vllm_causallms.py

import copy
from importlib.util import find_spec
from typing import List, Literal, Optional, Tuple, Union

from tqdm import tqdm

from lm_eval.api.instance import Instance
from lm_eval.api.model import TemplateLM
from lm_eval.api.registry import register_model
from lm_eval.models.utils import Collator, divide
from lm_eval.utils import (
    eval_logger,
    get_rolling_token_windows,
    make_disjoint_window,
)


try:
    import ray
    from ray.util.multiprocessing import Pool
    from vllm import LLM, SamplingParams
    from vllm.transformers_utils.tokenizer import get_tokenizer
except ModuleNotFoundError:
    pass

eval_logger = eval_logger


# adapted from https://github.com/vllm-project/vllm/issues/367#issuecomment-1788341727
def run_inference_one_model(
    model_args: dict, sampling_params, requests: List[List[int]]
):
    llm = LLM(**model_args)
    return llm.generate(prompt_token_ids=requests, sampling_params=sampling_params)


@register_model("vllm")
class VLLM(TemplateLM):
    _DEFAULT_MAX_LENGTH = 2048

    def __init__(
        self,
        pretrained="gpt2",
        dtype: Literal["float16", "bfloat16", "float32", "auto"] = "auto",
        revision: Optional[str] = None,
        trust_remote_code: Optional[bool] = False,
        tokenizer: Optional[str] = None,
        tokenizer_mode: Literal["auto", "slow"] = "auto",
        tokenizer_revision: Optional[str] = None,
        add_bos_token: Optional[bool] = False,
        tensor_parallel_size: int = 1,
        quantization: Optional[str] = None,
        max_gen_toks: int = 256,
        swap_space: int = 4,
        batch_size: Union[str, int] = 1,
        max_batch_size=None,
        max_length: int = None,
        max_model_len: int = None,
        seed: int = 1234,
        gpu_memory_utilization: float = 0.9,
        device: str = "cuda",
        data_parallel_size: int = 1,
    ):
        super().__init__()

        if not find_spec("vllm"):
            raise Exception(
                "attempted to use 'vllm' LM type, but package `vllm` is not installed. "
                "Please install vllm via `pip install lm-eval[vllm]` or `pip install -e .[vllm]`"
            )

        assert "cuda" in device or device is None, "vLLM only supports CUDA"
        assert (
            max_length is None or max_model_len is None
        ), "Either max_length or max_model_len may be provided, but not both"

        self._max_length = max_model_len if max_model_len is not None else max_length
        self.tensor_parallel_size = int(tensor_parallel_size)
        self.data_parallel_size = int(data_parallel_size)
        self.model_args = {
            "model": pretrained,
            "gpu_memory_utilization": float(gpu_memory_utilization),
            "revision": revision,
            "dtype": dtype,
            "tokenizer": tokenizer,
            "tokenizer_mode": tokenizer_mode,
            "tokenizer_revision": tokenizer_revision,
            "trust_remote_code": trust_remote_code,
            "tensor_parallel_size": int(tensor_parallel_size),
            "max_model_len": int(self._max_length) if self._max_length else None,
            "swap_space": int(swap_space),
            "quantization": quantization,
            "seed": int(seed),
        }
        self.batch_size = (
            "auto"
            if isinstance(batch_size, str) and "auto" in batch_size
            else batch_size
        )
        if self.data_parallel_size <= 1:
            self.model = LLM(**self.model_args)
        else:
            self.model_args["worker_use_ray"] = True
            self.batch_size = "auto"
            eval_logger.info("Manual batching is not compatible with data parallelism.")

            from transformers import AutoConfig

            self._config = AutoConfig.from_pretrained(
                pretrained, trust_remote_code=trust_remote_code, revision=revision
            )
        self.tokenizer = get_tokenizer(
            tokenizer if tokenizer else pretrained,
            tokenizer_mode=tokenizer_mode,
            trust_remote_code=trust_remote_code,
            tokenizer_revision=tokenizer_revision,
        )
        self.add_bos_token = add_bos_token

        self._max_gen_toks = max_gen_toks

    @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
        if self.data_parallel_size <= 1:
            return self.model.llm_engine.model_config.max_model_len
        else:
            seqlen_config_attrs = ("n_positions", "max_position_embeddings", "n_ctx")
            for attr in seqlen_config_attrs:
                if hasattr(self._config, attr):
                    return getattr(self._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 self._max_gen_toks

    def tok_encode(
        self,
        string: str,
        left_truncate_len=None,
        add_special_tokens=None,
        truncation=False,
    ):
        """ """
        if not add_special_tokens:
            add_special_tokens = False or self.add_bos_token
        encoding = self.tokenizer.encode(
            string, add_special_tokens=add_special_tokens, truncation=truncation
        )

        # 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 _model_generate(
        self,
        requests: List[List[int]] = None,
        generate: bool = False,
        max_tokens: int = None,
        stop: Optional[List[str]] = None,
        **kwargs,
    ):
        if generate:
            kwargs = self.modify_gen_kwargs(kwargs)
            sampling_params = SamplingParams(max_tokens=max_tokens, stop=stop, **kwargs)
        else:
            sampling_params = SamplingParams(
                temperature=0, prompt_logprobs=1, max_tokens=1
            )
        if self.data_parallel_size > 1:
            requests = [list(x) for x in divide(requests, self.data_parallel_size)]
            inputs = [(self.model_args, sampling_params, req) for req in requests]

            with Pool(self.data_parallel_size) as pool:
                results = pool.starmap(run_inference_one_model, inputs)
            # Invoke ray.shutdown() to prevent hang-ups if subsequent calls required.
            ray.shutdown()
            # flatten results
            return [item for sublist in results for item in sublist]

        outputs = self.model.generate(
            prompt_token_ids=requests,
            sampling_params=sampling_params,
            use_tqdm=True if self.batch_size == "auto" else False,
        )
        return outputs

    def loglikelihood_rolling(self, requests: List[Instance]) -> List[float]:
        loglikelihoods = []

        for (string,) in tqdm([req.args for req in requests]):
            rolling_token_windows = list(
                map(
                    make_disjoint_window,
                    get_rolling_token_windows(
                        token_list=self.tok_encode(string),
                        prefix_token=self.eot_token_id,
                        max_seq_len=self.max_length - 1,
                        context_len=1,
                    ),
                )
            )

            rolling_token_windows = [(None,) + x for x in rolling_token_windows]

            string_nll = self._loglikelihood_tokens(
                rolling_token_windows,
            )

            # discard is_greedy
            string_nll = [x[0] for x in string_nll]

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

    def generate_until(self, requests: List[Instance]) -> List[str]:
        res = []

        # batch tokenize contexts
        context, all_gen_kwargs = zip(*(req.args for req in requests))
        context_encoding = self.tokenizer(context, add_special_tokens=False).input_ids
        requests = [
            ((a, b), c) for a, b, c in zip(context, context_encoding, all_gen_kwargs)
        ]

        def _collate_gen(_requests):
            # 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
            return -len(_requests[0][1]), _requests[0][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.
        re_ords = Collator(requests, _collate_gen, group_by="gen_kwargs")
        chunks = re_ords.get_batched(
            n=int(self.batch_size) if self.batch_size != "auto" else 0, batch_fn=None
        )

        pbar = tqdm(
            total=len(requests),
            disable=(self.rank != 0),
            desc="Running generate_until requests",
        )
        # for each different set of kwargs, we execute all requests, by batch.
        for chunk in chunks:
            context_and_encoding, all_gen_kwargs = zip(*chunk)
            context, context_encoding = zip(*context_and_encoding)
            # 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 = [until]
                    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 {gen_kwargs}"
                )
            if not until:
                until = [self.tokenizer.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

            # set the max length in tokens of inputs ("context_enc")
            # max len for inputs = max length, minus room to generate the max new tokens
            max_ctx_len = self.max_length - max_gen_toks
            context_encoding = [x[-max_ctx_len:] for x in context_encoding]

            # perform batched generation
            cont = self._model_generate(
                requests=context_encoding,
                generate=True,
                max_tokens=max_gen_toks,
                stop=until,
                **kwargs,
            )

            # cache generations
            for output, context in zip(cont, context):
                generated_text = output.outputs[0].text
                res.append(generated_text)
                self.cache_hook.add_partial(
                    "generate_until", (context, gen_kwargs), generated_text
                )
                pbar.update(1)

        pbar.close()
        # reorder all group of results back to original unsorted form
        return re_ords.get_original(res)

    def _loglikelihood_tokens(
        self,
        requests: List[Tuple[Tuple[str, str], List[int], List[int]]],
        disable_tqdm: bool = False,
    ) -> List[Tuple[float, bool]]:
        res = []

        def _collate(x):
            toks = x[1] + x[2]
            return -len(toks), tuple(toks)

        # Reorder requests by length and batch
        re_ord = Collator(requests, sort_fn=_collate)
        chunks = re_ord.get_batched(
            n=int(self.batch_size) if self.batch_size != "auto" else 0, batch_fn=None
        )

        pbar = tqdm(
            total=len(requests),
            disable=disable_tqdm,
            desc="Running loglikelihood requests",
        )
        for chunk in chunks:
            inputs = []
            ctxlens = []
            for cache_key, context_enc, continuation_enc in chunk:
                inp = (context_enc + continuation_enc)[-(self.max_length) :]
                ctxlen = len(context_enc) - max(
                    0, len(context_enc) + len(continuation_enc) - (self.max_length)
                )

                inputs.append(inp)
                ctxlens.append(ctxlen)

            outputs = self._model_generate(requests=inputs, generate=False)

            for output, ctxlen, (cache_key, _, _), inp in zip(
                outputs, ctxlens, chunk, inputs
            ):
                answer = self._parse_logprobs(
                    tokens=inp,
                    outputs=output,
                    ctxlen=ctxlen,
                )

                res.append(answer)

                # partial caching
                if cache_key is not None:
                    self.cache_hook.add_partial("loglikelihood", cache_key, answer)
                pbar.update(1)
        pbar.close()
        return re_ord.get_original(res)

    @staticmethod
    def _parse_logprobs(tokens: List, outputs, ctxlen: int) -> Tuple[float, bool]:
        """Process logprobs and tokens.

        :param tokens: list
            Input tokens (potentially left-truncated)
        :param outputs: RequestOutput
            Contains prompt_logprobs
        :param ctxlen: int
            Length of context (so we can slice them away and only keep the predictions)
        :return:
            continuation_logprobs: float
                Log probabilities of continuation tokens
            is_greedy: bool
                Whether argmax matches given continuation exactly
        """

        # The first entry of prompt_logprobs is None because the model has no previous tokens to condition on.
        continuation_logprobs_dicts = outputs.prompt_logprobs

        # Calculate continuation_logprobs
        # assume ctxlen always >= 1
        continuation_logprobs = sum(
            logprob_dict.get(token)
            for token, logprob_dict in zip(
                tokens[ctxlen:], continuation_logprobs_dicts[ctxlen:]
            )
        )

        # Determine if is_greedy
        is_greedy = True
        for token, logprob_dict in zip(
            tokens[ctxlen:], continuation_logprobs_dicts[ctxlen:]
        ):
            # Get the token with the maximum log probability from the logprob_dict
            if logprob_dict:  # Ensure the logprob_dict is not None
                top_token = max(logprob_dict, key=logprob_dict.get)
                if top_token != token:
                    is_greedy = False
                    break

        return continuation_logprobs, is_greedy

    @staticmethod
    def modify_gen_kwargs(kwargs: dict) -> dict:
        # sampling_params
        do_sample = kwargs.pop("do_sample", None)
        if do_sample is False or "temperature" not in kwargs:
            kwargs["temperature"] = 0.0
        # hf defaults
        kwargs["skip_special_tokens"] = kwargs.get("skip_special_tokens", False)
        kwargs["spaces_between_special_tokens"] = kwargs.get(
            "spaces_between_special_tokens", False
        )
        return kwargs