huggingface.py 37.7 KB
Newer Older
1
2
import torch
import transformers
3
4
5
6
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
)
7
from peft import __version__ as PEFT_VERSION, PeftModel
8
9

import copy
10
from collections import defaultdict
11
from tqdm import tqdm
12
from pathlib import Path
13
14
15
16
17
18
19
20
21
22

import torch.nn.functional as F

from lm_eval import utils
from lm_eval.logger import eval_logger
from lm_eval.api.model import LM
from lm_eval.api.registry import register_model

from lm_eval.utils import MultiTokenEOSCriteria, stop_sequences_criteria

Benjamin Fattori's avatar
Benjamin Fattori committed
23
from accelerate import Accelerator, find_executable_batch_size
24
from typing import List, Optional, Union
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49


def _get_accelerate_args(
    device_map_option: Optional[str] = "auto",
    max_memory_per_gpu: Optional[Union[int, str]] = None,
    max_cpu_memory: Optional[Union[int, str]] = None,
    offload_folder: Optional[str] = "./offload",
) -> dict:
    """Returns the kwargs needed to apply `accelerate` in `AutoModel.from_pretrained`."""
    max_memory = {}
    if max_memory_per_gpu is not None:
        max_memory_per_gpu_map = {
            device_idx: max_memory_per_gpu
            for device_idx in range(torch.cuda.device_count())
        }
        max_memory.update(max_memory_per_gpu_map)
    if max_cpu_memory is not None:
        max_memory["cpu"] = max_cpu_memory

    args = {}
    if max_memory:
        args["max_memory"] = max_memory
    args["device_map"] = device_map_option
    args["offload_folder"] = offload_folder
    return args
50
51


52
@register_model("hf-auto", "hf", "huggingface")
53
class HFLM(LM):
54
55
56
57
58
59
60
    """
    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.
    """

61
    AUTO_MODEL_CLASS = None
62
    _DEFAULT_MAX_LENGTH = 2048
haileyschoelkopf's avatar
haileyschoelkopf committed
63

64
65
    def __init__(
        self,
66
67
68
69
70
71
        pretrained: Optional[str] = "gpt2",
        revision: Optional[str] = "main",
        subfolder: Optional[str] = None,
        tokenizer: Optional[str] = None,
        max_length: Optional[int] = None,
        device: Optional[str] = "cuda",
72
        dtype: Optional[Union[str, torch.dtype]] = "auto",
Benjamin Fattori's avatar
Benjamin Fattori committed
73
74
        batch_size: Optional[Union[int, str]] = 1,
        max_batch_size: Optional[int] = 64,
75
76
        low_cpu_mem_usage: Optional[bool] = True,
        trust_remote_code: Optional[bool] = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
77
        use_fast_tokenizer: Optional[bool] = True,
78
        # arguments used for splitting a model across GPUs naively.
79
80
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
81
82
83
84
        device_map_option: Optional[str] = "auto",
        max_memory_per_gpu: Optional[Union[int, str]] = None,
        max_cpu_memory: Optional[Union[int, str]] = None,
        offload_folder: Optional[str] = "./offload",
85
86
87
88
89
90
91
92
        # PEFT and quantization options
        peft: Optional[str] = None,
        load_in_8bit: Optional[bool] = False,
        load_in_4bit: Optional[bool] = False,
        bnb_4bit_quant_type: Optional[str] = None,
        bnb_4bit_compute_dtype: Optional[Union[str, torch.dtype]] = None,
        gptq: Optional[Union[bool, str]] = False,
        gptq_use_triton: Optional[bool] = False,
93
94
95
96
97
    ):
        super().__init__()

        assert isinstance(device, str)
        assert isinstance(pretrained, str)
Benjamin Fattori's avatar
Benjamin Fattori committed
98
        assert isinstance(batch_size, (int, str))
99
100

        gpus = torch.cuda.device_count()
101
        accelerator = Accelerator()
haileyschoelkopf's avatar
haileyschoelkopf committed
102

103
        if not (parallelize or accelerator.num_processes > 1):
104
            # use user-passed device
105
            device_list = set(
baberabb's avatar
add mps  
baberabb committed
106
                ["cuda", "cpu", "mps"]
107
108
                + [f"cuda:{i}" for i in range(torch.cuda.device_count())]
            )
109
            if device:
110
                if device not in device_list:
111
112
113
                    device = int(device)
                self._device = torch.device(device)
                eval_logger.info(f"Using device '{device}'")
114
115
                if device == "mps":
                    eval_logger.info(
baberabb's avatar
baberabb committed
116
                        "MPS is still in beta and only supports float32; setting dtype to float32."
117
                    )
118
119
120
121
122
123
124
125
            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")
                )
126
        else:
127
128
129
130
            if device != "cuda":
                eval_logger.info(
                    f"Using `accelerate launch` or `parallelize=True`, device '{device}' will be overridden when placing model."
                )
131
            # TODO: include in warning that `load_in_8bit` etc. affect this too
132
133
134
            self._device = device

        model_kwargs = {}
135
        if parallelize:
136
137
138
139
140
141
            model_kwargs = _get_accelerate_args(
                device_map_option,
                max_memory_per_gpu,
                max_cpu_memory,
                offload_folder,
            )
142
143
144
145
146
147
148

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

        self._config = transformers.AutoConfig.from_pretrained(
            pretrained,
            revision=revision,
149
            trust_remote_code=trust_remote_code,
150
151
152
153
        )

        if getattr(self._config, "model_type") in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
            self.AUTO_MODEL_CLASS = transformers.AutoModelForCausalLM
154
155
156
157
158
159
160
161
162
163
164
165
        elif (
            not getattr(self._config, "model_type")
            in MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES
        ):
            if not trust_remote_code:
                eval_logger.warning(
                    "HF model type is neither marked as CausalLM or Seq2SeqLM. \
                This is expected if your model requires `trust_remote_code=True` but may be an error otherwise."
                )
            # if model type is neither in HF transformers causal or seq2seq model registries
            # then we default to AutoModelForCausalLM
            self.AUTO_MODEL_CLASS = transformers.AutoModelForCausalLM
166
        else:
haileyschoelkopf's avatar
haileyschoelkopf committed
167
            self.AUTO_MODEL_CLASS = transformers.AutoModelForSeq2SeqLM
168

haileyschoelkopf's avatar
haileyschoelkopf committed
169
170
171
172
        assert self.AUTO_MODEL_CLASS in [
            transformers.AutoModelForCausalLM,
            transformers.AutoModelForSeq2SeqLM,
        ]
173

174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
        if not gptq:
            if load_in_4bit:
                assert (
                    transformers.__version__ >= "4.30.0"
                ), "load_in_4bit requires transformers >= 4.30.0"
            if transformers.__version__ >= "4.30.0":
                model_kwargs["load_in_4bit"] = load_in_4bit
                if load_in_4bit:
                    if bnb_4bit_quant_type:
                        model_kwargs["bnb_4bit_quant_type"] = bnb_4bit_quant_type
                    if bnb_4bit_compute_dtype:
                        model_kwargs["bnb_4bit_compute_dtype"] = utils.get_dtype(
                            bnb_4bit_compute_dtype
                        )
            self._model = self.AUTO_MODEL_CLASS.from_pretrained(
                pretrained,
                revision=revision,
                torch_dtype=utils.get_dtype(dtype),
                low_cpu_mem_usage=low_cpu_mem_usage,
                trust_remote_code=trust_remote_code,
                load_in_8bit=load_in_8bit,
                **model_kwargs,
            )
        else:
gk's avatar
gk committed
198
199
200
201
202
203
204
            try:
                from auto_gptq import AutoGPTQForCausalLM
            except ModuleNotFoundError:
                raise Exception(
                    "Tried to load auto_gptq, but auto-gptq is not installed ",
                    "please install auto-gptq via pip install lm-eval[gptq] or pip install -e .[gptq]",
                )
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223

            self._model = AutoGPTQForCausalLM.from_quantized(
                pretrained,
                model_basename=None if gptq is True else Path(gptq).stem,
                low_cpu_mem_usage=low_cpu_mem_usage,
                trust_remote_code=trust_remote_code,
                use_safetensors=True if gptq is True else gptq.endswith(".safetensors"),
                use_triton=gptq_use_triton,
                warmup_triton=gptq_use_triton,
                **model_kwargs,
            )

        if peft:
            if load_in_4bit:
                assert PEFT_VERSION >= "0.4.0", "load_in_4bit requires peft >= 0.4.0"
            self._model = PeftModel.from_pretrained(
                self._model, peft, revision=revision
            )

224
        # forever after, access self._model through self.model property
225
        self.model.eval()
226
227
228
        self.model.tie_weights()
        if gpus <= 1 and not parallelize:
            # place model onto device, if not using HF Accelerate in any form
229
230
231
232
233
234
            try:
                self.model.to(self.device)
            except ValueError:
                eval_logger.info(
                    "Failed to place model onto specified device. This may be because the model is quantized via `bitsandbytes`. If the desired GPU is being used, this message is safe to ignore."
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
235

236
237
238
        self.tokenizer = transformers.AutoTokenizer.from_pretrained(
            pretrained if tokenizer is None else tokenizer,
            revision=revision,
239
            trust_remote_code=trust_remote_code,
haileyschoelkopf's avatar
haileyschoelkopf committed
240
            use_fast=use_fast_tokenizer,
241
242
243
        )

        self.vocab_size = self.tokenizer.vocab_size
haileyschoelkopf's avatar
haileyschoelkopf committed
244
        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
245

246
247
        self._max_length = max_length

Benjamin Fattori's avatar
Benjamin Fattori committed
248
249
250
251
252
253
254
255
256
257
        self.batch_schedule = 1
        self.batch_sizes = {}
        self.max_batch_size = max_batch_size

        if str(batch_size).startswith("auto"):
            batch_size = batch_size.split(":")
            self.batch_size_per_gpu = batch_size[0]
            self.batch_schedule = float(batch_size[1]) if len(batch_size) > 1 else 1
        else:
            self.batch_size_per_gpu = int(batch_size)
258
259
260
261
262
263
264
265
266
267
268

        # multigpu data-parallel support when launched with accelerate
        if gpus > 1:
            if parallelize:
                if accelerator.num_processes > 1:
                    raise RuntimeError(
                        "Attempted to use both a HF Accelerate `device_map` and to launch via `accelerate launch`. If this is the case, please either remove `parallelize=True` from --model_args or launch outside of the Accelerate launcher."
                    )
                else:
                    pass
            elif gpus > accelerator.num_processes:
269
                # TODO: make sure there's still never an edge case where we unintentionally default to CPU
270
271
272
273
274
275
276
277
278
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."
                )
                self._rank = accelerator.local_process_index
                self._world_size = accelerator.num_processes
                # manually set model to use gpu, for case where many GPUs available but
                # only seek to use one
                self._device = (
                    torch.device(f"cuda:{accelerator.local_process_index}")
                    if torch.cuda.is_available()
                    else torch.device("cpu")
                )
285
286
287
288
289
290
                try:
                    self.model.to(self.device)
                except ValueError:
                    eval_logger.info(
                        "Failed to place model onto specified device. This may be because the model is quantized via `bitsandbytes`. If the desired GPU is being used, this message is safe to ignore."
                    )
291
            else:
292
                self._model = accelerator.prepare_model(self.model, evaluation_mode=True)
293
294
295
296
297
298
299
300
                self._device = torch.device(f"cuda:{accelerator.local_process_index}")
                self.accelerator = accelerator

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

                self._rank = self.accelerator.local_process_index
                self._world_size = self.accelerator.num_processes
haileyschoelkopf's avatar
haileyschoelkopf committed
301

302
303
304
305
306
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

307
308
309
310
311
312
313
314
    @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

315
316
317
318
319
320
321
    @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

    @property
    def max_length(self):
322
323
324
325
326
327
328
329
330
331
332
        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
333

334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
    @property
    def max_gen_toks(self):
        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

Benjamin Fattori's avatar
Benjamin Fattori committed
354
355
356
357
358
359
    def _detect_batch_size(self, requests=None, pos=0):
        if requests:
            _, context_enc, continuation_enc = requests[pos]
            max_length = len(
                (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1]
            )
360
361
            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
362
363
        else:
            max_length = self.max_length
lintangsutawika's avatar
lintangsutawika committed
364

Benjamin Fattori's avatar
Benjamin Fattori committed
365
366
367
        # 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):
368
369
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                length = max(max_context_enc, max_cont_enc)
lintangsutawika's avatar
lintangsutawika committed
370
371
372
                batched_conts = torch.ones(
                    (batch_size, length), device=self.device
                ).long()
373
374
                test_batch = torch.ones((batch_size, length), device=self.device).long()
                call_kwargs = {
lintangsutawika's avatar
lintangsutawika committed
375
376
377
                    "attn_mask": test_batch,
                    "labels": batched_conts,
                }
378
379
            else:
                call_kwargs = {}
lintangsutawika's avatar
lintangsutawika committed
380
381
382
                test_batch = torch.ones(
                    (batch_size, max_length), device=self.device
                ).long()
Benjamin Fattori's avatar
Benjamin Fattori committed
383
            for _ in range(5):
384
                out = F.log_softmax(self._model_call(test_batch, **call_kwargs), dim=-1)
lintangsutawika's avatar
lintangsutawika committed
385
386
                out = out  # Identity process so that it passes pre-commit

Benjamin Fattori's avatar
Benjamin Fattori committed
387
388
389
390
            return batch_size

        batch_size = forward_batch()

391
392
393
394
395
396
397
398
399
400
401
        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)
            utils.clear_torch_cache()
            return batch_size

        utils.clear_torch_cache()
Benjamin Fattori's avatar
Benjamin Fattori committed
402
403
        return batch_size

404
    def tok_encode(self, string: str, left_truncate_len=None):
haileyschoelkopf's avatar
haileyschoelkopf committed
405
        """ """
406
407
408
409
410
411
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
            add_special_tokens = False
        elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
            add_special_tokens = True

        encoding = self.tokenizer.encode(string, add_special_tokens=add_special_tokens)
haileyschoelkopf's avatar
haileyschoelkopf committed
412

413
414
415
        # 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
416

417
418
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
    def tok_batch_encode(
        self, strings: List[str], padding_side="left", left_truncate_len=None
    ):
        # 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

        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
            add_special_tokens = False
        elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
            add_special_tokens = True

        encoding = self.tokenizer(
            strings,
            padding="longest",
            return_tensors="pt",
            add_special_tokens=add_special_tokens,
        )
        if left_truncate_len:
            encoding["input_ids"] = encoding["input_ids"][:, -left_truncate_len:]
            encoding["attention_mask"] = encoding["attention_mask"][
                :, -left_truncate_len:
            ]
        self.tokenizer.padding_side = old_padding_side

        return encoding["input_ids"], encoding["attention_mask"]

446
447
448
449
450
451
452
453
    def tok_decode(self, tokens):
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
            return self.tokenizer.decode(tokens)
        elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
            return self.tokenizer.decode(tokens, skip_special_tokens=True)

    def _model_call(self, inps, attn_mask=None, labels=None):
        """
haileyschoelkopf's avatar
haileyschoelkopf committed
454
        :param inps: torch.Tensor
455
456
457
458
459
460
461
462
463
464
465
466
467
            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():
468
469
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
470
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
haileyschoelkopf's avatar
haileyschoelkopf committed
471
472
473
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
474
475
476
477
478
479
480
481
482
483
484
485
486
            else:
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                return self.model(inps).logits

    def _model_generate(self, context, max_length, stop, **generation_kwargs):
        # we require users to pass do_sample=True explicitly
        # for non-greedy gen. This should be reevaluated when considering beam search.
        if "do_sample" not in generation_kwargs.keys():
            generation_kwargs["do_sample"] = False
        # build stopping criteria
        stopping_criteria = stop_sequences_criteria(
            self.tokenizer, stop, 1, context.shape[0]
        )
487
488
489
490
491
492
493
494
        return self.model.generate(
            context,
            max_length=max_length,
            stopping_criteria=stopping_criteria,
            pad_token_id=self.eot_token_id,
            use_cache=True,
            **generation_kwargs,
        )
495
496
497

    def _select_cont_toks(self, logits, contlen=None, inplen=None):
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
498
499
500
            assert (
                contlen and inplen
            ), "Must pass input len and cont. len to select scored logits for causal LM"
501
502
503
504
            # discard right-padding.
            # also discard the input/context tokens. we'll only score continuations.
            logits = logits[inplen - contlen : inplen]
        elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
505
506
507
508
            assert (
                contlen and not inplen
            ), "Selecting scored logits for Seq2SeqLM requires only cont. len"
            # only discard right-padding.
509
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
510
511
            logits = logits[:contlen]

512
513
        return logits

514
515
516
517
518
519
520
521
522
523
524
    def _encode_pair(self, context, continuation):
        n_spaces = len(context) - len(context.rstrip())
        if n_spaces > 0:
            continuation = context[-n_spaces:] + continuation
            context = context[:-n_spaces]
        whole_enc = self.tok_encode(context + continuation)
        context_enc = self.tok_encode(context)
        context_enc_len = len(context_enc)
        continuation_enc = whole_enc[context_enc_len:]
        return context_enc, continuation_enc

525
526
527
528
529
    def loglikelihood(self, requests):
        new_reqs = []
        for context, continuation in [req.args for req in requests]:
            if context == "":
                # end of text as context
530
531
532
                context_enc, continuation_enc = [self.eot_token_id], self.tok_encode(
                    continuation
                )
533
            else:
534
                context_enc, continuation_enc = self._encode_pair(context, continuation)
535
536
537
538
539
540
541

            new_reqs.append(((context, continuation), context_enc, continuation_enc))

        return self._loglikelihood_tokens(new_reqs)

    def loglikelihood_rolling(self, requests):
        loglikelihoods = []
Benjamin Fattori's avatar
Benjamin Fattori committed
542
543
544
545
546
547
548
549
550

        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

551
552
553
554
555
556
        for (string,) in tqdm([req.args for req in requests], disable=(self.rank != 0)):
            rolling_token_windows = list(
                map(
                    utils.make_disjoint_window,
                    utils.get_rolling_token_windows(
                        token_list=self.tok_encode(string),
haileyschoelkopf's avatar
haileyschoelkopf committed
557
                        prefix_token=self.eot_token_id,
558
559
560
561
562
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
563
564

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
            rolling_token_windows = [(None,) + x for x in rolling_token_windows]

            pad_amnt = 0
            if self.world_size > 1:
                # We pad out the external document-level iterator so the inner iterator doesn't hang
                mytensor = torch.tensor(len(rolling_token_windows), device=self.device)
                gathered = (
                    self.accelerator.gather(mytensor).cpu().detach().numpy().tolist()
                )

                pad_amnt = max(gathered) - gathered[self.rank]
                if pad_amnt > 0:
                    rolling_token_windows += pad_amnt * [rolling_token_windows[0]]

            string_nll = self._loglikelihood_tokens(
lintangsutawika's avatar
lintangsutawika committed
580
581
582
                rolling_token_windows,
                disable_tqdm=True,
                override_bs=adaptive_batch_size,
583
584
585
586
587
588
589
590
591
592
593
594
595
            )

            if (self.world_size > 1) and (pad_amnt > 0):
                string_nll = [x[0] for x in string_nll[:-pad_amnt]]
            else:
                # discard is_greedy
                string_nll = [x[0] for x in string_nll]

            string_nll = sum(string_nll)
            loglikelihoods.append(string_nll)

        return loglikelihoods

Benjamin Fattori's avatar
Benjamin Fattori committed
596
    def _loglikelihood_tokens(self, requests, disable_tqdm=False, override_bs=None):
597
598
599
600
601
602
603
604
605
606
607
608
609
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

        def _collate(x):
            # 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

            toks = x[1] + x[2]
            return -len(toks), tuple(toks)
lintangsutawika's avatar
lintangsutawika committed
610

611
        re_ord = utils.Reorderer(requests, _collate)
Benjamin Fattori's avatar
Benjamin Fattori committed
612
613
614
615

        n_reordered_requests = len(re_ord.get_reordered())
        # automatic (variable) batch size detection for vectorization
        # pull longest context sample from request
lintangsutawika's avatar
lintangsutawika committed
616

Benjamin Fattori's avatar
Benjamin Fattori committed
617
618
619
620
        def _batch_scheduler(pos):
            sched = pos // int(n_reordered_requests / self.batch_schedule)
            if sched in self.batch_sizes:
                return self.batch_sizes[sched]
lintangsutawika's avatar
lintangsutawika committed
621
622
623
            if (len(self.batch_sizes) > 1) and (
                self.batch_sizes[sched - 1] == self.max_batch_size
            ):
624
625
626
                # if previous batch size is already maximal, skip recomputation
                self.batch_sizes[sched] = self.max_batch_size
                return self.batch_sizes[sched]
Benjamin Fattori's avatar
Benjamin Fattori committed
627
628
629
            print(
                f"Passed argument batch_size = auto:{self.batch_schedule}. Detecting largest batch size"
            )
lintangsutawika's avatar
lintangsutawika committed
630
631
632
            self.batch_sizes[sched] = self._detect_batch_size(
                re_ord.get_reordered(), pos
            )
Benjamin Fattori's avatar
Benjamin Fattori committed
633
            print(f"Determined largest batch size: {self.batch_sizes[sched]}")
lintangsutawika's avatar
lintangsutawika committed
634
            return self.batch_sizes[sched]
Benjamin Fattori's avatar
Benjamin Fattori committed
635

636
637
        for chunk in utils.chunks(
            tqdm(re_ord.get_reordered(), disable=(disable_tqdm or (self.rank != 0))),
Benjamin Fattori's avatar
Benjamin Fattori committed
638
639
640
641
642
643
            n=self.batch_size
            if self.batch_size != "auto"
            else override_bs
            if override_bs is not None
            else 0,
            fn=_batch_scheduler
lintangsutawika's avatar
lintangsutawika committed
644
645
646
            if self.batch_size == "auto"
            and n_reordered_requests > 0
            and not override_bs
Benjamin Fattori's avatar
Benjamin Fattori committed
647
            else None,
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
        ):
            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
668
                # how this all works (illustrated on a causal decoder-only setup):
669
670
671
672
673
674
675
676
677
678
679
                #          CTX      CONT
                # inp    0 1 2 3|4 5 6 7 8 9   <- last token is deleted by inp[:, :-1]
                # model  \               \
                # logits   1 2 3|4 5 6 7 8 9   <- the ctx half gets tossed out by the
                # cont_toks      4 5 6 7 8 9      [:, -len(continuation_enc):, :self.vocab_size] slice

                # when too long to fit in context, truncate from the left
                if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
                    inp = torch.tensor(
                        (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1],
                        dtype=torch.long,
680
681
                        device=self.device,
                    )
682
683
684
685
686
                    (inplen,) = inp.shape
                elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                    inp = torch.tensor(
                        (context_enc)[-self.max_length :],
                        dtype=torch.long,
haileyschoelkopf's avatar
haileyschoelkopf committed
687
                        device=self.device,
688
                    )
689
                    (inplen,) = inp.shape
690
691
692
693

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

694
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
695
                        (continuation_enc)[-self.max_length :],
696
697
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
698
                        dtype=torch.long,
699
700
                        device=self.device,
                    )
701
702
                    (contlen,) = cont.shape

703
704
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
705
706
707
708
709
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
710

haileyschoelkopf's avatar
haileyschoelkopf committed
711
712
713
714
715
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
716
717
718
719

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
720

721
722
723
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
724
725
726
                batched_inps = utils.pad_and_concat(
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
727
728
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                # TODO: left-pad encoder inps and mask?
haileyschoelkopf's avatar
haileyschoelkopf committed
729
730
731
732
733
734
735
736
737
738
739
740
741
                batched_inps = utils.pad_and_concat(
                    padding_len_inp, inps
                )  # [batch, padding_len_inp]
                batched_conts = utils.pad_and_concat(
                    padding_len_cont, conts
                )  # [batch, padding_len_cont]
                batched_encoder_mask = utils.pad_and_concat(
                    padding_len_inp, encoder_attns
                )  # [batch, padding_len_inp]
                call_kwargs = {
                    "attn_mask": batched_encoder_mask,
                    "labels": batched_conts,
                }
742
743
744

            multi_logits = F.log_softmax(
                self._model_call(batched_inps, **call_kwargs), dim=-1
745
            )  # [batch, padding_length (inp or cont), vocab]
746
747
748
749
750
751

            for (cache_key, _, _), logits, inplen, cont_toks in zip(
                chunk, multi_logits, inplens, cont_toks_list
            ):
                # Slice to original seq length
                contlen = len(cont_toks)
haileyschoelkopf's avatar
haileyschoelkopf committed
752
                # take only logits in the continuation
753
                # (discard context toks if decoder-only ; discard right-padding)
754
755
                # 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
756
                ctx_len = (
757
                    inplen + (logits.shape[0] - padding_len_inp)
haileyschoelkopf's avatar
haileyschoelkopf committed
758
759
760
                    if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                    else None
                )
761
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
762
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
763
764
765

                # Check if per-token argmax is exactly equal to continuation
                greedy_tokens = logits.argmax(dim=-1)
766
767
768
                cont_toks = torch.tensor(
                    cont_toks, dtype=torch.long, device=self.device
                ).unsqueeze(
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
                    0
                )  # [1, seq]
                max_equal = (greedy_tokens == cont_toks).all()

                # Obtain log-probs at the corresponding continuation token indices
                # last_token_slice = logits[:, -1, :].squeeze(0).tolist()
                logits = torch.gather(logits, 2, cont_toks.unsqueeze(-1)).squeeze(
                    -1
                )  # [1, seq]

                # Answer: (log prob, is-exact-match)
                answer = (float(logits.sum()), bool(max_equal))

                res.append(answer)

haileyschoelkopf's avatar
haileyschoelkopf committed
784
785
                self.cache_hook.add_partial("loglikelihood", cache_key, answer)

786
787
788
        return re_ord.get_original(res)

    def greedy_until(self, requests):
789
790
        res = defaultdict(list)
        re_ords = {}
791
792

        def _collate(x):
793
794
795
796
797
798
            # 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
799
            toks = self.tok_encode(x[0])
haileyschoelkopf's avatar
haileyschoelkopf committed
800
            return -len(toks), x[0]
801

802
803
804
        # 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.
805
806
        grouper = utils.Grouper(requests, lambda x: str(x.args[1]))
        for key, reqs in grouper.get_grouped().items():
807
            # within each set of reqs for given kwargs, we reorder by token length, descending.
808
            re_ords[key] = utils.Reorderer([req.args for req in reqs], _collate)
809

810
811
812
        pbar = tqdm(total=len(requests), disable=(self.rank != 0))

        # for each different set of kwargs, we execute all requests, by batch.
813
814
        for key, re_ord in re_ords.items():
            for chunk in utils.chunks(
haileyschoelkopf's avatar
haileyschoelkopf committed
815
                re_ord.get_reordered(),
816
817
818
                self.batch_size,
            ):
                contexts, all_gen_kwargs = zip(*chunk)
819
820
821
822
                # 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.
823
824
825
826
827
828
829
830
831
                until = None
                if isinstance(gen_kwargs, dict):
                    kwargs = copy.deepcopy(gen_kwargs)  # edge case for repeats > 1
                    if "until" in kwargs.keys():
                        until = kwargs.pop("until")
                        if isinstance(until, str):
                            until = [kwargs]
                        elif not isinstance(until, list):
                            raise ValueError(
832
                                f"Expected `kwargs['until']` to be of type Union[str,list] but got {until}"
833
834
835
                            )
                else:
                    raise ValueError(
836
                        f"Expected `kwargs` to be of type `dict` but got {kwargs}"
837
838
839
840
841
842
843
844
                    )
                if not until:
                    until = [self.tok_decode(self.eot_token_id)]
                if "max_gen_toks" in kwargs.keys():
                    max_gen_toks = kwargs.pop("max_gen_toks")
                else:
                    max_gen_toks = self.max_gen_toks
                # first stop sequence is used to halt generation upon encountering
845
                primary_until = [until[0]]
846

847
                # set the max length in tokens of inputs ("context_enc")
haileyschoelkopf's avatar
haileyschoelkopf committed
848
                if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
849
850
851
852
853
                    # max len for inputs = max length, minus room to generate the max new tokens
                    max_ctx_len = self.max_length - max_gen_toks
                elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                    # max len for inputs = encoder's whole max_length
                    max_ctx_len = self.max_length
854

855
                # encode, pad, and truncate contexts for this batch
856
857
858
859
860
861
                context_enc, attn_masks = self.tok_batch_encode(
                    contexts, left_truncate_len=max_ctx_len
                )
                context_enc = context_enc.to(self.device)
                attn_masks = attn_masks.to(self.device)

862
                if "max_length" not in kwargs:
Lintang Sutawika's avatar
Lintang Sutawika committed
863
                    kwargs["max_length"] = context_enc.shape[1] + max_gen_toks
864

865
                # perform batched generation
866
867
868
869
870
871
                cont = self._model_generate(
                    context=context_enc,
                    attention_mask=attn_masks,
                    stop=primary_until,
                    **kwargs,
                )
872

873
874
875
876
877
                cont_toks_list = cont.tolist()
                for cont_toks, context in zip(cont_toks_list, contexts):
                    # discard context + left-padding toks if using causal decoder-only LM
                    if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
                        cont_toks = cont_toks[context_enc.shape[1] :]
878

879
                    s = self.tok_decode(cont_toks)
880

881
882
                    # use secondary stop seqs to cut off should-have-been-stopped content post-hoc
                    for term in until:
883
884
885
                        if len(term) > 0:
                            # ignore '' separator,
                            # for seq2seq case where self.tok_decode(self.eot_token_id) = ''
886
                            s = s.split(term)[0]
887

888
                    res[key].append(s)
889

890
891
892
893
                    self.cache_hook.add_partial(
                        "greedy_until", (context, gen_kwargs), s
                    )
                    pbar.update(1)
894
            # reorder this group of results back to original unsorted form
895
            res[key] = re_ord.get_original(res[key])
896

897
        pbar.close()
898

899
        return grouper.get_original(res)