
69
70
        tokenizer: Optional[
            Union[
                str,
                transformers.PreTrainedTokenizer,
                transformers.PreTrainedTokenizerFast,
            ]
        ] = None,
lintangsutawika's avatar
lintangsutawika committed
71
        truncation: Optional[bool] = False,
Baber Abbasi's avatar
Baber Abbasi committed
72
        logits_cache: bool = True,
73
74
        max_length: Optional[int] = None,
        device: Optional[str] = "cuda",
75
        dtype: Optional[Union[str, torch.dtype]] = "auto",
Benjamin Fattori's avatar
Benjamin Fattori committed
76
77
        batch_size: Optional[Union[int, str]] = 1,
        max_batch_size: Optional[int] = 64,
78
        trust_remote_code: Optional[bool] = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
79
        use_fast_tokenizer: Optional[bool] = True,
80
        add_bos_token: Optional[bool] = False,
81
        prefix_token_id: Optional[int] = None,
82
        # arguments used for splitting a model across GPUs naively.
83
84
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
85
86
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
87
        offload_folder: Optional[Union[str, os.PathLike]] = "./offload",
88
        # PEFT, delta weights and quantization options
89
        peft: Optional[str] = None,
90
        delta: Optional[str] = None,
91
        autogptq: Optional[Union[bool, str]] = False,
92
        gptqmodel: Optional[bool] = False,
93
        gguf_file: Optional[str] = None,
94
        **kwargs,
Ethan Smith's avatar
Ethan Smith committed
95
    ) -> None:
96
        super().__init__()
97
98
99
100
        # 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."
101
            )
102
            assert not parallelize, "`parallelize=True` is not compatible with passing pre-initialized model to `pretrained`"
103
104
105
            self._model = pretrained
            self._device = self._model.device
            self._config = self._model.config
Baber Abbasi's avatar
Baber Abbasi committed
106
            gpus = 0
107

108
        else:
109
110
111
112
113
            assert isinstance(device, str)
            assert isinstance(pretrained, str)
            assert isinstance(batch_size, (int, str))

            gpus = torch.cuda.device_count()
Jeevan's avatar
Jeevan committed
114
115
            accelerator_kwargs = InitProcessGroupKwargs(timeout=timedelta(weeks=52))
            accelerator = Accelerator(kwargs_handlers=[accelerator_kwargs])
116
117
            if accelerator.num_processes > 1:
                self.accelerator = accelerator
118

119
120
121
            if "npu" in accelerator.device.type:
                gpus = torch.npu.device_count()

Nathan Habib's avatar
Nathan Habib committed
122
            # using one process with no model parallelism
123
124
125
126
            if not (parallelize or accelerator.num_processes > 1):
                # use user-passed device
                device_list = set(
                    ["cuda", "cpu"]
127
                    + [f"cuda:{i}" for i in range(gpus)]
128
                    + ["mps", "mps:0"]
129
                    + [f"npu:{i}" for i in range(gpus)]
130
                )
131
                if device and device in device_list:
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
                    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
148
            else:  # Parallelism managed by accelerate
149
150
151
152
153
                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
154
155
156
157
158
                self._device = (
                    self.accelerator.device
                    if hasattr(self, "accelerator")
                    else torch.device(device)
                )
159

Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
160
            revision = str(revision)  # cast to string if not already one
161
162
            # TODO: update this to be less of a hack once subfolder is fixed in HF
            revision = revision + ("/" + subfolder if subfolder is not None else "")
163

164
            self._get_config(
165
166
167
                pretrained,
                revision=revision,
                trust_remote_code=trust_remote_code,
168
                gguf_file=gguf_file,
169
170
            )

171
            # determine which of 'causal' and 'seq2seq' backends to use for HF models
172
173
174
        self._get_backend(
            config=self.config, backend=backend, trust_remote_code=trust_remote_code
        )
175

176
177
178
179
180
181
182
        # 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,
183
            gguf_file=gguf_file,
184
185
        )

186
187
188
189
190
191
192
193
        # 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,
194
                gpus=gpus,
195
196
197
198
                max_memory_per_gpu=max_memory_per_gpu,
                max_cpu_memory=max_cpu_memory,
                offload_folder=offload_folder,
                peft=peft,
199
                delta=delta,
200
                autogptq=autogptq,
201
                gptqmodel=gptqmodel,
202
                gguf_file=gguf_file,
203
                **kwargs,
204
205
            )

206
        # access self._model through self.model property outside this method
207
208
209
        if isinstance(self.model, torch.nn.Module):
            self.model.eval()
            self.model.tie_weights()
haileyschoelkopf's avatar
haileyschoelkopf committed
210

lintangsutawika's avatar
lintangsutawika committed
211
        self.truncation = truncation
Baber Abbasi's avatar
Baber Abbasi committed
212
        self.logits_cache = logits_cache
213
        self.vocab_size = self.tokenizer.vocab_size
214
        # select (or create) a pad token to use
215
        self.tokenizer = configure_pad_token(self.tokenizer, model_config=self.config)
216

217
        self.add_bos_token = add_bos_token
218
        if "gemma" in getattr(self.config, "model_type", ""):
219
            self.add_bos_token = True
220
            eval_logger.info(
221
                f"Model type is '{self.config.model_type}', part of the Gemma family--a BOS token will be used as Gemma underperforms without it."
222
223
            )

224
        self._max_length = max_length
225
226
227
228
        self.pretrained = pretrained
        self.delta = delta
        self.peft = peft
        self.revision = revision
Benjamin Fattori's avatar
Benjamin Fattori committed
229
230
231
232
233
234
235
236
237
238
        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)
239

240
        if isinstance(pretrained, str):
Nathan Habib's avatar
Nathan Habib committed
241
242
243
244
245
246
247
248
249
250
251
252
            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."
                        )
253
254
            # multigpu data-parallel support when launched with accelerate
            if gpus > 1:
Nathan Habib's avatar
Nathan Habib committed
255
256
257
258
                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."
259
                        )
Nathan Habib's avatar
Nathan Habib committed
260
                    elif gpus > accelerator.num_processes:
261
262
263
264
265
266
                        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
267
268
269
270
271
                        if self.accelerator.is_local_main_process:
                            eval_logger.info(
                                f"Using {gpus} devices with data parallelism"
                            )

272
                    self._device = torch.device(f"{accelerator.device}")
273
                    self.accelerator = accelerator
274

275
276
                    self._rank = self.accelerator.local_process_index
                    self._world_size = self.accelerator.num_processes
Nathan Habib's avatar
Nathan Habib committed
277
278
279
280
                else:
                    # if we aren't launching via accelerate, ditch
                    self._rank = 0
                    self._world_size = 1
281
282
283
284
285
286
287
        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
288

289
        self.custom_prefix_token_id = prefix_token_id
290
291
292
293
        if prefix_token_id is not None:
            eval_logger.info(
                f"Loglikelihood prefix token id used in evaluation: {self.prefix_token_id}"
            )
294

Nathan Habib's avatar
Nathan Habib committed
295
296
    def _get_accelerate_args(
        self,
297
        parallelize: Optional[bool] = None,
Nathan Habib's avatar
Nathan Habib committed
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
        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
344
                        k: v for k, v in max_memory_all_gpus.items()
Nathan Habib's avatar
Nathan Habib committed
345
                    }
Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
346
                else:
Nathan Habib's avatar
Nathan Habib committed
347
348
349
350
351
352
353
354
                    # 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
355
            args["device_map"] = "auto" if device_map is None else device_map
Nathan Habib's avatar
Nathan Habib committed
356
            eval_logger.info(
357
                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
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
            )

            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

383
384
385
386
387
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

388
389
390
391
392
393
394
395
    @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

396
397
398
399
400
    @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

401
402
403
404
405
406
407
408
409
    @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

410
411
    @property
    def max_length(self):
412
413
414
415
416
417
418
419
420
421
422
        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
423

424
    @property
Ethan Smith's avatar
Ethan Smith committed
425
    def max_gen_toks(self) -> int:
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
        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
444
445
446
447
    @property
    def tokenizer_name(self) -> str:
        return self.tokenizer.name_or_path.replace("/", "__")

448
449
    def _get_backend(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
450
        config: Union[transformers.PretrainedConfig, transformers.AutoConfig],
451
        backend: Literal["default", "causal", "seq2seq"] = "default",
452
453
454
455
        trust_remote_code: Optional[bool] = False,
    ) -> None:
        """
        Helper method during initialization.
456
        Determines the backend ("causal" (decoder-only) or "seq2seq" (encoder-decoder)) model type to be used.
457
        sets `self.AUTO_MODEL_CLASS` appropriately if not already set.
458
459
460

        **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!**
461
        """
462

463
464
465
466
467
        assert backend in ["default", "causal", "seq2seq"]

        if backend != "default":
            # if we've settled on non-default backend, use that manually
            if backend == "causal":
468
                self.backend = backend
469
            elif backend == "seq2seq":
470
                self.backend = backend
471
            eval_logger.info(
472
                f"Overrode HF model backend type, and using type '{self.backend}'"
473
474
475
476
477
478
479
480
481
482
            )
        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.
483
                self.backend = "seq2seq"
484
                eval_logger.debug(f"Using model type '{self.backend}'")
485
486
487
            elif (
                getattr(self.config, "model_type") in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
            ):
488
                self.backend = "causal"
489
                eval_logger.debug(f"Using model type '{self.backend}'")
490
491
492
493
494
            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."
495
                        "Setting backend to causal"
496
497
                    )
                # if model type is neither in HF transformers causal or seq2seq model registries
498
499
500
                # then we default to assuming AutoModelForCausalLM
                self.backend = "causal"
                eval_logger.info(
501
                    f"Model type cannot be determined. Using default model type '{self.backend}'"
502
                )
503

504
505
506
507
508
        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
509
510
511
512
513
514

    def _get_config(
        self,
        pretrained: str,
        revision: str = "main",
        trust_remote_code: bool = False,
515
        gguf_file: Optional[str] = None,
516
    ) -> None:
517
        """Return the model config for HuggingFace models"""
518
519
520
521
        self._config = transformers.AutoConfig.from_pretrained(
            pretrained,
            revision=revision,
            trust_remote_code=trust_remote_code,
522
            gguf_file=gguf_file,
523
524
525
526
527
528
529
530
531
532
533
534
        )

    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,
535
        gpus: Optional[int] = None,
536
537
538
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
        offload_folder: Optional[str] = "./offload",
539
        # PEFT, delta weights and quantization options
540
        peft: Optional[str] = None,
541
        delta: Optional[str] = None,
542
        autogptq: Optional[Union[bool, str]] = False,
543
        gptqmodel: Optional[bool] = False,
544
        gguf_file: Optional[str] = None,
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
        **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
561
562
563
564
565
566
567
568
        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,
569
            )
Nathan Habib's avatar
Nathan Habib committed
570
        )
571

572
        if not autogptq and not gptqmodel:
573
574
575
576
577
578
579
            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):
580
                        model_kwargs["bnb_4bit_compute_dtype"] = get_dtype(
581
582
                            model_kwargs["bnb_4bit_compute_dtype"]
                        )
Nathan Habib's avatar
Nathan Habib committed
583

584
585
586
            self._model = self.AUTO_MODEL_CLASS.from_pretrained(
                pretrained,
                revision=revision,
587
                torch_dtype=get_dtype(dtype),
588
                trust_remote_code=trust_remote_code,
589
                gguf_file=gguf_file,
590
591
592
                **model_kwargs,
            )
        else:
593
594
595
            if autogptq and gptqmodel:
                raise ValueError(
                    "Cannot use both 'autogptq' and 'gptqmodel' options at the same time."
596
597
                )

598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
            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
                )
629

630
631
632
633
634
        if peft and delta:
            raise ValueError(
                "Cannot use both 'peft' and 'delta' options at the same time."
            )

635
636
        if peft:
            if model_kwargs.get("load_in_4bit", None):
WoosungMyung's avatar
WoosungMyung committed
637
638
                if version.parse(PEFT_VERSION) < version.parse("0.4.0"):
                    raise AssertionError("load_in_4bit requires peft >= 0.4.0")
639
640
            if self._model.config.vocab_size != len(self.tokenizer):
                # resize model for LoRAs with added tokens
641
642
643
                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..."
                )
644
                self._model.resize_token_embeddings(len(self.tokenizer))
645
646
647
            self._model = PeftModel.from_pretrained(
                self._model, peft, revision=revision
            )
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
        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
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686

        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,
687
        gguf_file: Optional[str] = None,
688
689
690
691
692
693
694
    ) -> 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.
        """
695
696
697
698
699
700
701
702
703
704
        kwargs = {
            "revision": revision,
            "trust_remote_code": trust_remote_code,
        }

        # gguf format embeds tokenizer and is not compatible with hf tokenizer `use_fast` param
        if gguf_file is not None:
            kwargs["gguf_file"] = gguf_file
        else:
            kwargs["use_fast"] = use_fast_tokenizer
705
706
707
708

        if tokenizer:
            if isinstance(tokenizer, str):
                self.tokenizer = transformers.AutoTokenizer.from_pretrained(
709
                    tokenizer, **kwargs
710
711
712
713
714
715
716
717
718
719
720
721
722
723
                )
            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(
724
                model_name, **kwargs
725
726
727
            )
        return None

Ethan Smith's avatar
Ethan Smith committed
728
    def _detect_batch_size(self, requests=None, pos: int = 0):
Benjamin Fattori's avatar
Benjamin Fattori committed
729
730
731
732
733
        if requests:
            _, context_enc, continuation_enc = requests[pos]
            max_length = len(
                (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1]
            )
734
735
            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
736
737
        else:
            max_length = self.max_length
738
739
            max_context_enc = max_length
            max_cont_enc = max_length
lintangsutawika's avatar
lintangsutawika committed
740

Benjamin Fattori's avatar
Benjamin Fattori committed
741
742
743
        # 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):
744
            if self.backend == "seq2seq":
745
                length = max(max_context_enc, max_cont_enc)
lintangsutawika's avatar
lintangsutawika committed
746
747
748
                batched_conts = torch.ones(
                    (batch_size, length), device=self.device
                ).long()
749
750
                test_batch = torch.ones((batch_size, length), device=self.device).long()
                call_kwargs = {
lintangsutawika's avatar
lintangsutawika committed
751
752
753
                    "attn_mask": test_batch,
                    "labels": batched_conts,
                }
754
755
            else:
                call_kwargs = {}
lintangsutawika's avatar
lintangsutawika committed
756
757
758
                test_batch = torch.ones(
                    (batch_size, max_length), device=self.device
                ).long()
Benjamin Fattori's avatar
Benjamin Fattori committed
759
            for _ in range(5):
760
                out = F.log_softmax(self._model_call(test_batch, **call_kwargs), dim=-1)  # noqa: F841
lintangsutawika's avatar
lintangsutawika committed
761

Benjamin Fattori's avatar
Benjamin Fattori committed
762
763
            return batch_size

764
765
766
767
768
769
770
        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
771

772
773
774
775
776
777
778
        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)
779
            clear_torch_cache()
780
781
            return batch_size

782
        clear_torch_cache()
Benjamin Fattori's avatar
Benjamin Fattori committed
783
784
        return batch_size

baberabb's avatar
baberabb committed
785
786
787
    def tok_encode(
        self, string: str, left_truncate_len=None, add_special_tokens=None
    ) -> List[int]:
haileyschoelkopf's avatar
haileyschoelkopf committed
788
        """ """
Lintang Sutawika's avatar
Lintang Sutawika committed
789
790
791
792
793
        # 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
794
        if add_special_tokens is None:
795
            if self.backend == "causal":
Lintang Sutawika's avatar
Lintang Sutawika committed
796
797
798
799
800
801
                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}
802

Lintang Sutawika's avatar
Lintang Sutawika committed
803
        encoding = self.tokenizer.encode(string, **special_tokens_kwargs)
haileyschoelkopf's avatar
haileyschoelkopf committed
804

805
806
807
        # 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
808

809
810
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
811
    def tok_batch_encode(
lintangsutawika's avatar
lintangsutawika committed
812
813
        self,
        strings: List[str],
lintangsutawika's avatar
lintangsutawika committed
814
        padding_side: str = "left",
815
816
        left_truncate_len: int = None,
        truncation: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
817
    ) -> Tuple[torch.Tensor, torch.Tensor]:
haileyschoelkopf's avatar
haileyschoelkopf committed
818
819
820
821
        # 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
822
        add_special_tokens = {}
823
        if self.backend == "causal":
Lintang Sutawika's avatar
Lintang Sutawika committed
824
            add_special_tokens = {"add_special_tokens": False or self.add_bos_token}
haileyschoelkopf's avatar
haileyschoelkopf committed
825
826
827

        encoding = self.tokenizer(
            strings,
lintangsutawika's avatar
lintangsutawika committed
828
            truncation=truncation,
haileyschoelkopf's avatar
haileyschoelkopf committed
829
830
            padding="longest",
            return_tensors="pt",
Lintang Sutawika's avatar
Lintang Sutawika committed
831
            **add_special_tokens,
haileyschoelkopf's avatar
haileyschoelkopf committed
832
833
        )
        if left_truncate_len:
834
835
836
837
838
839
            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
840
841
842
843
844
845
846
847
            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
848
849
    def tok_decode(self, tokens, skip_special_tokens=True):
        return self.tokenizer.decode(tokens, skip_special_tokens=skip_special_tokens)
850
851
852

    def _model_call(self, inps, attn_mask=None, labels=None):
        """
haileyschoelkopf's avatar
haileyschoelkopf committed
853
        :param inps: torch.Tensor
854
855
856
857
858
859
860
861
862
863
864
865
866
            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():
867
868
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
869
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
haileyschoelkopf's avatar
haileyschoelkopf committed
870
871
872
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
873
874
875
876
877
            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
878
        # temperature = 0.0 if not set
879
880
881
        # 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
882
        generation_kwargs["temperature"] = generation_kwargs.get("temperature", 0.0)
883
        do_sample = generation_kwargs.get("do_sample", None)
884
885
886
887
888

        # 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
889
890
        if do_sample is False and generation_kwargs.get("temperature") == 0.0:
            generation_kwargs.pop("temperature")
891
892
        # build stopping criteria
        stopping_criteria = stop_sequences_criteria(
893
            self.tokenizer, stop, context.shape[1], context.shape[0]
894
        )
895
        return self.model.generate(
896
            input_ids=context,
897
898
            max_length=max_length,
            stopping_criteria=stopping_criteria,
899
            pad_token_id=self.tokenizer.pad_token_id,
900
901
902
            use_cache=True,
            **generation_kwargs,
        )
903

Baber Abbasi's avatar
Baber Abbasi committed
904
905
906
    def _select_cont_toks(
        self, logits: torch.Tensor, contlen: int = None, inplen: int = None
    ) -> torch.Tensor:
907
        if self.backend == "causal":
haileyschoelkopf's avatar
haileyschoelkopf committed
908
909
910
            assert (
                contlen and inplen
            ), "Must pass input len and cont. len to select scored logits for causal LM"
911
912
913
            # discard right-padding.
            # also discard the input/context tokens. we'll only score continuations.
            logits = logits[inplen - contlen : inplen]
914
        elif self.backend == "seq2seq":
haileyschoelkopf's avatar
haileyschoelkopf committed
915
916
917
918
            assert (
                contlen and not inplen
            ), "Selecting scored logits for Seq2SeqLM requires only cont. len"
            # only discard right-padding.
919
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
920
921
            logits = logits[:contlen]

922
923
        return logits

924
925
926
    def loglikelihood_rolling(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[float]:
Benjamin Fattori's avatar
Benjamin Fattori committed
927
928
929
930
931
932
933
934
        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

935
936
937
938
939
940
941
942
943
        # 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)),
            )
944
        ):
945
            rolling_token_windows: List[Tuple[List[int], List[int]]] = list(
946
947
948
949
                map(
                    utils.make_disjoint_window,
                    utils.get_rolling_token_windows(
                        token_list=self.tok_encode(string),
950
                        prefix_token=self.prefix_token_id,
951
952
953
954
955
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
956
957

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
958
            windows = [(None,) + x for x in rolling_token_windows]
959

960
961
962
            # Store windows with their request index
            all_windows.extend((req_idx, window) for window in windows)
            request_window_counts.append(len(windows))
963

964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
        # 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),
984
            )
985
986
            # Store results with their request indices
            all_nlls.extend(zip(batch_indices, batch_nlls))
987

988
989
990
        # Remove padding if necessary
        if (self.world_size > 1) and (pad_amnt > 0):
            all_nlls = all_nlls[:-pad_amnt]
991

992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
        # 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
            )
1007

1008
        return loglikelihoods
Zhiwei Zhuang's avatar
Zhiwei Zhuang committed
1009

1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
    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
1023
        self.batch_sizes[sched] = self._detect_batch_size(n_reordered_requests, pos)
1024
1025
        print(f"Determined largest batch size: {self.batch_sizes[sched]}")
        return self.batch_sizes[sched]
1026

Ethan Smith's avatar
Ethan Smith committed
1027
    def _loglikelihood_tokens(
baberabb's avatar
baberabb committed
1028
1029
1030
1031
1032
        self,
        requests: List[Tuple[Tuple[str, str], List[int], List[int]]],
        disable_tqdm: bool = False,
        override_bs: int = None,
    ) -> List[Tuple[float, bool]]:
1033
1034
1035
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

Baber Abbasi's avatar
Baber Abbasi committed
1036
        def _collate(req: Tuple[Tuple[str, str], List[int], List[int]]):
Baber Abbasi's avatar
Baber Abbasi committed
1037
            """Defines the key for the sorted method"""
1038
1039
1040
1041
1042
1043
1044
            # 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
1045
            toks = req[1] + req[2]
1046
1047
            return -len(toks), tuple(toks)

Baber Abbasi's avatar
Baber Abbasi committed
1048
1049
1050
        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
1051
            # 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
1052
1053
1054
1055
1056
1057
1058
1059
            # 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"
1060
            if self.backend == "causal" and self.logits_cache
Baber Abbasi's avatar
Baber Abbasi committed
1061
1062
1063
            else None,
            group_fn=_lookup_one_token_cont,
        )
Benjamin Fattori's avatar
Benjamin Fattori committed
1064
1065
1066

        # automatic (variable) batch size detection for vectorization
        # pull longest context sample from request
Baber Abbasi's avatar
Baber Abbasi committed
1067
1068
1069
        n_reordered_requests = len(re_ord)
        batch_size = (
            self.batch_size
1070
1071
1072
            if self.batch_size != "auto"
            else override_bs
            if override_bs is not None
Baber Abbasi's avatar
Baber Abbasi committed
1073
1074
1075
1076
            else 0
        )
        batch_fn = (
            self._batch_scheduler
1077
1078
1079
            if self.batch_size == "auto"
            and n_reordered_requests > 0
            and not override_bs
Baber Abbasi's avatar
Baber Abbasi committed
1080
            else None
1081
1082
        )

Baber Abbasi's avatar
Baber Abbasi committed
1083
        chunks = re_ord.get_batched(n=batch_size, batch_fn=batch_fn)
1084
1085
1086
1087
1088
        pbar = tqdm(
            total=len(requests),
            disable=(disable_tqdm or (self.rank != 0)),
            desc="Running loglikelihood requests",
        )
haileyschoelkopf's avatar
haileyschoelkopf committed
1089
        for chunk in chunks:
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
            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
1109
                # how this all works (illustrated on a causal decoder-only setup):
1110
1111
1112
1113
1114
1115
1116
                #          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
1117
                if self.backend == "causal":
1118
1119
1120
1121
1122
1123
1124
                    total_length = len(context_enc) + len(continuation_enc)
                    if total_length > self.max_length + 1:
                        eval_logger.warn(
                            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."
                        )
1125
1126
1127
                    inp = torch.tensor(
                        (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1],
                        dtype=torch.long,
1128
1129
                        device=self.device,
                    )
1130
                    (inplen,) = inp.shape
1131
                elif self.backend == "seq2seq":
1132
1133
1134
                    inp = torch.tensor(
                        (context_enc)[-self.max_length :],
                        dtype=torch.long,
haileyschoelkopf's avatar
haileyschoelkopf committed
1135
                        device=self.device,
1136
                    )
1137
                    (inplen,) = inp.shape
1138
1139
1140
1141

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

1142
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
1143
                        (continuation_enc)[-self.max_length :],
1144
1145
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
1146
                        dtype=torch.long,
1147
1148
                        device=self.device,
                    )
1149
1150
                    (contlen,) = cont.shape

1151
1152
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
1153
1154
1155
1156
1157
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
1158

haileyschoelkopf's avatar
haileyschoelkopf committed
1159
1160
1161
1162
1163
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
1164
1165
1166
1167

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
1168

1169
1170
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
1171
            if self.backend == "causal":
1172
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1173
1174
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
1175
            elif self.backend == "seq2seq":
1176
                # TODO: left-pad encoder inps and mask?
1177
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1178
1179
                    padding_len_inp, inps
                )  # [batch, padding_len_inp]
1180
                batched_conts = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1181
1182
                    padding_len_cont, conts
                )  # [batch, padding_len_cont]
1183
                batched_encoder_mask = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1184
1185
1186
1187
1188
1189
                    padding_len_inp, encoder_attns
                )  # [batch, padding_len_inp]
                call_kwargs = {
                    "attn_mask": batched_encoder_mask,
                    "labels": batched_conts,
                }
1190
1191
1192

            multi_logits = F.log_softmax(
                self._model_call(batched_inps, **call_kwargs), dim=-1
1193
            )  # [batch, padding_length (inp or cont), vocab]
1194

Baber Abbasi's avatar
Baber Abbasi committed
1195
            for (request_str, ctx_tokens, _), logits, inplen, cont_toks in zip(
1196
1197
1198
1199
                chunk, multi_logits, inplens, cont_toks_list
            ):
                # Slice to original seq length
                contlen = len(cont_toks)
haileyschoelkopf's avatar
haileyschoelkopf committed
1200
                # take only logits in the continuation
1201
                # (discard context toks if decoder-only ; discard right-padding)
1202
1203
                # 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
1204
                ctx_len = (
1205
                    inplen + (logits.shape[0] - padding_len_inp)
1206
                    if self.backend == "causal"
haileyschoelkopf's avatar
haileyschoelkopf committed
1207
1208
                    else None
                )
1209
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
1210
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
1211
1212
1213
1214

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

Baber Abbasi's avatar
Baber Abbasi committed
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
                # 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)

1242
1243
1244
1245
1246
1247
1248
                    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
1249
                    pbar.update(1)
haileyschoelkopf's avatar
haileyschoelkopf committed
1250
1251

        pbar.close()
haileyschoelkopf's avatar
haileyschoelkopf committed
1252

1253
1254
        return re_ord.get_original(res)

1255
1256
1257
    def generate_until(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1258
        res = []
1259

Baber Abbasi's avatar
Baber Abbasi committed
1260
        def _collate(req: Tuple[str, dict]):
Baber Abbasi's avatar
Baber Abbasi committed
1261
            """Defines the key for the sorted method"""
1262
1263
1264
1265
1266
1267
            # 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
1268
1269
            toks = self.tok_encode(req[0])
            return -len(toks), req[0]
1270

1271
1272
        pbar = tqdm(
            total=len(requests),
1273
            disable=(disable_tqdm or (self.rank != 0)),
1274
1275
            desc="Running generate_until requests",
        )
Baber Abbasi's avatar
Baber Abbasi committed
1276
        adaptive_batch_size = None
1277
1278
1279
1280
1281
1282
        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
1283
        # for each different set of kwargs, we execute all requests, by batch.
Baber Abbasi's avatar
Baber Abbasi committed
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
        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
        )
1296

Baber Abbasi's avatar
Baber Abbasi committed
1297
1298
1299
        # 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
1300
1301
1302
1303
1304
1305
1306
        # 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
1307
        chunks = re_ords.get_batched(n=batch_size, batch_fn=batch_fn)
1308
        eos = self.tok_decode(self.eot_token_id, skip_special_tokens=False)
Baber Abbasi's avatar
Baber Abbasi committed
1309
1310
1311
1312
1313
1314
1315
1316
        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
1317
1318
                # add EOS token to stop sequences
                until = handle_stop_sequences(kwargs.pop("until", None), eos=eos)
Baber Abbasi's avatar
Baber Abbasi committed
1319
1320
            else:
                raise ValueError(
Baber Abbasi's avatar
Baber Abbasi committed
1321
                    f"Expected `kwargs` to be of type `dict` but got {type(gen_kwargs)}"
1322
                )
Baber Abbasi's avatar
Baber Abbasi committed
1323
1324
1325
1326
1327
1328
            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")
1329
            if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
1330
1331
                # max len for inputs = max length, minus room to generate the max new tokens
                max_ctx_len = self.max_length - max_gen_toks
1332
1333
1334
                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})."
1335
            elif self.backend == "seq2seq":
Baber Abbasi's avatar
Baber Abbasi committed
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
                # 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)
1347

Baber Abbasi's avatar
Baber Abbasi committed
1348
1349
            if "max_length" not in kwargs:
                kwargs["max_length"] = context_enc.shape[1] + max_gen_toks
1350

Baber Abbasi's avatar
Baber Abbasi committed
1351
1352
1353
1354
1355
1356
1357
            # perform batched generation
            cont = self._model_generate(
                context=context_enc,
                attention_mask=attn_masks,
                stop=until,
                **kwargs,
            )
1358

Baber Abbasi's avatar
Baber Abbasi committed
1359
1360
1361
            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
1362
                if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
1363
                    cont_toks = cont_toks[context_enc.shape[1] :]
1364

Baber Abbasi's avatar
Baber Abbasi committed
1365
                s = self.tok_decode(cont_toks)
1366

Baber Abbasi's avatar
Baber Abbasi committed
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
                # 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)
1380

1381
        pbar.close()
1382

Baber Abbasi's avatar
Baber Abbasi committed
1383
        return res
1384

Baber Abbasi's avatar
Baber Abbasi committed
1385
1386
1387
    def apply_chat_template(
        self, chat_history: List[Dict[str, str]], add_generation_prompt: bool = True
    ) -> str:
KonradSzafer's avatar
KonradSzafer committed
1388
1389
1390
        """
        Method to apply a chat template to a list of chat history between user and model.
        """
1391
1392
        try:
            chat_templated = self.tokenizer.apply_chat_template(
Baber Abbasi's avatar
Baber Abbasi committed
1393
1394
1395
1396
                chat_history,
                tokenize=False,
                add_generation_prompt=add_generation_prompt,
                continue_final_message=not add_generation_prompt,
1397
1398
1399
1400
1401
1402
1403
            )
        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
1404
1405
1406
1407
                chat_history,
                tokenize=False,
                add_generation_prompt=add_generation_prompt,
                continue_final_message=not add_generation_prompt,
1408
1409
1410
            )

        return chat_templated
KonradSzafer's avatar
KonradSzafer committed
1411

1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
    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
1436
                eval_logger.debug(
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
                    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