huggingface.py

import torch
import transformers
from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES

import copy
from collections import defaultdict
from tqdm import tqdm

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
from typing import List, Optional, Union


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


@register_model("hf-auto", "hf", "huggingface")
class HFLM(LM):
    """
    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.
    """

    AUTO_MODEL_CLASS = None
    _DEFAULT_MAX_LENGTH = 2048

    def __init__(
        self,
        device: Optional[Union[str, int]] = "cuda",
        pretrained="gpt2",
        revision="main",
        low_cpu_mem_usage=None,
        max_length=None,
        subfolder=None,
        tokenizer=None,
        batch_size=1,
        dtype: Optional[Union[str, torch.dtype]] = "auto",
        # arguments used for splitting a model across GPUs naively.
        # only used if `parallelize=True`.
        parallelize: Optional[bool] = False,
        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",
    ):
        super().__init__()

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

        gpus = torch.cuda.device_count()
        accelerator = Accelerator()

        if not (parallelize or accelerator.num_processes > 1):
            # use user-passed device
            device_list = set(
                ["cuda", "cpu"]
                + [f"cuda:{i}" for i in range(torch.cuda.device_count())]
            )
            if device:
                if device not in device_list:
                    device = int(device)
                self._device = torch.device(device)
                eval_logger.info(f"Using device '{device}'")
            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")
                )
        else:
            eval_logger.info(
                f"Using `accelerate launch` or `parallelize=True`, device '{device}' will be overridden when placing model."
            )
            # TODO: include in warning that `load_in_8bit` etc. affect this too
            self._device = device

        model_kwargs = {}
        if parallelize:
            model_kwargs = _get_accelerate_args(
                device_map_option,
                max_memory_per_gpu,
                max_cpu_memory,
                offload_folder,
            )

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

        # get config
        self._config = transformers.AutoConfig.from_pretrained(
            pretrained,
            revision=revision,
        )

        if getattr(self._config, "model_type") in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
            self.AUTO_MODEL_CLASS = transformers.AutoModelForCausalLM
        else:
            self.AUTO_MODEL_CLASS = transformers.AutoModelForSeq2SeqLM

        assert self.AUTO_MODEL_CLASS in [
            transformers.AutoModelForCausalLM,
            transformers.AutoModelForSeq2SeqLM,
        ]

        self._model = self.AUTO_MODEL_CLASS.from_pretrained(
            pretrained,
            revision=revision,
            low_cpu_mem_usage=low_cpu_mem_usage,
            **model_kwargs,
            torch_dtype=utils.get_dtype(dtype),
        )
        # forever after, access self._model through self.model property
        self.model.eval()
        self.model.tie_weights()
        if gpus <= 1 and not parallelize:
            # place model onto device, if not using HF Accelerate in any form
            self.model.to(self.device)

        self.tokenizer = transformers.AutoTokenizer.from_pretrained(
            pretrained if tokenizer is None else tokenizer,
            revision=revision,
        )

        self.vocab_size = self.tokenizer.vocab_size
        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id

        self._max_length = max_length

        # multithreading and batching
        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
            else:
                if gpus > accelerator.num_processes:
                    # TODO: make sure there's still never an edge case where we unintentionally default to CPU
                    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

                self._model = accelerator.prepare(self.model)
                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")
                # 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")
                # )
                # self.model.to(self.device)
            # else:
            #     self._model = accelerator.prepare(self.model)
            #     self._device = torch.device(f"cuda:{accelerator.local_process_index}")
            #     self.accelerator = accelerator

            #     self._rank = self.accelerator.local_process_index
            #     self._world_size = self.accelerator.num_processes

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

    @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

    @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):
        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

    @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):
        """ """
        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)

        # left-truncate the encoded context to be at most `left_truncate_len` tokens long
        if left_truncate_len:
            encoding = encoding[-left_truncate_len:]

        return encoding

    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"]

    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):
        """
        :param inps: torch.Tensor
            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():
            if attn_mask is not None or labels is not None:
                assert attn_mask is not None and labels is not None
                assert self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM
                return self.model(
                    input_ids=inps, attention_mask=attn_mask, labels=labels
                ).logits
            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]
        )
        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,
        )

    def _select_cont_toks(self, logits, contlen=None, inplen=None):
        if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
            assert (
                contlen and inplen
            ), "Must pass input len and cont. len to select scored logits for causal LM"
            # 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:
            assert (
                contlen and not inplen
            ), "Selecting scored logits for Seq2SeqLM requires only cont. len"
            # only discard right-padding.
            # the logits input to this fn only contain decoder-side tokens.
            logits = logits[:contlen]

        return logits

    def loglikelihood(self, requests):
        new_reqs = []
        for context, continuation in [req.args for req in requests]:
            if context == "":
                # end of text as context
                context_enc = [self.eot_token_id]
            else:
                context_enc = self.tok_encode(context)

            continuation_enc = self.tok_encode(continuation)

            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),
                        prefix_token=self.eot_token_id,
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )

            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
            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

                # how this all works (illustrated on a causal decoder-only setup):
                #          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,
                        device=self.device,
                    )
                    (inplen,) = inp.shape
                elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                    inp = torch.tensor(
                        (context_enc)[-self.max_length :],
                        dtype=torch.long,
                        device=self.device,
                    )
                    (inplen,) = inp.shape

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

                    cont = torch.tensor(
                        (continuation_enc)[-self.max_length :],
                        # TODO: left-shift these?
                        # TODO: our code assumes we never end up truncating conts for either model type
                        dtype=torch.long,
                        device=self.device,
                    )
                    (contlen,) = cont.shape

                    conts.append(cont)

                    padding_len_cont = (
                        max(padding_len_cont, contlen)
                        if padding_len_cont is not None
                        else contlen
                    )

                padding_len_inp = (
                    max(padding_len_inp, inplen)
                    if padding_len_inp is not None
                    else inplen
                )

                inps.append(inp)  # [1, inp_length]
                cont_toks_list.append(continuation_enc)
                inplens.append(inplen)

            # create encoder attn mask and batched conts, if seq2seq
            call_kwargs = {}
            if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
                batched_inps = utils.pad_and_concat(
                    padding_len_inp, inps, padding_side="right"
                )  # [batch, padding_len_inp]
            elif self.AUTO_MODEL_CLASS == transformers.AutoModelForSeq2SeqLM:
                # TODO: left-pad encoder inps and mask?
                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,
                }

            multi_logits = F.log_softmax(
                self._model_call(batched_inps, **call_kwargs), dim=-1
            )  # [batch, padding_length (inp or cont), vocab]

            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)
                # take only logits in the continuation
                # (discard context toks if decoder-only ; discard right-padding)
                ctx_len = (
                    inplen
                    if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM
                    else None
                )
                logits = self._select_cont_toks(logits, contlen=contlen, inplen=ctx_len)
                logits = logits.unsqueeze(0)  # [1, seq, vocab]

                # Check if per-token argmax is exactly equal to continuation
                greedy_tokens = logits.argmax(dim=-1)
                cont_toks = torch.tensor(
                    cont_toks, dtype=torch.long, device=self.device
                ).unsqueeze(
                    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)

                self.cache_hook.add_partial("loglikelihood", cache_key, answer)

        return re_ord.get_original(res)

    def greedy_until(self, requests):
        res = defaultdict(list)
        re_ords = {}

        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 = self.tok_encode(x[0])
            return -len(toks), x[0]

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

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

        # for each different set of kwargs, we execute all requests, by batch.
        for key, re_ord in re_ords.items():
            for chunk in utils.chunks(
                re_ord.get_reordered(),
                self.batch_size,
            ):
                contexts, all_gen_kwargs = zip(*chunk)
                # we assume all gen kwargs in the batch are the same
                # this is safe to assume because the `grouper` object ensures it.
                gen_kwargs = all_gen_kwargs[0]
                # unpack our keyword arguments.
                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(
                                f"Expected `kwargs['until']` to be of type Union[str,list] but got {until}"
                            )
                else:
                    raise ValueError(
                        f"Expected `kwargs` to be of type `dict` but got {kwargs}"
                    )
                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
                (primary_until) = until[0]

                # set the max length in tokens of inputs ("context_enc")
                if self.AUTO_MODEL_CLASS == transformers.AutoModelForCausalLM:
                    # 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

                # encode, pad, and truncate contexts for this batch
                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)

                # perform batched generation
                cont = self._model_generate(
                    context=context_enc,
                    attention_mask=attn_masks,
                    max_length=context_enc.shape[1] + max_gen_toks,
                    stop=primary_until,
                    **kwargs,
                )

                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] :]

                    s = self.tok_decode(cont_toks)

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

                    res[key].append(s)

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

        pbar.close()

        return grouper.get_original(res)