huggingface.py 39.7 KB
Newer Older
1
import os
Herbie Bradley's avatar
Herbie Bradley committed
2
from typing import List, Optional, Union
3

4
import torch
Herbie Bradley's avatar
Herbie Bradley committed
5
import torch.nn.functional as F
6
import transformers
Herbie Bradley's avatar
Herbie Bradley committed
7
8
9
10
from accelerate import Accelerator, DistributedType, find_executable_batch_size
from peft import PeftModel
from peft import __version__ as PEFT_VERSION
from tqdm import tqdm
11
12
13
14
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
)
15
16
17
18
19
20

from lm_eval import utils
from lm_eval.api.model import LM
from lm_eval.api.registry import register_model
from lm_eval.utils import MultiTokenEOSCriteria, stop_sequences_criteria

21
eval_logger = utils.eval_logger
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46


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
47
48


49
@register_model("hf-auto", "hf", "huggingface")
50
class HFLM(LM):
51
52
53
54
55
56
57
    """
    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.
    """

58
    AUTO_MODEL_CLASS = None
59
    _DEFAULT_MAX_LENGTH = 2048
haileyschoelkopf's avatar
haileyschoelkopf committed
60

61
62
    def __init__(
        self,
63
64
65
66
        pretrained: Optional[str] = "gpt2",
        revision: Optional[str] = "main",
        subfolder: Optional[str] = None,
        tokenizer: Optional[str] = None,
lintangsutawika's avatar
lintangsutawika committed
67
        truncation: Optional[bool] = False,
68
69
        max_length: Optional[int] = None,
        device: Optional[str] = "cuda",
70
        dtype: Optional[Union[str, torch.dtype]] = "auto",
Benjamin Fattori's avatar
Benjamin Fattori committed
71
72
        batch_size: Optional[Union[int, str]] = 1,
        max_batch_size: Optional[int] = 64,
73
74
        low_cpu_mem_usage: Optional[bool] = True,
        trust_remote_code: Optional[bool] = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
75
        use_fast_tokenizer: Optional[bool] = True,
lintangsutawika's avatar
lintangsutawika committed
76
        cache_dir: Optional[Union[str, os.PathLike]] = None,
77
        # arguments used for splitting a model across GPUs naively.
78
79
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
80
81
82
83
        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",
84
85
86
87
88
89
90
91
        # 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,
Ethan Smith's avatar
Ethan Smith committed
92
    ) -> None:
93
94
95
96
        super().__init__()

        assert isinstance(device, str)
        assert isinstance(pretrained, str)
Benjamin Fattori's avatar
Benjamin Fattori committed
97
        assert isinstance(batch_size, (int, str))
98
99

        gpus = torch.cuda.device_count()
100
        accelerator = Accelerator()
haileyschoelkopf's avatar
haileyschoelkopf committed
101

102
        if not (parallelize or accelerator.num_processes > 1):
103
            # use user-passed device
104
            device_list = set(
105
                ["cuda", "cpu"]
106
                + [f"cuda:{i}" for i in range(torch.cuda.device_count())]
107
                + ["mps", "mps:0"]
108
            )
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
                if device in ("mps", "mps:0") and "dev" not in torch.__version__:
115
                    eval_logger.info(
116
117
118
                        "MPS: Setting dtype to float32. To use float16 with MPS, please install a nightly build of "
                        "PyTorch: pip3 install --pre torch torchvision torchaudio --index-url "
                        "https://download.pytorch.org/whl/nightly/cpu"
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
        )

lintangsutawika's avatar
lintangsutawika committed
245
246
        self.truncation = truncation

247
        self.vocab_size = self.tokenizer.vocab_size
haileyschoelkopf's avatar
haileyschoelkopf committed
248
        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
249

250
251
        self._max_length = max_length

Benjamin Fattori's avatar
Benjamin Fattori committed
252
253
254
255
256
257
258
259
260
261
        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)
262
263
264
265
266
267
268
269
270
271
272

        # 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:
273
                # TODO: make sure there's still never an edge case where we unintentionally default to CPU
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
                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")
                )
289
290
291
292
293
294
                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."
                    )
295
            else:
Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
296
                assert accelerator.distributed_type in [
lintangsutawika's avatar
lintangsutawika committed
297
298
                    DistributedType.FSDP,
                    DistributedType.MULTI_GPU,
Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
299
                ], "Unsupported distributed type provided. Only DDP and FSDP are supported."
300
                if accelerator.distributed_type == DistributedType.FSDP:
Hailey Schoelkopf's avatar
Hailey Schoelkopf committed
301
                    self._model = accelerator.prepare(self.model)
302
303
                else:
                    self._model = accelerator.prepare_model(
lintangsutawika's avatar
lintangsutawika committed
304
                        self.model, evaluation_mode=True
305
                    )
306
307
308
309
310
311
312
313
                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
314

315
316
317
318
319
    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

320
321
322
323
324
325
326
327
    @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

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

347
    @property
Ethan Smith's avatar
Ethan Smith committed
348
    def max_gen_toks(self) -> int:
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
        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

Ethan Smith's avatar
Ethan Smith committed
367
    def _detect_batch_size(self, requests=None, pos: int = 0):
Benjamin Fattori's avatar
Benjamin Fattori committed
368
369
370
371
372
        if requests:
            _, context_enc, continuation_enc = requests[pos]
            max_length = len(
                (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1]
            )
373
374
            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
375
376
        else:
            max_length = self.max_length
lintangsutawika's avatar
lintangsutawika committed
377

Benjamin Fattori's avatar
Benjamin Fattori committed
378
379
380
        # 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):
381
382
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                length = max(max_context_enc, max_cont_enc)
lintangsutawika's avatar
lintangsutawika committed
383
384
385
                batched_conts = torch.ones(
                    (batch_size, length), device=self.device
                ).long()
386
387
                test_batch = torch.ones((batch_size, length), device=self.device).long()
                call_kwargs = {
lintangsutawika's avatar
lintangsutawika committed
388
389
390
                    "attn_mask": test_batch,
                    "labels": batched_conts,
                }
391
392
            else:
                call_kwargs = {}
lintangsutawika's avatar
lintangsutawika committed
393
394
395
                test_batch = torch.ones(
                    (batch_size, max_length), device=self.device
                ).long()
Benjamin Fattori's avatar
Benjamin Fattori committed
396
            for _ in range(5):
397
                out = F.log_softmax(self._model_call(test_batch, **call_kwargs), dim=-1)
lintangsutawika's avatar
lintangsutawika committed
398
399
                out = out  # Identity process so that it passes pre-commit

Benjamin Fattori's avatar
Benjamin Fattori committed
400
401
402
403
            return batch_size

        batch_size = forward_batch()

404
405
406
407
408
409
410
411
412
413
414
        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
415
416
        return batch_size

417
    def tok_encode(self, string: str, left_truncate_len=None, add_special_tokens=None):
haileyschoelkopf's avatar
haileyschoelkopf committed
418
        """ """
419
420
421
422
423
        if add_special_tokens is None:
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
                add_special_tokens = False
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                add_special_tokens = True
424
425

        encoding = self.tokenizer.encode(string, add_special_tokens=add_special_tokens)
haileyschoelkopf's avatar
haileyschoelkopf committed
426

427
428
429
        # 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
430

431
432
        return encoding

haileyschoelkopf's avatar
haileyschoelkopf committed
433
    def tok_batch_encode(
lintangsutawika's avatar
lintangsutawika committed
434
435
436
        self,
        strings: List[str],
        padding_side: str = "left",
437
438
        left_truncate_len: int = None,
        truncation: bool = False,
haileyschoelkopf's avatar
haileyschoelkopf committed
439
440
441
442
443
444
445
446
447
448
449
450
    ):
        # 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,
lintangsutawika's avatar
lintangsutawika committed
451
            truncation=truncation,
haileyschoelkopf's avatar
haileyschoelkopf committed
452
453
454
455
456
457
458
459
460
461
462
463
464
            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"]

465
466
467
468
469
470
471
472
    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
473
        :param inps: torch.Tensor
474
475
476
477
478
479
480
481
482
483
484
485
486
            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():
487
488
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
489
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
haileyschoelkopf's avatar
haileyschoelkopf committed
490
491
492
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
493
494
495
496
497
498
499
500
501
502
503
504
505
            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]
        )
506
        return self.model.generate(
507
            input_ids=context,
508
509
510
511
512
513
            max_length=max_length,
            stopping_criteria=stopping_criteria,
            pad_token_id=self.eot_token_id,
            use_cache=True,
            **generation_kwargs,
        )
514
515
516

    def _select_cont_toks(self, logits, contlen=None, inplen=None):
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
517
518
519
            assert (
                contlen and inplen
            ), "Must pass input len and cont. len to select scored logits for causal LM"
520
521
522
523
            # 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
524
525
526
527
            assert (
                contlen and not inplen
            ), "Selecting scored logits for Seq2SeqLM requires only cont. len"
            # only discard right-padding.
528
            # the logits input to this fn only contain decoder-side tokens.
haileyschoelkopf's avatar
haileyschoelkopf committed
529
530
            logits = logits[:contlen]

531
532
        return logits

533
534
535
536
537
    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]
538
539
540
541
542
543

        whole_enc = self.tok_encode(context + continuation, add_special_tokens=False)
        context_enc = self.tok_encode(context, add_special_tokens=False)

        # whole_enc = self.tok_encode(context + continuation)
        # context_enc = self.tok_encode(context, add_special_tokens=False)
544
545
546
547
        context_enc_len = len(context_enc)
        continuation_enc = whole_enc[context_enc_len:]
        return context_enc, continuation_enc

548
549
550
551
552
    def loglikelihood(self, requests):
        new_reqs = []
        for context, continuation in [req.args for req in requests]:
            if context == "":
                # end of text as context
553
554
555
                context_enc, continuation_enc = [self.eot_token_id], self.tok_encode(
                    continuation
                )
556
            else:
557
                context_enc, continuation_enc = self._encode_pair(context, continuation)
558
559
560
561
562
563
564

            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
565
566
567
568
569
570
571
572
573

        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

574
        for (string,) in tqdm([req.args for req in requests], disable=(self.rank != 0)):
Herbie Bradley's avatar
Herbie Bradley committed
575
576
577
578
            if len(string) == 0:
                loglikelihoods.append(float("-inf"))
                continue

579
580
581
582
583
            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
584
                        prefix_token=self.eot_token_id,
585
586
587
588
589
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
590
591

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
            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
607
608
609
                rolling_token_windows,
                disable_tqdm=True,
                override_bs=adaptive_batch_size,
610
611
612
613
614
615
616
617
618
619
620
621
622
            )

            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

623
624
625
626
627
628
629
630
631
632
633
634
635
    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
636
        self.batch_sizes[sched] = self._detect_batch_size(n_reordered_requests, pos)
637
638
        print(f"Determined largest batch size: {self.batch_sizes[sched]}")
        return self.batch_sizes[sched]
639

Ethan Smith's avatar
Ethan Smith committed
640
641
642
    def _loglikelihood_tokens(
        self, requests, disable_tqdm: bool = False, override_bs=None
    ):
643
644
645
646
647
648
649
650
651
652
653
654
655
        # 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
656

657
        re_ord = utils.Reorderer(requests, _collate)
Benjamin Fattori's avatar
Benjamin Fattori committed
658
659
660
661

        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
662

663
664
        chunks = utils.chunks(
            re_ord.get_reordered(),
Benjamin Fattori's avatar
Benjamin Fattori committed
665
666
667
668
669
            n=self.batch_size
            if self.batch_size != "auto"
            else override_bs
            if override_bs is not None
            else 0,
670
            fn=self._batch_scheduler
lintangsutawika's avatar
lintangsutawika committed
671
672
673
            if self.batch_size == "auto"
            and n_reordered_requests > 0
            and not override_bs
Benjamin Fattori's avatar
Benjamin Fattori committed
674
            else None,
675
676
        )

haileyschoelkopf's avatar
haileyschoelkopf committed
677
        pbar = tqdm(total=len(requests), disable=(disable_tqdm or (self.rank != 0)))
haileyschoelkopf's avatar
haileyschoelkopf committed
678
        for chunk in chunks:
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
            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
698
                # how this all works (illustrated on a causal decoder-only setup):
699
700
701
702
703
704
705
706
707
708
709
                #          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,
710
711
                        device=self.device,
                    )
712
713
714
715
716
                    (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
717
                        device=self.device,
718
                    )
719
                    (inplen,) = inp.shape
720
721
722
723

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

724
                    cont = torch.tensor(
haileyschoelkopf's avatar
haileyschoelkopf committed
725
                        (continuation_enc)[-self.max_length :],
726
727
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
728
                        dtype=torch.long,
729
730
                        device=self.device,
                    )
731
732
                    (contlen,) = cont.shape

733
734
                    conts.append(cont)

haileyschoelkopf's avatar
haileyschoelkopf committed
735
736
737
738
739
                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )
740

haileyschoelkopf's avatar
haileyschoelkopf committed
741
742
743
744
745
                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )
746
747
748
749

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)
haileyschoelkopf's avatar
haileyschoelkopf committed
750

751
752
753
            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
haileyschoelkopf's avatar
haileyschoelkopf committed
754
755
756
                batched_inps = utils.pad_and_concat(
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
757
758
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                # TODO: left-pad encoder inps and mask?
haileyschoelkopf's avatar
haileyschoelkopf committed
759
760
761
762
763
764
765
766
767
768
769
770
771
                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,
                }
772
773
774

            multi_logits = F.log_softmax(
                self._model_call(batched_inps, **call_kwargs), dim=-1
775
            )  # [batch, padding_length (inp or cont), vocab]
776
777
778
779
780
781

            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
782
                # take only logits in the continuation
783
                # (discard context toks if decoder-only ; discard right-padding)
784
785
                # 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
786
                ctx_len = (
787
                    inplen + (logits.shape[0] - padding_len_inp)
haileyschoelkopf's avatar
haileyschoelkopf committed
788
789
790
                    if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                    else None
                )
791
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
haileyschoelkopf's avatar
haileyschoelkopf committed
792
                logits = logits.unsqueeze(0)  # [1, seq, vocab]
793
794
795

                # Check if per-token argmax is exactly equal to continuation
                greedy_tokens = logits.argmax(dim=-1)
796
797
798
                cont_toks = torch.tensor(
                    cont_toks, dtype=torch.long, device=self.device
                ).unsqueeze(
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
                    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
814
                self.cache_hook.add_partial("loglikelihood", cache_key, answer)
haileyschoelkopf's avatar
haileyschoelkopf committed
815
816
817
                pbar.update(1)

        pbar.close()
haileyschoelkopf's avatar
haileyschoelkopf committed
818

819
820
        return re_ord.get_original(res)

821
    def generate_until(self, requests):
822
823
        res = defaultdict(list)
        re_ords = {}
824
825

        def _collate(x):
826
827
828
829
830
831
            # 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
832
            toks = self.tok_encode(x[0])
haileyschoelkopf's avatar
haileyschoelkopf committed
833
            return -len(toks), x[0]
834

835
836
837
        # 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.
838
839
        grouper = utils.Grouper(requests, lambda x: str(x.args[1]))
        for key, reqs in grouper.get_grouped().items():
840
            # within each set of reqs for given kwargs, we reorder by token length, descending.
841
            re_ords[key] = utils.Reorderer([req.args for req in reqs], _collate)
842

843
        pbar = tqdm(total=len(requests), disable=(self.rank != 0))
844
845
846
847
848
849
        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
850
        # for each different set of kwargs, we execute all requests, by batch.
851
        for key, re_ord in re_ords.items():
852
            chunks = utils.chunks(
haileyschoelkopf's avatar
haileyschoelkopf committed
853
                re_ord.get_reordered(),
854
855
856
857
858
859
860
861
                n=self.batch_size
                if self.batch_size != "auto"
                else adaptive_batch_size
                if adaptive_batch_size is not None
                else 0,
                fn=self._batch_scheduler
                if self.batch_size == "auto" and not adaptive_batch_size
                else None,
862
            )
haileyschoelkopf's avatar
haileyschoelkopf committed
863
            for chunk in chunks:
864
                contexts, all_gen_kwargs = zip(*chunk)
865
866
867
868
                # 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.
869
870
871
872
873
874
875
876
877
                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(
878
                                f"Expected `kwargs['until']` to be of type Union[str,list] but got {until}"
879
880
881
                            )
                else:
                    raise ValueError(
882
                        f"Expected `kwargs` to be of type `dict` but got {kwargs}"
883
884
885
886
887
888
889
                    )
                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
890

891
                # set the max length in tokens of inputs ("context_enc")
haileyschoelkopf's avatar
haileyschoelkopf committed
892
                if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
893
894
895
896
897
                    # 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
898

899
                # encode, pad, and truncate contexts for this batch
900
                context_enc, attn_masks = self.tok_batch_encode(
lintangsutawika's avatar
lintangsutawika committed
901
902
903
                    contexts,
                    left_truncate_len=max_ctx_len,
                    truncation=self.truncation,
904
905
906
907
                )
                context_enc = context_enc.to(self.device)
                attn_masks = attn_masks.to(self.device)

908
                if "max_length" not in kwargs:
Lintang Sutawika's avatar
Lintang Sutawika committed
909
                    kwargs["max_length"] = context_enc.shape[1] + max_gen_toks
910

911
                # perform batched generation
912
913
914
                cont = self._model_generate(
                    context=context_enc,
                    attention_mask=attn_masks,
915
                    stop=until,
916
917
                    **kwargs,
                )
918

919
920
921
922
923
                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] :]
924

925
                    s = self.tok_decode(cont_toks)
926

927
928
                    # use secondary stop seqs to cut off should-have-been-stopped content post-hoc
                    for term in until:
929
930
931
                        if len(term) > 0:
                            # ignore '' separator,
                            # for seq2seq case where self.tok_decode(self.eot_token_id) = ''
932
                            s = s.split(term)[0]
933

934
                    res[key].append(s)
935

936
                    self.cache_hook.add_partial(
937
                        "generate_until", (context, gen_kwargs), s
938
939
                    )
                    pbar.update(1)
940
            # reorder this group of results back to original unsorted form
941
            res[key] = re_ord.get_original(res[key])
942

943
        pbar.close()
944

945
        return grouper.get_original(res)