huggingface.py 61.6 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 Any, 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
21
22
from packaging import version
from peft import PeftModel
from peft import __version__ as PEFT_VERSION
from tqdm import tqdm
23
24
25
26
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
)
27
28

from lm_eval import utils
baberabb's avatar
baberabb committed
29
from lm_eval.api.instance import Instance
30
from lm_eval.api.model import TemplateLM
31
from lm_eval.api.registry import register_model
32
33
34
from lm_eval.models.utils import (
    Collator,
    clear_torch_cache,
35
    configure_pad_token,
36
    get_dtype,
37
    handle_stop_sequences,
38
39
40
    pad_and_concat,
    stop_sequences_criteria,
)
41

42

Lintang Sutawika's avatar
Lintang Sutawika committed
43
eval_logger = logging.getLogger(__name__)
44

lintangsutawika's avatar
lintangsutawika committed
45

46
@register_model("hf-auto", "hf", "huggingface")
47
class HFLM(TemplateLM):
48
49
50
51
52
53
54
    """
    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.
    """

55
    AUTO_MODEL_CLASS = None
56
    _DEFAULT_MAX_LENGTH = 2048
haileyschoelkopf's avatar
haileyschoelkopf committed
57

58
59
    def __init__(
        self,
60
        pretrained: Union[str, transformers.PreTrainedModel],
61
        backend: Literal["default", "causal", "seq2seq"] = "default",
Baber Abbasi's avatar
Baber Abbasi committed
62
        # override whether the model should be treated as decoder-only (causal) or encoder-decoder (seq2seq)
63
64
        revision: Optional[str] = "main",
        subfolder: Optional[str] = None,
65
66
67
68
69
70
71
        tokenizer: Optional[
            Union[
                str,
                transformers.PreTrainedTokenizer,
                transformers.PreTrainedTokenizerFast,
            ]
        ] = None,
lintangsutawika's avatar
lintangsutawika committed
72
        truncation: Optional[bool] = False,
Baber Abbasi's avatar
Baber Abbasi committed
73
        logits_cache: bool = True,
74
75
        max_length: Optional[int] = None,
        device: Optional[str] = "cuda",
76
        dtype: Optional[Union[str, torch.dtype]] = "auto",
77
        softmax_dtype: Optional[Union[str, torch.dtype]] = None,
Benjamin Fattori's avatar
Benjamin Fattori committed
78
79
        batch_size: Optional[Union[int, str]] = 1,
        max_batch_size: Optional[int] = 64,
80
        trust_remote_code: Optional[bool] = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
81
        use_fast_tokenizer: Optional[bool] = True,
82
        add_bos_token: Optional[bool] = False,
83
        prefix_token_id: Optional[int] = None,
84
        # arguments used for splitting a model across GPUs naively.
85
86
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
87
88
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
89
        offload_folder: Optional[Union[str, os.PathLike]] = "./offload",
90
        # PEFT, delta weights and quantization options
91
        peft: Optional[str] = None,
92
        delta: Optional[str] = None,
93
        autogptq: Optional[Union[bool, str]] = False,
94
        gptqmodel: Optional[bool] = False,
95
        gguf_file: Optional[str] = None,
96
        **kwargs,
Ethan Smith's avatar
Ethan Smith committed
97
    ) -> None:
98
        super().__init__()
99
100
101
102
        # 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."
103
            )
Baber Abbasi's avatar
Baber Abbasi committed
104
105
106
            assert not parallelize, (
                "`parallelize=True` is not compatible with passing pre-initialized model to `pretrained`"
            )
107
108
109
            self._model = pretrained
            self._device = self._model.device
            self._config = self._model.config
Baber Abbasi's avatar
Baber Abbasi committed
110
            gpus = 0
111

112
        else:
113
114
115
116
117
            assert isinstance(device, str)
            assert isinstance(pretrained, str)
            assert isinstance(batch_size, (int, str))

            gpus = torch.cuda.device_count()
Jeevan's avatar
Jeevan committed
118
119
            accelerator_kwargs = InitProcessGroupKwargs(timeout=timedelta(weeks=52))
            accelerator = Accelerator(kwargs_handlers=[accelerator_kwargs])
120
121
            if accelerator.num_processes > 1:
                self.accelerator = accelerator
122

123
124
125
            if "npu" in accelerator.device.type:
                gpus = torch.npu.device_count()

Nathan Habib's avatar
Nathan Habib committed
126
            # using one process with no model parallelism
127
128
129
130
            if not (parallelize or accelerator.num_processes > 1):
                # use user-passed device
                device_list = set(
                    ["cuda", "cpu"]
131
                    + [f"cuda:{i}" for i in range(gpus)]
132
                    + ["mps", "mps:0"]
133
                    + [f"npu:{i}" for i in range(gpus)]
134
                )
135
                if device and device in device_list:
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
                    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
152
            else:  # Parallelism managed by accelerate
153
154
155
156
157
                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
158
159
160
161
162
                self._device = (
                    self.accelerator.device
                    if hasattr(self, "accelerator")
                    else torch.device(device)
                )
163

Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
164
            revision = str(revision)  # cast to string if not already one
165
166
            # TODO: update this to be less of a hack once subfolder is fixed in HF
            revision = revision + ("/" + subfolder if subfolder is not None else "")
167

168
            self._get_config(
169
170
171
                pretrained,
                revision=revision,
                trust_remote_code=trust_remote_code,
172
                gguf_file=gguf_file,
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
185
186
        # 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,
187
            gguf_file=gguf_file,
188
            add_bos_token=add_bos_token,
189
190
        )

191
192
193
194
195
196
197
198
        # 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,
199
                gpus=gpus,
200
201
202
203
                max_memory_per_gpu=max_memory_per_gpu,
                max_cpu_memory=max_cpu_memory,
                offload_folder=offload_folder,
                peft=peft,
204
                delta=delta,
205
                autogptq=autogptq,
206
                gptqmodel=gptqmodel,
207
                gguf_file=gguf_file,
208
                quantization_config=getattr(self.config, "quantization_config", None),
209
                **kwargs,
210
211
            )

212
        # access self._model through self.model property outside this method
213
214
215
        if isinstance(self.model, torch.nn.Module):
            self.model.eval()
            self.model.tie_weights()
haileyschoelkopf's avatar
haileyschoelkopf committed
216

lintangsutawika's avatar
lintangsutawika committed
217
        self.truncation = truncation
Baber Abbasi's avatar
Baber Abbasi committed
218
        self.logits_cache = logits_cache
219
        self.vocab_size = self.tokenizer.vocab_size
220
        # select (or create) a pad token to use
221
        self.tokenizer = configure_pad_token(self.tokenizer, model_config=self.config)
222

223
        self.add_bos_token = add_bos_token
224
        if "gemma" in getattr(self.config, "model_type", ""):
225
            self.add_bos_token = True
226
            eval_logger.info(
227
                f"Model type is '{self.config.model_type}', part of the Gemma family--a BOS token will be used as Gemma underperforms without it."
228
229
            )

230
        self._max_length = max_length
231
232
233
234
        self.pretrained = pretrained
        self.delta = delta
        self.peft = peft
        self.revision = revision
Benjamin Fattori's avatar
Benjamin Fattori committed
235
236
237
        self.batch_schedule = 1
        self.batch_sizes = {}
        self.max_batch_size = max_batch_size
238
239
240
        self.softmax_dtype = (
            get_dtype(softmax_dtype) if softmax_dtype is not None else None
        )
Benjamin Fattori's avatar
Benjamin Fattori committed
241
242
243
244
245
246
247

        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)
248

249
        if isinstance(pretrained, str):
Nathan Habib's avatar
Nathan Habib committed
250
251
252
253
254
255
256
257
258
259
260
261
            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."
                        )
262
263
            # multigpu data-parallel support when launched with accelerate
            if gpus > 1:
Nathan Habib's avatar
Nathan Habib committed
264
265
266
267
                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."
268
                        )
Nathan Habib's avatar
Nathan Habib committed
269
                    elif gpus > accelerator.num_processes:
270
271
272
273
274
275
                        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
276
277
278
279
280
                        if self.accelerator.is_local_main_process:
                            eval_logger.info(
                                f"Using {gpus} devices with data parallelism"
                            )

281
                    self._device = torch.device(f"{accelerator.device}")
282
                    self.accelerator = accelerator
283

284
285
                    self._rank = self.accelerator.local_process_index
                    self._world_size = self.accelerator.num_processes
Nathan Habib's avatar
Nathan Habib committed
286
287
288
289
                else:
                    # if we aren't launching via accelerate, ditch
                    self._rank = 0
                    self._world_size = 1
290
291
292
293
294
295
296
        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
297

298
        self.custom_prefix_token_id = prefix_token_id
299
300
301
302
        if prefix_token_id is not None:
            eval_logger.info(
                f"Loglikelihood prefix token id used in evaluation: {self.prefix_token_id}"
            )
303

Nathan Habib's avatar
Nathan Habib committed
304
305
    def _get_accelerate_args(
        self,
306
        parallelize: Optional[bool] = None,
Nathan Habib's avatar
Nathan Habib committed
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
344
345
346
347
348
349
350
351
352
        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
353
                        k: v for k, v in max_memory_all_gpus.items()
Nathan Habib's avatar
Nathan Habib committed
354
                    }
Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
355
                else:
Nathan Habib's avatar
Nathan Habib committed
356
357
358
359
360
361
362
363
                    # 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
364
            args["device_map"] = "auto" if device_map is None else device_map
Nathan Habib's avatar
Nathan Habib committed
365
            eval_logger.info(
366
                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
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
            )

            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

392
393
394
395
396
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

397
398
399
400
401
402
403
404
    @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

405
406
407
408
409
    @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

410
411
412
413
414
415
416
417
418
    @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

419
420
    @property
    def max_length(self):
421
422
423
424
425
426
427
428
429
430
431
        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
432

433
    @property
Ethan Smith's avatar
Ethan Smith committed
434
    def max_gen_toks(self) -> int:
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
        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
453
454
455
456
    @property
    def tokenizer_name(self) -> str:
        return self.tokenizer.name_or_path.replace("/", "__")

457
458
    def _get_backend(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
459
        config: Union[transformers.PretrainedConfig, transformers.AutoConfig],
460
        backend: Literal["default", "causal", "seq2seq"] = "default",
461
462
463
464
        trust_remote_code: Optional[bool] = False,
    ) -> None:
        """
        Helper method during initialization.
465
        Determines the backend ("causal" (decoder-only) or "seq2seq" (encoder-decoder)) model type to be used.
466
        sets `self.AUTO_MODEL_CLASS` appropriately if not already set.
467
468
469

        **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!**
470
        """
471

472
473
474
475
476
        assert backend in ["default", "causal", "seq2seq"]

        if backend != "default":
            # if we've settled on non-default backend, use that manually
            if backend == "causal":
477
                self.backend = backend
478
            elif backend == "seq2seq":
479
                self.backend = backend
480
            eval_logger.info(
481
                f"Overrode HF model backend type, and using type '{self.backend}'"
482
483
484
485
486
487
488
489
490
491
            )
        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.
492
                self.backend = "seq2seq"
493
                eval_logger.debug(f"Using model type '{self.backend}'")
494
495
496
            elif (
                getattr(self.config, "model_type") in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
            ):
497
                self.backend = "causal"
498
                eval_logger.debug(f"Using model type '{self.backend}'")
499
500
501
502
503
            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."
504
                        "Setting backend to causal"
505
506
                    )
                # if model type is neither in HF transformers causal or seq2seq model registries
507
508
509
                # then we default to assuming AutoModelForCausalLM
                self.backend = "causal"
                eval_logger.info(
510
                    f"Model type cannot be determined. Using default model type '{self.backend}'"
511
                )
512

513
514
515
516
517
        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
518
519
520
521
522
523

    def _get_config(
        self,
        pretrained: str,
        revision: str = "main",
        trust_remote_code: bool = False,
524
        gguf_file: Optional[str] = None,
525
    ) -> None:
526
        """Return the model config for HuggingFace models"""
527
528
529
530
        self._config = transformers.AutoConfig.from_pretrained(
            pretrained,
            revision=revision,
            trust_remote_code=trust_remote_code,
531
            gguf_file=gguf_file,
532
533
534
535
536
537
538
539
540
541
542
543
        )

    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,
544
        gpus: Optional[int] = None,
545
546
547
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
        offload_folder: Optional[str] = "./offload",
548
        # PEFT, delta weights and quantization options
549
        peft: Optional[str] = None,
550
        delta: Optional[str] = None,
551
        autogptq: Optional[Union[bool, str]] = False,
552
        gptqmodel: Optional[bool] = False,
553
        gguf_file: Optional[str] = None,
554
        quantization_config: Optional[Dict[str, Any]] = None,
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
        **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
571
572
573
574
575
576
577
578
        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,
579
            )
Nathan Habib's avatar
Nathan Habib committed
580
        )
581

582
        if not autogptq and not gptqmodel:
583
            if model_kwargs.get("load_in_4bit", None):
Baber Abbasi's avatar
Baber Abbasi committed
584
585
586
                assert transformers.__version__ >= "4.30.0", (
                    "load_in_4bit requires transformers >= 4.30.0"
                )
587
588
589
            if transformers.__version__ >= "4.30.0":
                if model_kwargs.get("load_in_4bit", None):
                    if model_kwargs.get("bnb_4bit_compute_dtype", None):
590
                        model_kwargs["bnb_4bit_compute_dtype"] = get_dtype(
591
592
                            model_kwargs["bnb_4bit_compute_dtype"]
                        )
Nathan Habib's avatar
Nathan Habib committed
593

594
595
596
            self._model = self.AUTO_MODEL_CLASS.from_pretrained(
                pretrained,
                revision=revision,
597
                torch_dtype=get_dtype(dtype),
598
                trust_remote_code=trust_remote_code,
599
                gguf_file=gguf_file,
600
                quantization_config=quantization_config,
601
602
603
                **model_kwargs,
            )
        else:
604
605
606
            if autogptq and gptqmodel:
                raise ValueError(
                    "Cannot use both 'autogptq' and 'gptqmodel' options at the same time."
607
608
                )

609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
            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
                )
640

641
642
643
644
645
        if peft and delta:
            raise ValueError(
                "Cannot use both 'peft' and 'delta' options at the same time."
            )

646
647
        if peft:
            if model_kwargs.get("load_in_4bit", None):
WoosungMyung's avatar
WoosungMyung committed
648
649
                if version.parse(PEFT_VERSION) < version.parse("0.4.0"):
                    raise AssertionError("load_in_4bit requires peft >= 0.4.0")
650
651
            if self._model.config.vocab_size != len(self.tokenizer):
                # resize model for LoRAs with added tokens
652
653
654
                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..."
                )
655
                self._model.resize_token_embeddings(len(self.tokenizer))
656
657
658
            self._model = PeftModel.from_pretrained(
                self._model, peft, revision=revision
            )
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
        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
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697

        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,
698
        gguf_file: Optional[str] = None,
699
        add_bos_token: Optional[bool] = False,
700
701
702
703
704
705
706
    ) -> 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.
        """
707
708
709
710
711
712
713
714
715
716
        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
717

718
719
720
        if add_bos_token:
            kwargs["add_bos_token"] = True

721
722
723
        if tokenizer:
            if isinstance(tokenizer, str):
                self.tokenizer = transformers.AutoTokenizer.from_pretrained(
724
                    tokenizer, **kwargs
725
726
727
728
729
730
731
732
733
734
735
736
737
738
                )
            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(
739
                model_name, **kwargs
740
741
742
            )
        return None

Ethan Smith's avatar
Ethan Smith committed
743
    def _detect_batch_size(self, requests=None, pos: int = 0):
Benjamin Fattori's avatar
Benjamin Fattori committed
744
745
746
747
748
        if requests:
            _, context_enc, continuation_enc = requests[pos]
            max_length = len(
                (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1]
            )
749
750
            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
751
752
        else:
            max_length = self.max_length
753
754
            max_context_enc = max_length
            max_cont_enc = max_length
lintangsutawika's avatar
lintangsutawika committed
755

Benjamin Fattori's avatar
Benjamin Fattori committed
756
757
758
        # 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):
759
            if self.backend == "seq2seq":
760
                length = max(max_context_enc, max_cont_enc)
lintangsutawika's avatar
lintangsutawika committed
761
762
763
                batched_conts = torch.ones(
                    (batch_size, length), device=self.device
                ).long()
764
765
                test_batch = torch.ones((batch_size, length), device=self.device).long()
                call_kwargs = {
lintangsutawika's avatar
lintangsutawika committed
766
767
768
                    "attn_mask": test_batch,
                    "labels": batched_conts,
                }
769
770
            else:
                call_kwargs = {}
lintangsutawika's avatar
lintangsutawika committed
771
772
773
                test_batch = torch.ones(
                    (batch_size, max_length), device=self.device
                ).long()
Benjamin Fattori's avatar
Benjamin Fattori committed
774
            for _ in range(5):
775
776
777
778
779
                out = F.log_softmax(  # noqa: F841
                    self._model_call(test_batch, **call_kwargs),
                    dim=-1,
                    dtype=self.softmax_dtype,
                )
lintangsutawika's avatar
lintangsutawika committed
780

Benjamin Fattori's avatar
Benjamin Fattori committed
781
782
            return batch_size

783
784
785
786
787
788
789
        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
790

791
792
793
794
795
796
797
        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)
798
            clear_torch_cache()
799
800
            return batch_size

801
        clear_torch_cache()
Benjamin Fattori's avatar
Benjamin Fattori committed
802
803
        return batch_size

baberabb's avatar
baberabb committed
804
805
806
    def tok_encode(
        self, string: str, left_truncate_len=None, add_special_tokens=None
    ) -> List[int]:
haileyschoelkopf's avatar
haileyschoelkopf committed
807
        """ """
Lintang Sutawika's avatar
Lintang Sutawika committed
808
809
810
811
812
        # 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
813
        if add_special_tokens is None:
814
            if self.backend == "causal":
Lintang Sutawika's avatar
Lintang Sutawika committed
815
816
817
818
819
820
                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}
821

Lintang Sutawika's avatar
Lintang Sutawika committed
822
        encoding = self.tokenizer.encode(string, **special_tokens_kwargs)
haileyschoelkopf's avatar
haileyschoelkopf committed
823

824
825
826
        # 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
827

828
829
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
830
    def tok_batch_encode(
lintangsutawika's avatar
lintangsutawika committed
831
832
        self,
        strings: List[str],
lintangsutawika's avatar
lintangsutawika committed
833
        padding_side: str = "left",
834
835
        left_truncate_len: int = None,
        truncation: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
836
    ) -> Tuple[torch.Tensor, torch.Tensor]:
haileyschoelkopf's avatar
haileyschoelkopf committed
837
838
839
840
        # 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
841
        add_special_tokens = {}
842
        if self.backend == "causal":
Lintang Sutawika's avatar
Lintang Sutawika committed
843
            add_special_tokens = {"add_special_tokens": False or self.add_bos_token}
haileyschoelkopf's avatar
haileyschoelkopf committed
844
845
846

        encoding = self.tokenizer(
            strings,
lintangsutawika's avatar
lintangsutawika committed
847
            truncation=truncation,
haileyschoelkopf's avatar
haileyschoelkopf committed
848
849
            padding="longest",
            return_tensors="pt",
Lintang Sutawika's avatar
Lintang Sutawika committed
850
            **add_special_tokens,
haileyschoelkopf's avatar
haileyschoelkopf committed
851
852
        )
        if left_truncate_len:
853
854
855
856
857
858
            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
859
860
861
862
863
864
865
866
            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
867
868
    def tok_decode(self, tokens, skip_special_tokens=True):
        return self.tokenizer.decode(tokens, skip_special_tokens=skip_special_tokens)
869
870
871

    def _model_call(self, inps, attn_mask=None, labels=None):
        """
haileyschoelkopf's avatar
haileyschoelkopf committed
872
        :param inps: torch.Tensor
873
874
875
876
877
878
879
880
881
882
883
884
885
            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():
886
887
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
888
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
haileyschoelkopf's avatar
haileyschoelkopf committed
889
890
891
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
892
            else:
893
894
895
896
                assert self.AUTO_MODEL_CLASS in (
                    transformers.AutoModelForCausalLM,
                    transformers.AutoModelForVision2Seq,
                )
897
898
899
                return self.model(inps).logits

    def _model_generate(self, context, max_length, stop, **generation_kwargs):
Baber Abbasi's avatar
Baber Abbasi committed
900
        # temperature = 0.0 if not set
901
902
903
        # 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
904
        generation_kwargs["temperature"] = generation_kwargs.get("temperature", 0.0)
905
        do_sample = generation_kwargs.get("do_sample", None)
906
907
908
909
910

        # 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
911
912
        if do_sample is False and generation_kwargs.get("temperature") == 0.0:
            generation_kwargs.pop("temperature")
913
914
        # build stopping criteria
        stopping_criteria = stop_sequences_criteria(
915
            self.tokenizer, stop, context.shape[1], context.shape[0]
916
        )
917
        return self.model.generate(
918
            input_ids=context,
919
920
            max_length=max_length,
            stopping_criteria=stopping_criteria,
921
            pad_token_id=self.tokenizer.pad_token_id,
922
923
924
            use_cache=True,
            **generation_kwargs,
        )
925

Baber Abbasi's avatar
Baber Abbasi committed
926
927
928
    def _select_cont_toks(
        self, logits: torch.Tensor, contlen: int = None, inplen: int = None
    ) -> torch.Tensor:
929
        if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
930
931
932
            assert contlen and inplen, (
                "Must pass input len and cont. len to select scored logits for causal LM"
            )
933
934
935
            # discard right-padding.
            # also discard the input/context tokens. we'll only score continuations.
            logits = logits[inplen - contlen : inplen]
936
        elif self.backend == "seq2seq":
Baber Abbasi's avatar
Baber Abbasi committed
937
938
939
            assert contlen and not inplen, (
                "Selecting scored logits for Seq2SeqLM requires only cont. len"
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
940
            # only discard right-padding.
941
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
942
943
            logits = logits[:contlen]

944
945
        return logits

946
947
948
    def loglikelihood_rolling(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[float]:
Benjamin Fattori's avatar
Benjamin Fattori committed
949
950
951
952
953
954
955
956
        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

957
958
959
960
961
962
963
964
965
        # 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)),
            )
966
        ):
967
            rolling_token_windows: List[Tuple[List[int], List[int]]] = list(
968
969
970
971
                map(
                    utils.make_disjoint_window,
                    utils.get_rolling_token_windows(
                        token_list=self.tok_encode(string),
972
                        prefix_token=self.prefix_token_id,
973
974
975
976
977
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
978
979

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
980
            windows = [(None,) + x for x in rolling_token_windows]
981

982
983
984
            # Store windows with their request index
            all_windows.extend((req_idx, window) for window in windows)
            request_window_counts.append(len(windows))
985

986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
        # 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),
1006
            )
1007
1008
            # Store results with their request indices
            all_nlls.extend(zip(batch_indices, batch_nlls))
1009

1010
1011
1012
        # Remove padding if necessary
        if (self.world_size > 1) and (pad_amnt > 0):
            all_nlls = all_nlls[:-pad_amnt]
1013

1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
        # 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
            )
1029

1030
        return loglikelihoods
Zhiwei Zhuang's avatar
Zhiwei Zhuang committed
1031

1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
    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
1045
        self.batch_sizes[sched] = self._detect_batch_size(n_reordered_requests, pos)
1046
1047
        print(f"Determined largest batch size: {self.batch_sizes[sched]}")
        return self.batch_sizes[sched]
1048

Ethan Smith's avatar
Ethan Smith committed
1049
    def _loglikelihood_tokens(
baberabb's avatar
baberabb committed
1050
1051
1052
1053
1054
        self,
        requests: List[Tuple[Tuple[str, str], List[int], List[int]]],
        disable_tqdm: bool = False,
        override_bs: int = None,
    ) -> List[Tuple[float, bool]]:
1055
1056
1057
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

Baber Abbasi's avatar
Baber Abbasi committed
1058
        def _collate(req: Tuple[Tuple[str, str], List[int], List[int]]):
Baber Abbasi's avatar
Baber Abbasi committed
1059
            """Defines the key for the sorted method"""
1060
1061
1062
1063
1064
1065
1066
            # 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
1067
            toks = req[1] + req[2]
1068
1069
            return -len(toks), tuple(toks)

Baber Abbasi's avatar
Baber Abbasi committed
1070
1071
1072
        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
1073
            # 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
1074
1075
1076
1077
1078
1079
1080
1081
            # 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"
1082
            if self.backend == "causal" and self.logits_cache
Baber Abbasi's avatar
Baber Abbasi committed
1083
1084
1085
            else None,
            group_fn=_lookup_one_token_cont,
        )
Benjamin Fattori's avatar
Benjamin Fattori committed
1086
1087
1088

        # automatic (variable) batch size detection for vectorization
        # pull longest context sample from request
Baber Abbasi's avatar
Baber Abbasi committed
1089
1090
1091
        n_reordered_requests = len(re_ord)
        batch_size = (
            self.batch_size
1092
1093
1094
            if self.batch_size != "auto"
            else override_bs
            if override_bs is not None
Baber Abbasi's avatar
Baber Abbasi committed
1095
1096
1097
1098
            else 0
        )
        batch_fn = (
            self._batch_scheduler
1099
1100
1101
            if self.batch_size == "auto"
            and n_reordered_requests > 0
            and not override_bs
Baber Abbasi's avatar
Baber Abbasi committed
1102
            else None
1103
1104
        )

Baber Abbasi's avatar
Baber Abbasi committed
1105
        chunks = re_ord.get_batched(n=batch_size, batch_fn=batch_fn)
1106
1107
1108
1109
1110
        pbar = tqdm(
            total=len(requests),
            disable=(disable_tqdm or (self.rank != 0)),
            desc="Running loglikelihood requests",
        )
haileyschoelkopf's avatar
haileyschoelkopf committed
1111
        for chunk in chunks:
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
            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
1131
                # how this all works (illustrated on a causal decoder-only setup):
1132
1133
1134
1135
1136
1137
1138
                #          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
1139
                if self.backend == "causal":
1140
1141
                    total_length = len(context_enc) + len(continuation_enc)
                    if total_length > self.max_length + 1:
1142
                        eval_logger.warning(
1143
1144
1145
1146
                            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."
                        )
1147
1148
1149
                    inp = torch.tensor(
                        (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1],
                        dtype=torch.long,
1150
1151
                        device=self.device,
                    )
1152
                    (inplen,) = inp.shape
1153
                elif self.backend == "seq2seq":
1154
1155
1156
                    inp = torch.tensor(
                        (context_enc)[-self.max_length :],
                        dtype=torch.long,
haileyschoelkopf's avatar
haileyschoelkopf committed
1157
                        device=self.device,
1158
                    )
1159
                    (inplen,) = inp.shape
1160
1161
1162
1163

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

1164
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
1165
                        (continuation_enc)[-self.max_length :],
1166
1167
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
1168
                        dtype=torch.long,
1169
1170
                        device=self.device,
                    )
1171
1172
                    (contlen,) = cont.shape

1173
1174
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
1175
1176
1177
1178
1179
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
1180

haileyschoelkopf's avatar
haileyschoelkopf committed
1181
1182
1183
1184
1185
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
1186
1187
1188
1189

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
1190

1191
1192
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
1193
            if self.backend == "causal":
1194
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1195
1196
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
1197
            elif self.backend == "seq2seq":
1198
                # TODO: left-pad encoder inps and mask?
1199
                batched_inps = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1200
1201
                    padding_len_inp, inps
                )  # [batch, padding_len_inp]
1202
                batched_conts = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1203
1204
                    padding_len_cont, conts
                )  # [batch, padding_len_cont]
1205
                batched_encoder_mask = pad_and_concat(
haileyschoelkopf's avatar
haileyschoelkopf committed
1206
1207
1208
1209
1210
1211
                    padding_len_inp, encoder_attns
                )  # [batch, padding_len_inp]
                call_kwargs = {
                    "attn_mask": batched_encoder_mask,
                    "labels": batched_conts,
                }
1212
1213

            multi_logits = F.log_softmax(
1214
1215
1216
                self._model_call(batched_inps, **call_kwargs),
                dim=-1,
                dtype=self.softmax_dtype,
1217
            )  # [batch, padding_length (inp or cont), vocab]
1218

Baber Abbasi's avatar
Baber Abbasi committed
1219
            for (request_str, ctx_tokens, _), logits, inplen, cont_toks in zip(
1220
1221
1222
1223
                chunk, multi_logits, inplens, cont_toks_list
            ):
                # Slice to original seq length
                contlen = len(cont_toks)
haileyschoelkopf's avatar
haileyschoelkopf committed
1224
                # take only logits in the continuation
1225
                # (discard context toks if decoder-only ; discard right-padding)
1226
1227
                # 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
1228
                ctx_len = (
1229
                    inplen + (logits.shape[0] - padding_len_inp)
1230
                    if self.backend == "causal"
haileyschoelkopf's avatar
haileyschoelkopf committed
1231
1232
                    else None
                )
1233
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
1234
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
1235
1236
1237
1238

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

Baber Abbasi's avatar
Baber Abbasi committed
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
                # 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]
1253
1254
1255
1256
1257
1258
                    # 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
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270

                    # 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)

1271
1272
1273
1274
1275
1276
1277
                    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
1278
                    pbar.update(1)
haileyschoelkopf's avatar
haileyschoelkopf committed
1279
1280

        pbar.close()
haileyschoelkopf's avatar
haileyschoelkopf committed
1281

1282
1283
        return re_ord.get_original(res)

1284
1285
1286
    def generate_until(
        self, requests: List[Instance], disable_tqdm: bool = False
    ) -> List[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1287
        res = []
1288

Baber Abbasi's avatar
Baber Abbasi committed
1289
        def _collate(req: Tuple[str, dict]):
Baber Abbasi's avatar
Baber Abbasi committed
1290
            """Defines the key for the sorted method"""
1291
1292
1293
1294
1295
1296
            # 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
1297
1298
            toks = self.tok_encode(req[0])
            return -len(toks), req[0]
1299

1300
1301
        pbar = tqdm(
            total=len(requests),
1302
            disable=(disable_tqdm or (self.rank != 0)),
1303
1304
            desc="Running generate_until requests",
        )
Baber Abbasi's avatar
Baber Abbasi committed
1305
        adaptive_batch_size = None
1306
1307
1308
1309
1310
1311
        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
1312
        # for each different set of kwargs, we execute all requests, by batch.
Baber Abbasi's avatar
Baber Abbasi committed
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
        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
        )
1325

Baber Abbasi's avatar
Baber Abbasi committed
1326
1327
1328
        # 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
1329
1330
1331
1332
1333
1334
1335
        # 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
1336
        chunks = re_ords.get_batched(n=batch_size, batch_fn=batch_fn)
1337
        eos = self.tok_decode(self.eot_token_id, skip_special_tokens=False)
Baber Abbasi's avatar
Baber Abbasi committed
1338
1339
1340
1341
1342
1343
1344
1345
        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
1346
1347
                # add EOS token to stop sequences
                until = handle_stop_sequences(kwargs.pop("until", None), eos=eos)
Baber Abbasi's avatar
Baber Abbasi committed
1348
1349
            else:
                raise ValueError(
Baber Abbasi's avatar
Baber Abbasi committed
1350
                    f"Expected `kwargs` to be of type `dict` but got {type(gen_kwargs)}"
1351
                )
Baber Abbasi's avatar
Baber Abbasi committed
1352
1353
1354
1355
1356
1357
            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")
1358
            if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
1359
1360
                # 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
1361
1362
1363
                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})."
                )
1364
            elif self.backend == "seq2seq":
Baber Abbasi's avatar
Baber Abbasi committed
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
                # 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)
1376

Baber Abbasi's avatar
Baber Abbasi committed
1377
1378
            if "max_length" not in kwargs:
                kwargs["max_length"] = context_enc.shape[1] + max_gen_toks
1379

Baber Abbasi's avatar
Baber Abbasi committed
1380
1381
1382
1383
1384
1385
1386
            # perform batched generation
            cont = self._model_generate(
                context=context_enc,
                attention_mask=attn_masks,
                stop=until,
                **kwargs,
            )
1387

Baber Abbasi's avatar
Baber Abbasi committed
1388
1389
1390
            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
1391
                if self.backend == "causal":
Baber Abbasi's avatar
Baber Abbasi committed
1392
                    cont_toks = cont_toks[context_enc.shape[1] :]
1393

Baber Abbasi's avatar
Baber Abbasi committed
1394
                s = self.tok_decode(cont_toks)
1395

Baber Abbasi's avatar
Baber Abbasi committed
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
                # 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)
1409

1410
        pbar.close()
1411

Baber Abbasi's avatar
Baber Abbasi committed
1412
        return res
1413

Baber Abbasi's avatar
Baber Abbasi committed
1414
1415
1416
    def apply_chat_template(
        self, chat_history: List[Dict[str, str]], add_generation_prompt: bool = True
    ) -> str:
KonradSzafer's avatar
KonradSzafer committed
1417
1418
1419
        """
        Method to apply a chat template to a list of chat history between user and model.
        """
1420
1421
        try:
            chat_templated = self.tokenizer.apply_chat_template(
Baber Abbasi's avatar
Baber Abbasi committed
1422
1423
1424
1425
                chat_history,
                tokenize=False,
                add_generation_prompt=add_generation_prompt,
                continue_final_message=not add_generation_prompt,
1426
1427
1428
1429
1430
1431
1432
            )
        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
1433
1434
1435
1436
                chat_history,
                tokenize=False,
                add_generation_prompt=add_generation_prompt,
                continue_final_message=not add_generation_prompt,
1437
1438
1439
            )

        return chat_templated
KonradSzafer's avatar
KonradSzafer committed
1440

1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
    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
1465
                eval_logger.debug(
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
                    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