huggingface.py 62.2 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,
Benjamin Fattori's avatar
Benjamin Fattori committed
79
80
        batch_size: Optional[Union[int, str]] = 1,
        max_batch_size: Optional[int] = 64,
81
        trust_remote_code: Optional[bool] = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
82
        use_fast_tokenizer: Optional[bool] = True,
83
        add_bos_token: Optional[bool] = False,
84
        prefix_token_id: Optional[int] = None,
85
        # arguments used for splitting a model across GPUs naively.
86
87
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
88
89
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
90
        offload_folder: Optional[Union[str, os.PathLike]] = "./offload",
91
        # PEFT, delta weights and quantization options
92
        peft: Optional[str] = None,
93
        delta: Optional[str] = None,
94
        autogptq: Optional[Union[bool, str]] = False,
95
        gptqmodel: Optional[bool] = False,
96
        gguf_file: Optional[str] = None,
97
        **kwargs,
Ethan Smith's avatar
Ethan Smith committed
98
    ) -> None:
99
        super().__init__()
100
101
102
103
        # 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."
104
            )
Baber Abbasi's avatar
Baber Abbasi committed
105
106
107
            assert not parallelize, (
                "`parallelize=True` is not compatible with passing pre-initialized model to `pretrained`"
            )
108
109
110
            self._model = pretrained
            self._device = self._model.device
            self._config = self._model.config
Baber Abbasi's avatar
Baber Abbasi committed
111
            gpus = 0
112

113
        else:
114
115
116
117
118
            assert isinstance(device, str)
            assert isinstance(pretrained, str)
            assert isinstance(batch_size, (int, str))

            gpus = torch.cuda.device_count()
Jeevan's avatar
Jeevan committed
119
120
            accelerator_kwargs = InitProcessGroupKwargs(timeout=timedelta(weeks=52))
            accelerator = Accelerator(kwargs_handlers=[accelerator_kwargs])
121
122
            if accelerator.num_processes > 1:
                self.accelerator = accelerator
123

124
125
126
            if "npu" in accelerator.device.type:
                gpus = torch.npu.device_count()

Nathan Habib's avatar
Nathan Habib committed
127
            # using one process with no model parallelism
128
129
130
131
            if not (parallelize or accelerator.num_processes > 1):
                # use user-passed device
                device_list = set(
                    ["cuda", "cpu"]
132
                    + [f"cuda:{i}" for i in range(gpus)]
133
                    + ["mps", "mps:0"]
134
                    + [f"npu:{i}" for i in range(gpus)]
135
                )
136
                if device and device in device_list:
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
                    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
153
            else:  # Parallelism managed by accelerate
154
155
156
157
158
                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
159
160
161
162
163
                self._device = (
                    self.accelerator.device
                    if hasattr(self, "accelerator")
                    else torch.device(device)
                )
164

Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
165
            revision = str(revision)  # cast to string if not already one
166

167
            self._get_config(
168
169
170
                pretrained,
                revision=revision,
                trust_remote_code=trust_remote_code,
171
                gguf_file=gguf_file,
172
                subfolder=subfolder,
173
174
            )

175
            # determine which of 'causal' and 'seq2seq' backends to use for HF models
176
177
178
        self._get_backend(
            config=self.config, backend=backend, trust_remote_code=trust_remote_code
        )
179

180
181
182
183
184
        # load tokenizer so we know tokenizer vocabulary size before loading model and PEFT
        self._create_tokenizer(
            pretrained,
            tokenizer,
            revision=revision,
185
            subfolder=subfolder,
186
187
            trust_remote_code=trust_remote_code,
            use_fast_tokenizer=use_fast_tokenizer,
188
            gguf_file=gguf_file,
189
            add_bos_token=add_bos_token,
190
191
        )

192
193
194
195
196
197
198
        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)

199
200
201
202
203
204
205
206
        # 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,
207
                gpus=gpus,
208
209
210
211
                max_memory_per_gpu=max_memory_per_gpu,
                max_cpu_memory=max_cpu_memory,
                offload_folder=offload_folder,
                peft=peft,
212
                delta=delta,
213
                autogptq=autogptq,
214
                gptqmodel=gptqmodel,
215
                gguf_file=gguf_file,
216
                quantization_config=quantization_config,
217
                subfolder=subfolder,
218
                **kwargs,
219
220
            )

221
        # access self._model through self.model property outside this method
222
223
224
        if isinstance(self.model, torch.nn.Module):
            self.model.eval()
            self.model.tie_weights()
haileyschoelkopf's avatar
haileyschoelkopf committed
225

lintangsutawika's avatar
lintangsutawika committed
226
        self.truncation = truncation
Baber Abbasi's avatar
Baber Abbasi committed
227
        self.logits_cache = logits_cache
228
        self.vocab_size = self.tokenizer.vocab_size
229
        # select (or create) a pad token to use
230
        self.tokenizer = configure_pad_token(self.tokenizer, model_config=self.config)
231

232
        self.add_bos_token = add_bos_token
233
        if "gemma" in getattr(self.config, "model_type", ""):
234
            self.add_bos_token = True
235
            eval_logger.info(
236
                f"Model type is '{self.config.model_type}', part of the Gemma family--a BOS token will be used as Gemma underperforms without it."
237
238
            )

239
        self._max_length = max_length
240
241
242
243
        self.pretrained = pretrained
        self.delta = delta
        self.peft = peft
        self.revision = revision
Benjamin Fattori's avatar
Benjamin Fattori committed
244
245
246
        self.batch_schedule = 1
        self.batch_sizes = {}
        self.max_batch_size = max_batch_size
247
248
249
        self.softmax_dtype = (
            get_dtype(softmax_dtype) if softmax_dtype is not None else None
        )
Benjamin Fattori's avatar
Benjamin Fattori committed
250
251
252
253
254
255
256

        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)
257

258
        if isinstance(pretrained, str):
Nathan Habib's avatar
Nathan Habib committed
259
260
261
262
263
264
265
266
267
268
269
270
            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."
                        )
271
272
            # multigpu data-parallel support when launched with accelerate
            if gpus > 1:
Nathan Habib's avatar
Nathan Habib committed
273
274
275
276
                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."
277
                        )
Nathan Habib's avatar
Nathan Habib committed
278
                    elif gpus > accelerator.num_processes:
279
280
281
282
283
284
                        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
285
286
287
288
289
                        if self.accelerator.is_local_main_process:
                            eval_logger.info(
                                f"Using {gpus} devices with data parallelism"
                            )

290
                    self._device = torch.device(f"{accelerator.device}")
291
                    self.accelerator = accelerator
292

293
294
                    self._rank = self.accelerator.local_process_index
                    self._world_size = self.accelerator.num_processes
Nathan Habib's avatar
Nathan Habib committed
295
296
297
298
                else:
                    # if we aren't launching via accelerate, ditch
                    self._rank = 0
                    self._world_size = 1
299
300
301
302
303
304
305
        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
306

307
        self.custom_prefix_token_id = prefix_token_id
308
309
310
311
        if prefix_token_id is not None:
            eval_logger.info(
                f"Loglikelihood prefix token id used in evaluation: {self.prefix_token_id}"
            )
312

Nathan Habib's avatar
Nathan Habib committed
313
314
    def _get_accelerate_args(
        self,
315
        parallelize: Optional[bool] = None,
Nathan Habib's avatar
Nathan Habib committed
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
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
        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
362
                        k: v for k, v in max_memory_all_gpus.items()
Nathan Habib's avatar
Nathan Habib committed
363
                    }
Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
364
                else:
Nathan Habib's avatar
Nathan Habib committed
365
366
367
368
369
370
371
372
                    # 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
373
            args["device_map"] = "auto" if device_map is None else device_map
Nathan Habib's avatar
Nathan Habib committed
374
            eval_logger.info(
375
                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
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
            )

            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

401
402
403
404
405
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

406
407
408
409
410
411
412
413
    @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

414
415
416
417
418
    @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

419
420
421
422
423
424
425
426
427
    @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

428
429
    @property
    def max_length(self):
430
431
432
433
434
435
436
437
438
439
440
        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
441

442
    @property
Ethan Smith's avatar
Ethan Smith committed
443
    def max_gen_toks(self) -> int:
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
        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
462
463
464
465
    @property
    def tokenizer_name(self) -> str:
        return self.tokenizer.name_or_path.replace("/", "__")

466
467
    def _get_backend(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
468
        config: Union[transformers.PretrainedConfig, transformers.AutoConfig],
469
        backend: Literal["default", "causal", "seq2seq"] = "default",
470
471
472
473
        trust_remote_code: Optional[bool] = False,
    ) -> None:
        """
        Helper method during initialization.
474
        Determines the backend ("causal" (decoder-only) or "seq2seq" (encoder-decoder)) model type to be used.
475
        sets `self.AUTO_MODEL_CLASS` appropriately if not already set.
476
477
478

        **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!**
479
        """
480

481
482
483
484
485
        assert backend in ["default", "causal", "seq2seq"]

        if backend != "default":
            # if we've settled on non-default backend, use that manually
            if backend == "causal":
486
                self.backend = backend
487
            elif backend == "seq2seq":
488
                self.backend = backend
489
            eval_logger.info(
490
                f"Overrode HF model backend type, and using type '{self.backend}'"
491
492
493
494
495
496
497
498
499
500
            )
        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.
501
                self.backend = "seq2seq"
502
                eval_logger.debug(f"Using model type '{self.backend}'")
503
504
505
            elif (
                getattr(self.config, "model_type") in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
            ):
506
                self.backend = "causal"
507
                eval_logger.debug(f"Using model type '{self.backend}'")
508
509
510
511
512
            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."
513
                        "Setting backend to causal"
514
515
                    )
                # if model type is neither in HF transformers causal or seq2seq model registries
516
517
518
                # then we default to assuming AutoModelForCausalLM
                self.backend = "causal"
                eval_logger.info(
519
                    f"Model type cannot be determined. Using default model type '{self.backend}'"
520
                )
521

522
523
524
525
526
        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
527
528
529
530
531
532

    def _get_config(
        self,
        pretrained: str,
        revision: str = "main",
        trust_remote_code: bool = False,
533
        gguf_file: Optional[str] = None,
534
        subfolder: str = "",
535
    ) -> None:
536
        """Return the model config for HuggingFace models"""
537
538
539
540
        self._config = transformers.AutoConfig.from_pretrained(
            pretrained,
            revision=revision,
            trust_remote_code=trust_remote_code,
541
            gguf_file=gguf_file,
542
            subfolder=subfolder,
543
544
545
546
547
548
549
550
551
552
553
554
        )

    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,
555
        gpus: Optional[int] = None,
556
557
558
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
        offload_folder: Optional[str] = "./offload",
559
        # PEFT, delta weights and quantization options
560
        peft: Optional[str] = None,
561
        delta: Optional[str] = None,
562
        autogptq: Optional[Union[bool, str]] = False,
563
        gptqmodel: Optional[bool] = False,
564
        gguf_file: Optional[str] = None,
565
        quantization_config: Optional["AutoQuantizationConfig"] = None,
566
        subfolder: str = "",
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
        **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
583
584
585
586
587
588
589
590
        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,
591
            )
Nathan Habib's avatar
Nathan Habib committed
592
        )
593

594
        if not autogptq and not gptqmodel:
595
            if model_kwargs.get("load_in_4bit", None):
Baber Abbasi's avatar
Baber Abbasi committed
596
597
598
                assert transformers.__version__ >= "4.30.0", (
                    "load_in_4bit requires transformers >= 4.30.0"
                )
599
600
601
            if transformers.__version__ >= "4.30.0":
                if model_kwargs.get("load_in_4bit", None):
                    if model_kwargs.get("bnb_4bit_compute_dtype", None):
602
                        model_kwargs["bnb_4bit_compute_dtype"] = get_dtype(
603
604
                            model_kwargs["bnb_4bit_compute_dtype"]
                        )
Nathan Habib's avatar
Nathan Habib committed
605

606
607
608
            self._model = self.AUTO_MODEL_CLASS.from_pretrained(
                pretrained,
                revision=revision,
609
                torch_dtype=get_dtype(dtype),
610
                trust_remote_code=trust_remote_code,
611
                gguf_file=gguf_file,
612
                quantization_config=quantization_config,
613
                subfolder=subfolder,
614
615
616
                **model_kwargs,
            )
        else:
617
618
619
            if autogptq and gptqmodel:
                raise ValueError(
                    "Cannot use both 'autogptq' and 'gptqmodel' options at the same time."
620
621
                )

622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
            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
                )
653

654
655
656
657
658
        if peft and delta:
            raise ValueError(
                "Cannot use both 'peft' and 'delta' options at the same time."
            )

659
        if peft:
660
661
662
            from peft import PeftModel
            from peft import __version__ as PEFT_VERSION

663
            if model_kwargs.get("load_in_4bit", None):
WoosungMyung's avatar
WoosungMyung committed
664
665
                if version.parse(PEFT_VERSION) < version.parse("0.4.0"):
                    raise AssertionError("load_in_4bit requires peft >= 0.4.0")
666
667
            if self._model.config.vocab_size != len(self.tokenizer):
                # resize model for LoRAs with added tokens
668
669
670
                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..."
                )
671
                self._model.resize_token_embeddings(len(self.tokenizer))
672
673
674
            self._model = PeftModel.from_pretrained(
                self._model, peft, revision=revision
            )
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
        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
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713

        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,
714
        gguf_file: Optional[str] = None,
715
        add_bos_token: Optional[bool] = False,
716
        subfolder: Optional[str] = "",
717
718
719
720
721
722
723
    ) -> 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.
        """
724
725
726
727
728
729
        kwargs = {
            "revision": revision,
            "trust_remote_code": trust_remote_code,
        }

        # gguf format embeds tokenizer and is not compatible with hf tokenizer `use_fast` param
730
        if not tokenizer and gguf_file is not None:
731
732
733
            kwargs["gguf_file"] = gguf_file
        else:
            kwargs["use_fast"] = use_fast_tokenizer
734

735
736
737
        if add_bos_token:
            kwargs["add_bos_token"] = True

738
739
740
        if subfolder:
            kwargs["subfolder"] = subfolder

741
742
743
        if tokenizer:
            if isinstance(tokenizer, str):
                self.tokenizer = transformers.AutoTokenizer.from_pretrained(
744
                    tokenizer, **kwargs
745
746
747
748
749
750
751
752
753
754
755
756
757
758
                )
            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(
759
                model_name, **kwargs
760
761
762
            )
        return None

Ethan Smith's avatar
Ethan Smith committed
763
    def _detect_batch_size(self, requests=None, pos: int = 0):
Benjamin Fattori's avatar
Benjamin Fattori committed
764
765
766
767
768
        if requests:
            _, context_enc, continuation_enc = requests[pos]
            max_length = len(
                (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1]
            )
769
770
            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
771
772
        else:
            max_length = self.max_length
773
774
            max_context_enc = max_length
            max_cont_enc = max_length
lintangsutawika's avatar
lintangsutawika committed
775

Benjamin Fattori's avatar
Benjamin Fattori committed
776
777
778
        # 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):
779
            if self.backend == "seq2seq":
780
                length = max(max_context_enc, max_cont_enc)
lintangsutawika's avatar
lintangsutawika committed
781
782
783
                batched_conts = torch.ones(
                    (batch_size, length), device=self.device
                ).long()
784
785
                test_batch = torch.ones((batch_size, length), device=self.device).long()
                call_kwargs = {
lintangsutawika's avatar
lintangsutawika committed
786
787
788
                    "attn_mask": test_batch,
                    "labels": batched_conts,
                }
789
790
            else:
                call_kwargs = {}
lintangsutawika's avatar
lintangsutawika committed
791
792
793
                test_batch = torch.ones(
                    (batch_size, max_length), device=self.device
                ).long()
Benjamin Fattori's avatar
Benjamin Fattori committed
794
            for _ in range(5):
795
796
797
798
799
                out = F.log_softmax(  # noqa: F841
                    self._model_call(test_batch, **call_kwargs),
                    dim=-1,
                    dtype=self.softmax_dtype,
                )
lintangsutawika's avatar
lintangsutawika committed
800

Benjamin Fattori's avatar
Benjamin Fattori committed
801
802
            return batch_size

803
804
805
806
807
808
809
        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
810

811
812
813
814
815
816
817
        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)
818
            clear_torch_cache()
819
820
            return batch_size

821
        clear_torch_cache()
Benjamin Fattori's avatar
Benjamin Fattori committed
822
823
        return batch_size

baberabb's avatar
baberabb committed
824
825
826
    def tok_encode(
        self, string: str, left_truncate_len=None, add_special_tokens=None
    ) -> List[int]:
haileyschoelkopf's avatar
haileyschoelkopf committed
827
        """ """
Lintang Sutawika's avatar
Lintang Sutawika committed
828
829
830
831
832
        # 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
833
        if add_special_tokens is None:
834
            if self.backend == "causal":
Lintang Sutawika's avatar
Lintang Sutawika committed
835
836
837
838
839
840
                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}
841

Lintang Sutawika's avatar
Lintang Sutawika committed
842
        encoding = self.tokenizer.encode(string, **special_tokens_kwargs)
haileyschoelkopf's avatar
haileyschoelkopf committed
843

844
845
846
        # 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
847

848
849
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
850
    def tok_batch_encode(
lintangsutawika's avatar
lintangsutawika committed
851
852
        self,
        strings: List[str],
lintangsutawika's avatar
lintangsutawika committed
853
        padding_side: str = "left",
854
855
        left_truncate_len: int = None,
        truncation: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
856
    ) -> Tuple[torch.Tensor, torch.Tensor]:
haileyschoelkopf's avatar
haileyschoelkopf committed
857
858
859
860
        # 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
861
        add_special_tokens = {}
862
        if self.backend == "causal":
Lintang Sutawika's avatar
Lintang Sutawika committed
863
            add_special_tokens = {"add_special_tokens": False or self.add_bos_token}
haileyschoelkopf's avatar
haileyschoelkopf committed
864
865
866

        encoding = self.tokenizer(
            strings,
lintangsutawika's avatar
lintangsutawika committed
867
            truncation=truncation,
haileyschoelkopf's avatar
haileyschoelkopf committed
868
869
            padding="longest",
            return_tensors="pt",
Lintang Sutawika's avatar
Lintang Sutawika committed
870
            **add_special_tokens,
haileyschoelkopf's avatar
haileyschoelkopf committed
871
872
        )
        if left_truncate_len:
873
874
875
876
877
878
            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
879
880
881
882
883
884
885
886
            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
887
888
    def tok_decode(self, tokens, skip_special_tokens=True):
        return self.tokenizer.decode(tokens, skip_special_tokens=skip_special_tokens)
889
890
891

    def _model_call(self, inps, attn_mask=None, labels=None):
        """
haileyschoelkopf's avatar
haileyschoelkopf committed
892
        :param inps: torch.Tensor
893
894
895
896
897
898
899
900
901
902
903
904
905
            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():
906
907
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
908
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
haileyschoelkopf's avatar
haileyschoelkopf committed
909
910
911
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
912
            else:
913
914
915
916
                assert self.AUTO_MODEL_CLASS in (
                    transformers.AutoModelForCausalLM,
                    transformers.AutoModelForVision2Seq,
                )
917
918
919
                return self.model(inps).logits

    def _model_generate(self, context, max_length, stop, **generation_kwargs):
Baber Abbasi's avatar
Baber Abbasi committed
920
        # temperature = 0.0 if not set
921
922
923
        # 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
924
        generation_kwargs["temperature"] = generation_kwargs.get("temperature", 0.0)
925
        do_sample = generation_kwargs.get("do_sample", None)
926
927
928
929
930

        # 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
931
932
        if do_sample is False and generation_kwargs.get("temperature") == 0.0:
            generation_kwargs.pop("temperature")
933
934
        # build stopping criteria
        stopping_criteria = stop_sequences_criteria(
935
            self.tokenizer, stop, context.shape[1], context.shape[0]
936
        )
937
        return self.model.generate(
938
            input_ids=context,
939
940
            max_length=max_length,
            stopping_criteria=stopping_criteria,
941
            pad_token_id=self.tokenizer.pad_token_id,
942
943
944
            use_cache=True,
            **generation_kwargs,
        )
945

Baber Abbasi's avatar
Baber Abbasi committed
946
947
948
    def _select_cont_toks(
        self, logits: torch.Tensor, contlen: int = None, inplen: int = None
    ) -> torch.Tensor:
949
        if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
950
951
952
            assert contlen and inplen, (
                "Must pass input len and cont. len to select scored logits for causal LM"
            )
953
954
955
            # discard right-padding.
            # also discard the input/context tokens. we'll only score continuations.
            logits = logits[inplen - contlen : inplen]
956
        elif self.backend == "seq2seq":
Baber Abbasi's avatar
Baber Abbasi committed
957
958
959
            assert contlen and not inplen, (
                "Selecting scored logits for Seq2SeqLM requires only cont. len"
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
960
            # only discard right-padding.
961
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
962
963
            logits = logits[:contlen]

964
965
        return logits

966
967
968
    def loglikelihood_rolling(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[float]:
Benjamin Fattori's avatar
Benjamin Fattori committed
969
970
971
972
973
974
975
976
        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

977
978
979
980
981
982
983
984
985
        # 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)),
            )
986
        ):
987
            rolling_token_windows: List[Tuple[List[int], List[int]]] = list(
988
989
990
991
                map(
                    utils.make_disjoint_window,
                    utils.get_rolling_token_windows(
                        token_list=self.tok_encode(string),
992
                        prefix_token=self.prefix_token_id,
993
994
995
996
997
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
998
999

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
1000
            windows = [(None,) + x for x in rolling_token_windows]
1001

1002
1003
1004
            # Store windows with their request index
            all_windows.extend((req_idx, window) for window in windows)
            request_window_counts.append(len(windows))
1005

1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
        # 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),
1026
            )
1027
1028
            # Store results with their request indices
            all_nlls.extend(zip(batch_indices, batch_nlls))
1029

1030
1031
1032
        # Remove padding if necessary
        if (self.world_size > 1) and (pad_amnt > 0):
            all_nlls = all_nlls[:-pad_amnt]
1033

1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
        # 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
            )
1049

1050
        return loglikelihoods
Zhiwei Zhuang's avatar
Zhiwei Zhuang committed
1051

1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
    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
1065
        self.batch_sizes[sched] = self._detect_batch_size(n_reordered_requests, pos)
1066
1067
        print(f"Determined largest batch size: {self.batch_sizes[sched]}")
        return self.batch_sizes[sched]
1068

Ethan Smith's avatar
Ethan Smith committed
1069
    def _loglikelihood_tokens(
baberabb's avatar
baberabb committed
1070
1071
1072
1073
1074
        self,
        requests: List[Tuple[Tuple[str, str], List[int], List[int]]],
        disable_tqdm: bool = False,
        override_bs: int = None,
    ) -> List[Tuple[float, bool]]:
1075
1076
1077
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

Baber Abbasi's avatar
Baber Abbasi committed
1078
        def _collate(req: Tuple[Tuple[str, str], List[int], List[int]]):
Baber Abbasi's avatar
Baber Abbasi committed
1079
            """Defines the key for the sorted method"""
1080
1081
1082
1083
1084
1085
1086
            # 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
1087
            toks = req[1] + req[2]
1088
1089
            return -len(toks), tuple(toks)

Baber Abbasi's avatar
Baber Abbasi committed
1090
1091
1092
        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
1093
            # 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
1094
1095
1096
1097
1098
1099
1100
1101
            # 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"
1102
            if self.backend == "causal" and self.logits_cache
Baber Abbasi's avatar
Baber Abbasi committed
1103
1104
1105
            else None,
            group_fn=_lookup_one_token_cont,
        )
Benjamin Fattori's avatar
Benjamin Fattori committed
1106
1107
1108

        # automatic (variable) batch size detection for vectorization
        # pull longest context sample from request
Baber Abbasi's avatar
Baber Abbasi committed
1109
1110
1111
        n_reordered_requests = len(re_ord)
        batch_size = (
            self.batch_size
1112
1113
1114
            if self.batch_size != "auto"
            else override_bs
            if override_bs is not None
Baber Abbasi's avatar
Baber Abbasi committed
1115
1116
1117
1118
            else 0
        )
        batch_fn = (
            self._batch_scheduler
1119
1120
1121
            if self.batch_size == "auto"
            and n_reordered_requests > 0
            and not override_bs
Baber Abbasi's avatar
Baber Abbasi committed
1122
            else None
1123
1124
        )

Baber Abbasi's avatar
Baber Abbasi committed
1125
        chunks = re_ord.get_batched(n=batch_size, batch_fn=batch_fn)
1126
1127
1128
1129
1130
        pbar = tqdm(
            total=len(requests),
            disable=(disable_tqdm or (self.rank != 0)),
            desc="Running loglikelihood requests",
        )
haileyschoelkopf's avatar
haileyschoelkopf committed
1131
        for chunk in chunks:
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
            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
1151
                # how this all works (illustrated on a causal decoder-only setup):
1152
1153
1154
1155
1156
1157
1158
                #          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
1159
                if self.backend == "causal":
1160
1161
                    total_length = len(context_enc) + len(continuation_enc)
                    if total_length > self.max_length + 1:
1162
                        eval_logger.warning(
1163
1164
1165
1166
                            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."
                        )
1167
1168
1169
                    inp = torch.tensor(
                        (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1],
                        dtype=torch.long,
1170
1171
                        device=self.device,
                    )
1172
                    (inplen,) = inp.shape
1173
                elif self.backend == "seq2seq":
1174
1175
1176
                    inp = torch.tensor(
                        (context_enc)[-self.max_length :],
                        dtype=torch.long,
haileyschoelkopf's avatar
haileyschoelkopf committed
1177
                        device=self.device,
1178
                    )
1179
                    (inplen,) = inp.shape
1180
1181
1182
1183

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

1184
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
1185
                        (continuation_enc)[-self.max_length :],
1186
1187
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
1188
                        dtype=torch.long,
1189
1190
                        device=self.device,
                    )
1191
1192
                    (contlen,) = cont.shape

1193
1194
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
1195
1196
1197
1198
1199
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
1200

haileyschoelkopf's avatar
haileyschoelkopf committed
1201
1202
1203
1204
1205
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
1206
1207
1208
1209

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
1210

1211
1212
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
1213
            if self.backend == "causal":
1214
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1215
1216
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
1217
            elif self.backend == "seq2seq":
1218
                # TODO: left-pad encoder inps and mask?
1219
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1220
1221
                    padding_len_inp, inps
                )  # [batch, padding_len_inp]
1222
                batched_conts = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1223
1224
                    padding_len_cont, conts
                )  # [batch, padding_len_cont]
1225
                batched_encoder_mask = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1226
1227
1228
1229
1230
1231
                    padding_len_inp, encoder_attns
                )  # [batch, padding_len_inp]
                call_kwargs = {
                    "attn_mask": batched_encoder_mask,
                    "labels": batched_conts,
                }
1232
1233

            multi_logits = F.log_softmax(
1234
1235
1236
                self._model_call(batched_inps, **call_kwargs),
                dim=-1,
                dtype=self.softmax_dtype,
1237
            )  # [batch, padding_length (inp or cont), vocab]
1238

Baber Abbasi's avatar
Baber Abbasi committed
1239
            for (request_str, ctx_tokens, _), logits, inplen, cont_toks in zip(
1240
1241
1242
1243
                chunk, multi_logits, inplens, cont_toks_list
            ):
                # Slice to original seq length
                contlen = len(cont_toks)
haileyschoelkopf's avatar
haileyschoelkopf committed
1244
                # take only logits in the continuation
1245
                # (discard context toks if decoder-only ; discard right-padding)
1246
1247
                # 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
1248
                ctx_len = (
1249
                    inplen + (logits.shape[0] - padding_len_inp)
1250
                    if self.backend == "causal"
haileyschoelkopf's avatar
haileyschoelkopf committed
1251
1252
                    else None
                )
1253
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
1254
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
1255
1256
1257
1258

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

Baber Abbasi's avatar
Baber Abbasi committed
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
                # 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]
1273
1274
1275
1276
1277
1278
                    # 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
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290

                    # 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)

1291
1292
1293
1294
1295
1296
1297
                    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
1298
                    pbar.update(1)
haileyschoelkopf's avatar
haileyschoelkopf committed
1299
1300

        pbar.close()
haileyschoelkopf's avatar
haileyschoelkopf committed
1301

1302
1303
        return re_ord.get_original(res)

1304
1305
1306
    def generate_until(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1307
        res = []
1308

Baber Abbasi's avatar
Baber Abbasi committed
1309
        def _collate(req: Tuple[str, dict]):
Baber Abbasi's avatar
Baber Abbasi committed
1310
            """Defines the key for the sorted method"""
1311
1312
1313
1314
1315
1316
            # 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
1317
1318
            toks = self.tok_encode(req[0])
            return -len(toks), req[0]
1319

1320
1321
        pbar = tqdm(
            total=len(requests),
1322
            disable=(disable_tqdm or (self.rank != 0)),
1323
1324
            desc="Running generate_until requests",
        )
Baber Abbasi's avatar
Baber Abbasi committed
1325
        adaptive_batch_size = None
1326
1327
1328
1329
1330
1331
        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
1332
        # for each different set of kwargs, we execute all requests, by batch.
Baber Abbasi's avatar
Baber Abbasi committed
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
        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
        )
1345

Baber Abbasi's avatar
Baber Abbasi committed
1346
1347
1348
        # 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
1349
1350
1351
1352
1353
1354
1355
        # 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
1356
        chunks = re_ords.get_batched(n=batch_size, batch_fn=batch_fn)
1357
        eos = self.tok_decode(self.eot_token_id, skip_special_tokens=False)
Baber Abbasi's avatar
Baber Abbasi committed
1358
1359
1360
1361
1362
1363
1364
1365
        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
1366
1367
                # add EOS token to stop sequences
                until = handle_stop_sequences(kwargs.pop("until", None), eos=eos)
Baber Abbasi's avatar
Baber Abbasi committed
1368
1369
            else:
                raise ValueError(
Baber Abbasi's avatar
Baber Abbasi committed
1370
                    f"Expected `kwargs` to be of type `dict` but got {type(gen_kwargs)}"
1371
                )
Baber Abbasi's avatar
Baber Abbasi committed
1372
1373
1374
1375
1376
1377
            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")
1378
            if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
1379
1380
                # 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
1381
1382
1383
                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})."
                )
1384
            elif self.backend == "seq2seq":
Baber Abbasi's avatar
Baber Abbasi committed
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
                # 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)
1396

Baber Abbasi's avatar
Baber Abbasi committed
1397
1398
            if "max_length" not in kwargs:
                kwargs["max_length"] = context_enc.shape[1] + max_gen_toks
1399

Baber Abbasi's avatar
Baber Abbasi committed
1400
1401
1402
1403
1404
1405
1406
            # perform batched generation
            cont = self._model_generate(
                context=context_enc,
                attention_mask=attn_masks,
                stop=until,
                **kwargs,
            )
1407

Baber Abbasi's avatar
Baber Abbasi committed
1408
1409
1410
            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
1411
                if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
1412
                    cont_toks = cont_toks[context_enc.shape[1] :]
1413

Baber Abbasi's avatar
Baber Abbasi committed
1414
                s = self.tok_decode(cont_toks)
1415

Baber Abbasi's avatar
Baber Abbasi committed
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
                # 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)
1429

1430
        pbar.close()
1431

Baber Abbasi's avatar
Baber Abbasi committed
1432
        return res
1433

Baber Abbasi's avatar
Baber Abbasi committed
1434
1435
1436
    def apply_chat_template(
        self, chat_history: List[Dict[str, str]], add_generation_prompt: bool = True
    ) -> str:
KonradSzafer's avatar
KonradSzafer committed
1437
1438
1439
        """
        Method to apply a chat template to a list of chat history between user and model.
        """
1440
1441
        try:
            chat_templated = self.tokenizer.apply_chat_template(
Baber Abbasi's avatar
Baber Abbasi committed
1442
1443
1444
1445
                chat_history,
                tokenize=False,
                add_generation_prompt=add_generation_prompt,
                continue_final_message=not add_generation_prompt,
1446
1447
1448
1449
1450
1451
1452
            )
        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
1453
1454
1455
1456
                chat_history,
                tokenize=False,
                add_generation_prompt=add_generation_prompt,
                continue_final_message=not add_generation_prompt,
1457
1458
1459
            )

        return chat_templated
KonradSzafer's avatar
KonradSzafer committed
1460

1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
    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
1485
                eval_logger.debug(
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
                    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