huggingface.py 33.6 KB
Newer Older
1
2
import os

3
4
import torch
import transformers
5
6
7
8
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
)
9
from peft import __version__ as PEFT_VERSION, PeftModel
10
11

import copy
12
from collections import defaultdict
13
from tqdm import tqdm
14
from pathlib import Path
15
16
17
18
19
20
21
22
23
24
25

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

from accelerate import Accelerator
26
from typing import List, Optional, Union
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51


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
52
53


54
@register_model("hf-auto", "hf", "huggingface")
55
class HFLM(LM):
56
57
58
59
60
61
62
    """
    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.
    """

63
    AUTO_MODEL_CLASS = None
64
    _DEFAULT_MAX_LENGTH = 2048
haileyschoelkopf's avatar
haileyschoelkopf committed
65

66
67
    def __init__(
        self,
68
69
70
71
72
73
        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",
74
        dtype: Optional[Union[str, torch.dtype]] = "auto",
75
76
77
        batch_size: Optional[int] = 1,
        low_cpu_mem_usage: Optional[bool] = True,
        trust_remote_code: Optional[bool] = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
78
        use_fast_tokenizer: Optional[bool] = True,
79
        cache_dir: Optional[Union[str,os.PathLike]] = None,
80
        # arguments used for splitting a model across GPUs naively.
81
82
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
83
84
85
86
        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",
87
88
89
90
91
92
93
94
        # 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,
95
96
97
98
99
100
101
102
    ):
        super().__init__()

        assert isinstance(device, str)
        assert isinstance(pretrained, str)
        assert isinstance(batch_size, int)

        gpus = torch.cuda.device_count()
103
        accelerator = Accelerator()
haileyschoelkopf's avatar
haileyschoelkopf committed
104

105
        if not (parallelize or accelerator.num_processes > 1):
106
            # use user-passed device
107
            device_list = set(
baberabb's avatar
add mps  
baberabb committed
108
                ["cuda", "cpu", "mps"]
109
110
                + [f"cuda:{i}" for i in range(torch.cuda.device_count())]
            )
111
            if device:
112
                if device not in device_list:
113
114
115
                    device = int(device)
                self._device = torch.device(device)
                eval_logger.info(f"Using device '{device}'")
116
117
                if device == "mps":
                    eval_logger.info(
baberabb's avatar
baberabb committed
118
                        "MPS is still in beta and only supports float32; setting dtype to float32."
119
                    )
120
121
122
123
124
125
126
127
            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")
                )
128
        else:
129
130
131
132
            if device != "cuda":
                eval_logger.info(
                    f"Using `accelerate launch` or `parallelize=True`, device '{device}' will be overridden when placing model."
                )
133
            # TODO: include in warning that `load_in_8bit` etc. affect this too
134
135
136
            self._device = device

        model_kwargs = {}
137
        if parallelize:
138
139
140
141
142
143
            model_kwargs = _get_accelerate_args(
                device_map_option,
                max_memory_per_gpu,
                max_cpu_memory,
                offload_folder,
            )
144
145
146
147
148
149
150

        # 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,
151
            trust_remote_code=trust_remote_code,
152
153
154
155
        )

        if getattr(self._config, "model_type") in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
            self.AUTO_MODEL_CLASS = transformers.AutoModelForCausalLM
156
157
158
159
160
161
162
163
164
165
166
167
        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
168
        else:
haileyschoelkopf's avatar
haileyschoelkopf committed
169
            self.AUTO_MODEL_CLASS = transformers.AutoModelForSeq2SeqLM
170

haileyschoelkopf's avatar
haileyschoelkopf committed
171
172
173
174
        assert self.AUTO_MODEL_CLASS in [
            transformers.AutoModelForCausalLM,
            transformers.AutoModelForSeq2SeqLM,
        ]
175

176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
        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
200
201
202
203
204
205
206
            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]",
                )
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225

            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
            )

226
        # forever after, access self._model through self.model property
227
        self.model.eval()
228
229
230
        self.model.tie_weights()
        if gpus <= 1 and not parallelize:
            # place model onto device, if not using HF Accelerate in any form
231
232
233
234
235
236
            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
237

238
239
240
        self.tokenizer = transformers.AutoTokenizer.from_pretrained(
            pretrained if tokenizer is None else tokenizer,
            revision=revision,
241
            trust_remote_code=trust_remote_code,
haileyschoelkopf's avatar
haileyschoelkopf committed
242
            use_fast=use_fast_tokenizer,
243
244
245
        )

        self.vocab_size = self.tokenizer.vocab_size
haileyschoelkopf's avatar
haileyschoelkopf committed
246
        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
247

248
249
        self._max_length = max_length

250
        # multithreading and batching
251
252
253
254
255
256
257
258
259
260
261
262
        self.batch_size_per_gpu = batch_size

        # 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:
263
                # TODO: make sure there's still never an edge case where we unintentionally default to CPU
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
                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")
                )
279
280
281
282
283
284
                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."
                    )
285
            else:
haileyschoelkopf's avatar
haileyschoelkopf committed
286
                self._model = accelerator.prepare(self.model)
287
288
289
290
291
292
293
294
                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
295

296
297
298
299
300
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

301
302
303
304
305
306
307
308
    @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

309
310
311
312
313
314
315
    @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):
316
317
318
319
320
321
322
323
324
325
326
        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
327

328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
    @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

    def tok_encode(self, string: str, left_truncate_len=None):
haileyschoelkopf's avatar
haileyschoelkopf committed
349
        """ """
350
351
352
353
354
355
        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
356

357
358
359
        # 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
360

361
362
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
    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"]

390
391
392
393
394
395
396
397
    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
398
        :param inps: torch.Tensor
399
400
401
402
403
404
405
406
407
408
409
410
411
            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():
412
413
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
414
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
haileyschoelkopf's avatar
haileyschoelkopf committed
415
416
417
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
418
419
420
421
422
423
424
425
426
427
428
429
430
            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]
        )
431
432
433
434
435
436
437
438
        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,
        )
439
440
441

    def _select_cont_toks(self, logits, contlen=None, inplen=None):
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
442
443
444
            assert (
                contlen and inplen
            ), "Must pass input len and cont. len to select scored logits for causal LM"
445
446
447
448
            # 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
449
450
451
452
            assert (
                contlen and not inplen
            ), "Selecting scored logits for Seq2SeqLM requires only cont. len"
            # only discard right-padding.
453
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
454
455
            logits = logits[:contlen]

456
457
        return logits

458
459
460
461
462
463
464
465
466
467
468
    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

469
470
471
472
473
    def loglikelihood(self, requests):
        new_reqs = []
        for context, continuation in [req.args for req in requests]:
            if context == "":
                # end of text as context
474
475
476
                context_enc, continuation_enc = [self.eot_token_id], self.tok_encode(
                    continuation
                )
477
            else:
478
                context_enc, continuation_enc = self._encode_pair(context, continuation)
479
480
481
482
483
484
485
486
487
488
489
490
491

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

        return self._loglikelihood_tokens(new_reqs)

    def loglikelihood_rolling(self, requests):
        loglikelihoods = []
        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
492
                        prefix_token=self.eot_token_id,
493
494
495
496
497
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
498
499

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
            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(
                rolling_token_windows, disable_tqdm=True
            )

            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

    def _loglikelihood_tokens(self, requests, disable_tqdm=False):
        # 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)

        # TODO: automatic (variable) batch size detection for vectorization
        re_ord = utils.Reorderer(requests, _collate)
        for chunk in utils.chunks(
            tqdm(re_ord.get_reordered(), disable=(disable_tqdm or (self.rank != 0))),
            self.batch_size,
        ):
            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
569
                # how this all works (illustrated on a causal decoder-only setup):
570
571
572
573
574
575
576
577
578
579
580
                #          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,
581
582
                        device=self.device,
                    )
583
584
585
586
587
                    (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
588
                        device=self.device,
589
                    )
590
                    (inplen,) = inp.shape
591
592
593
594

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

595
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
596
                        (continuation_enc)[-self.max_length :],
597
598
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
599
                        dtype=torch.long,
600
601
                        device=self.device,
                    )
602
603
                    (contlen,) = cont.shape

604
605
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
606
607
608
609
610
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
611

haileyschoelkopf's avatar
haileyschoelkopf committed
612
613
614
615
616
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
617
618
619
620

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
621

622
623
624
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
625
626
627
                batched_inps = utils.pad_and_concat(
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
628
629
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                # TODO: left-pad encoder inps and mask?
haileyschoelkopf's avatar
haileyschoelkopf committed
630
631
632
633
634
635
636
637
638
639
640
641
642
                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,
                }
643
644
645

            multi_logits = F.log_softmax(
                self._model_call(batched_inps, **call_kwargs), dim=-1
646
            )  # [batch, padding_length (inp or cont), vocab]
647
648
649
650
651
652

            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
653
                # take only logits in the continuation
654
                # (discard context toks if decoder-only ; discard right-padding)
655
656
                # 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
657
                ctx_len = (
658
                    inplen + (logits.shape[0] - padding_len_inp)
haileyschoelkopf's avatar
haileyschoelkopf committed
659
660
661
                    if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                    else None
                )
662
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
663
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
664
665
666

                # Check if per-token argmax is exactly equal to continuation
                greedy_tokens = logits.argmax(dim=-1)
667
668
669
                cont_toks = torch.tensor(
                    cont_toks, dtype=torch.long, device=self.device
                ).unsqueeze(
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
                    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
685
686
                self.cache_hook.add_partial("loglikelihood", cache_key, answer)

687
688
689
        return re_ord.get_original(res)

    def greedy_until(self, requests):
690
691
        res = defaultdict(list)
        re_ords = {}
692
693

        def _collate(x):
694
695
696
697
698
699
            # 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
700
            toks = self.tok_encode(x[0])
haileyschoelkopf's avatar
haileyschoelkopf committed
701
            return -len(toks), x[0]
702

703
704
705
        # 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.
706
707
        grouper = utils.Grouper(requests, lambda x: str(x.args[1]))
        for key, reqs in grouper.get_grouped().items():
708
            # within each set of reqs for given kwargs, we reorder by token length, descending.
709
            re_ords[key] = utils.Reorderer([req.args for req in reqs], _collate)
710

711
712
713
        pbar = tqdm(total=len(requests), disable=(self.rank != 0))

        # for each different set of kwargs, we execute all requests, by batch.
714
715
        for key, re_ord in re_ords.items():
            for chunk in utils.chunks(
haileyschoelkopf's avatar
haileyschoelkopf committed
716
                re_ord.get_reordered(),
717
718
719
                self.batch_size,
            ):
                contexts, all_gen_kwargs = zip(*chunk)
720
721
722
723
                # 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.
724
725
726
727
728
729
730
731
732
                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(
733
                                f"Expected `kwargs['until']` to be of type Union[str,list] but got {until}"
734
735
736
                            )
                else:
                    raise ValueError(
737
                        f"Expected `kwargs` to be of type `dict` but got {kwargs}"
738
739
740
741
742
743
744
745
                    )
                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
746
                primary_until = [until[0]]
747

748
                # set the max length in tokens of inputs ("context_enc")
haileyschoelkopf's avatar
haileyschoelkopf committed
749
                if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
750
751
752
753
754
                    # 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
755

756
                # encode, pad, and truncate contexts for this batch
757
758
759
760
761
762
                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)

763
                if "max_length" not in kwargs:
Lintang Sutawika's avatar
Lintang Sutawika committed
764
                    kwargs["max_length"] = context_enc.shape[1] + max_gen_toks
765

766
                # perform batched generation
767
768
769
770
771
772
                cont = self._model_generate(
                    context=context_enc,
                    attention_mask=attn_masks,
                    stop=primary_until,
                    **kwargs,
                )
773

774
775
776
777
778
                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] :]
779

780
                    s = self.tok_decode(cont_toks)
781

782
783
                    # use secondary stop seqs to cut off should-have-been-stopped content post-hoc
                    for term in until:
784
785
786
                        if len(term) > 0:
                            # ignore '' separator,
                            # for seq2seq case where self.tok_decode(self.eot_token_id) = ''
787
                            s = s.split(term)[0]
788

789
                    res[key].append(s)
790

791
792
793
794
                    self.cache_hook.add_partial(
                        "greedy_until", (context, gen_kwargs), s
                    )
                    pbar.update(1)
795
            # reorder this group of results back to original unsorted form
796
            res[key] = re_ord.get_original(res[key])
797

798
        pbar.close()
799

800
        return grouper.get_original(res)