openai_completions.py

import copy
import os
import time
from collections import defaultdict
from importlib.util import find_spec
from typing import List, Optional, Tuple

import transformers
from tqdm import tqdm

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


def get_result(response, ctxlen: int) -> Tuple[float, bool]:
    """Process results from OpenAI API response.

    :param response: dict
        OpenAI API Response
    :param ctxlen: int
        Length of context (so we can slice them away and only keep the predictions)
    :return:
        continuation_logprobs: np.array
            Log probabilities of continuation tokens
        is_greedy: bool
            whether argmax matches given continuation exactly
    """
    is_greedy = True
    logprobs = response.logprobs.token_logprobs
    continuation_logprobs = sum(logprobs[ctxlen:])

    for i in range(ctxlen, len(response.logprobs.token_logprobs)):
        token = response.logprobs.token_logprobs[i]
        top_tokens = response.logprobs.top_logprobs[i]
        top_token = max(top_tokens.keys(), key=lambda x: top_tokens[x])
        if top_token != token:
            is_greedy = False
            break

    return continuation_logprobs, is_greedy


def oa_completion(**kwargs):
    """Query OpenAI API for completion.

    Retry with back-off until they respond
    """
    if not find_spec("openai") or not find_spec("tiktoken"):
        raise Exception(
            "attempted to use 'openai' LM type, but package `openai` or `tiktoken` are not installed. "
            "Please install these via `pip install lm-eval[openai]` or `pip install -e .[openai]`"
        )
    else:
        import openai

    backoff_time = 3
    while True:
        try:
            return openai.completions.create(**kwargs)
        except openai.OpenAIError:
            import traceback

            traceback.print_exc()
            time.sleep(backoff_time)
            backoff_time *= 1.5


@register_model("openai-completions")
class OpenaiCompletionsLM(LM):
    REQ_CHUNK_SIZE = 20
    _DEFAULT_MAX_LENGTH = 2048

    def __init__(
        self,
        model: str = "text-davinci-003",
        truncate: bool = False,
        max_gen_toks: int = 256,
        batch_size: int = 1,
        seed: int = 1234,
        max_length: Optional[int] = None,
    ) -> None:
        """

        :param engine: str
            OpenAI API engine (e.g. davinci)
        :param truncate: bool
            Truncate input if too long (if False and input is too long, throw error)
        """
        super().__init__()
        self.seed = seed
        try:
            import openai  # noqa: E401
            import tiktoken
        except ModuleNotFoundError:
            raise Exception(
                "attempted to use 'openai' LM type, but package `openai` or `tiktoken` are not installed. \
    please install these via `pip install lm-eval[openai]` or `pip install -e .[openai]`",
            )
        self.model = model
        self.tokenizer = tiktoken.encoding_for_model(self.model)
        self.vocab_size = self.tokenizer.n_vocab
        self.truncate = truncate
        self.end_of_text_token_id = self.tokenizer.eot_token
        self._max_gen_toks = max_gen_toks
        self._max_length = max_length

        # Read from environment variable OPENAI_API_KEY
        openai.api_key = os.environ["OPENAI_API_KEY"]

    @property
    def eot_token_id(self):
        return self.end_of_text_token_id

    @property
    def max_length(self) -> int:
        if self._max_length:
            return self._max_length
        else:
            return self._DEFAULT_MAX_LENGTH

    @property
    def max_gen_toks(self) -> int:
        return self._max_gen_toks

    @property
    def batch_size(self):
        # Isn't used because we override _loglikelihood_tokens
        raise NotImplementedError()

    @property
    def device(self):
        # Isn't used because we override _loglikelihood_tokens
        raise NotImplementedError()

    def tok_encode(self, string: str) -> List[int]:
        return self.tokenizer.encode(string)

    def tok_decode(self, tokens: List[int]) -> str:
        return self.tokenizer.decode(tokens)

    def _encode_pair(
        self, context: str, continuation: str
    ) -> Tuple[List[int], List[int]]:
        n_spaces = len(context) - len(context.rstrip())
        if n_spaces > 0:
            continuation = context[-n_spaces:] + continuation
            context = context[:-n_spaces]
        whole_enc = self.tok_encode(context + continuation)
        context_enc = self.tok_encode(context)
        context_enc_len = len(context_enc)
        continuation_enc = whole_enc[context_enc_len:]
        return context_enc, continuation_enc

    def loglikelihood(self, requests) -> List[Tuple[float, bool]]:
        new_reqs = []
        for context, continuation in [req.args for req in requests]:
            if context == "":
                # end of text as context
                context_enc, continuation_enc = (
                    [self.eot_token_id],
                    self.tok_encode(continuation),
                )
            else:
                context_enc, continuation_enc = self._encode_pair(context, continuation)

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

        return self._loglikelihood_tokens(new_reqs)

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

        def _collate(x):
            # this doesn't efficiently handle last-token differences yet, but those are kinda annoying because
            # it's not guaranteed that the 100 or so logprobs we get to see actually contain all the continuations
            # we care about, and so we need some kind of backup for when it isn't
            toks = x[1] + x[2]
            return -len(toks), tuple(toks)

        re_ord = utils.Reorderer(requests, _collate)

        for chunk in tqdm(
            list(utils.chunks(re_ord.get_reordered(), self.REQ_CHUNK_SIZE)),
            disable=disable_tqdm,
        ):
            inps = []
            ctxlens = []
            for cache_key, context_enc, continuation_enc in chunk:
                # max_length+1 because the API takes up to 2049 tokens, including the first context token
                inp = (context_enc + continuation_enc)[-(self.max_length + 1) :]
                # TODO: the logic is much simpler if we just look at the length of continuation tokens
                ctxlen = len(context_enc) - max(
                    0, len(context_enc) + len(continuation_enc) - (self.max_length + 1)
                )

                inps.append(inp)
                ctxlens.append(ctxlen)

            response = oa_completion(
                model=self.model,
                prompt=inps,
                echo=True,
                max_tokens=0,
                temperature=0.0,
                logprobs=10,
                seed=self.seed,
            )

            for resp, ctxlen, (cache_key, context_enc, continuation_enc) in zip(
                response.choices, ctxlens, chunk
            ):
                answer = get_result(resp, ctxlen)

                res.append(answer)

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

    def generate_until(self, requests) -> List[str]:
        if not requests:
            return []
        res = []
        requests = [req.args for req in requests]

        def _collate(x):
            toks = self.tok_encode(x[0])
            return len(toks), x[0]

        re_ord = utils.Reorderer(requests, _collate)

        def sameuntil_chunks(xs, size):
            ret = []
            lastuntil = xs[0][1]
            for x in xs:
                if len(ret) >= size or x[1] != lastuntil:
                    yield ret, lastuntil
                    ret = []
                    lastuntil = x[1]
                ret.append(x)

            if ret:
                yield ret, lastuntil

        # todo: more intelligent batching for heterogeneous `until`
        for chunk, request_args in tqdm(
            list(sameuntil_chunks(re_ord.get_reordered(), self.REQ_CHUNK_SIZE))
        ):
            inps = []
            for context, _ in chunk:
                context_enc = self.tok_encode(context)
                inp = context_enc[-(self.max_length - self.max_gen_toks) :]
                inps.append(inp)

            until = request_args.pop("until", ["<|endoftext|>"])
            request_args.pop("do_sample", None)
            request_args["temperature"] = request_args.get("temperature", 0)

            response = oa_completion(
                model=self.model,
                prompt=inps,
                max_tokens=self.max_gen_toks,
                stop=until,
                seed=self.seed,
                **request_args,
            )
            for resp, (context, args_) in zip(response.choices, chunk):
                s = getattr(resp, "text")

                until_ = until

                for term in until_:
                    if len(term) > 0:
                        s = s.split(term)[0]

                # partial caching
                self.cache_hook.add_partial(
                    "generate_until", (context, {"until": until_}), s
                )

                res.append(s)
        return re_ord.get_original(res)

    def _model_call(self, inps):
        # Isn't used because we override _loglikelihood_tokens
        raise NotImplementedError()

    def _model_generate(self, context, max_length, eos_token_id):
        # Isn't used because we override generate_until
        raise NotImplementedError()

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

        for (string,) in tqdm([req.args for req in requests]):
            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]

            string_nll = self._loglikelihood_tokens(
                rolling_token_windows,
                disable_tqdm=True,
            )

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

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


def oa_chat_completion(client, **kwargs):
    """Query OpenAI API for chat completion.

    Retry with back-off until they respond
    """
    if not find_spec("openai") or not find_spec("tiktoken"):
        raise Exception(
            "attempted to use 'openai' LM type, but package `openai` or `tiktoken` are not installed. "
            "Please install these via `pip install lm-eval[openai]` or `pip install -e .[openai]`"
        )
    else:
        import openai

    async def _get_completions(**kwargs):
        chat_completions = await client.chat.completions.create(**kwargs)
        return chat_completions

    backoff_time = 3
    while True:
        try:
            return client.chat.completions.create(**kwargs)
        except openai.OpenAIError:
            import traceback

            traceback.print_exc()
            time.sleep(backoff_time)
            backoff_time *= 1.5


@register_model("openai-chat-completions", "local-chat-completions")
class OpenaiChatCompletionsLM(LM):
    def __init__(
        self,
        model: str = "gpt-3.5-turbo",  # GPT model or Local model using HuggingFace model paths
        base_url: str = None,
        truncate: bool = False,
        revision: Optional[str] = "main",
        trust_remote_code: Optional[bool] = False,
        use_fast_tokenizer: Optional[bool] = True,
        **kwargs,
    ) -> None:
        """

        :param model: str
            Implements an OpenAI-style chat completion API for
            accessing both OpenAI OR locally-hosted models using
            HuggingFace Tokenizer
            OpenAI API model (e.g. gpt-3.5-turbo)
            using the **gen_kwargs passed on init
        :param truncate: bool
            Truncate input if too long (if False and input is too long, throw error)
        """
        super().__init__()
        try:
            import openai  # noqa: E401
            import tiktoken
        except ModuleNotFoundError:
            raise Exception(
                "attempted to use 'openai' LM type, but package `openai` or `tiktoken` are not installed. \
    please install these via `pip install lm-eval[openai]` or `pip install -e .[openai]`",
            )
        self.model = model
        self.base_url = base_url
        self.truncate = truncate

        # if we have a local model, use HF tokenizer over tiktoken
        if self.base_url:
            self.revision = revision
            self.trust_remote_code = trust_remote_code
            self.use_fast_tokenizer = use_fast_tokenizer

            self.tokenizer = transformers.AutoTokenizer.from_pretrained(
                self.model,
                revision=self.revision,
                trust_remote_code=self.trust_remote_code,
                use_fast_tokenizer=self.use_fast_tokenizer,
            )
            self.vocab_size = self.tokenizer.vocab
            self.end_of_text_token_id = self.tokenizer.eos_token
        else:
            self.tokenizer = tiktoken.encoding_for_model(self.model)
            self.vocab_size = self.tokenizer.n_vocab
            self.end_of_text_token_id = self.tokenizer.eot_token

        # Read from environment variable OPENAI_API_KEY
        # Set to EMPTY for local
        if self.base_url:
            self.client = openai.OpenAI(base_url=self.base_url)
        else:
            self.client = openai.OpenAI()  # openai.AsyncOpenAI()

    @property
    def eot_token_id(self):
        return self.end_of_text_token_id

    @property
    def max_length(self) -> int:
        # Note: the OpenAI API supports up to 2049 tokens, with the first token being the first input token
        return 2048

    @property
    def max_gen_toks(self) -> int:
        return 256

    @property
    def batch_size(self):
        # Isn't used because we override _loglikelihood_tokens
        raise NotImplementedError()

    @property
    def device(self):
        # Isn't used because we override _loglikelihood_tokens
        raise NotImplementedError()

    def tok_encode(self, string: str) -> List[int]:
        return self.tokenizer.encode(string)

    def tok_decode(self, tokens: List[int]) -> str:
        return self.tokenizer.decode(tokens)

    def _encode_pair(
        self, context: str, continuation: str
    ) -> Tuple[List[int], List[int]]:
        n_spaces = len(context) - len(context.rstrip())
        if n_spaces > 0:
            continuation = context[-n_spaces:] + continuation
            context = context[:-n_spaces]
        whole_enc = self.tok_encode(context + continuation)
        context_enc = self.tok_encode(context)
        context_enc_len = len(context_enc)
        continuation_enc = whole_enc[context_enc_len:]
        return context_enc, continuation_enc

    def generate_until(self, requests) -> List[str]:
        res = defaultdict(list)
        re_ords = {}

        def _collate(x):
            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)

        def sameuntil_chunks(xs, size):
            ret = []
            lastuntil = xs[0][1]
            for x in xs:
                if len(ret) >= size or x[1] != lastuntil:
                    yield ret, lastuntil
                    ret = []
                    lastuntil = x[1]
                ret.append(x)

            if ret:
                yield ret, lastuntil

        pbar = tqdm(total=len(requests), disable=(self.rank != 0))
        for key, re_ord in re_ords.items():
            # n needs to be 1 because messages in
            # chat completion are not batch but
            # is regarded as a single conversation.
            chunks = utils.chunks(re_ord.get_reordered(), n=1)
            for chunk in chunks:
                contexts, all_gen_kwargs = zip(*chunk)
                inps = [{"role": "user", "content": context} for context in contexts]

                gen_kwargs = all_gen_kwargs[0]
                until = None
                if isinstance(gen_kwargs, dict):
                    kwargs = copy.deepcopy(gen_kwargs)  # edge case for repeats > 1
                    if "do_sample" in kwargs.keys():
                        kwargs.pop("do_sample")
                    if "until" in kwargs.keys():
                        until = kwargs.pop("until")
                        if isinstance(until, str):
                            until = [kwargs]
                        elif not isinstance(until, list):
                            raise ValueError(
                                f"Expected repr(kwargs['until']) to be of type Union[str, list] but got {until}"
                            )
                else:
                    raise ValueError(
                        f"Expected repr(kwargs) to be of type repr(dict) but got {kwargs}"
                    )

                response = oa_chat_completion(
                    client=self.client, messages=inps, model=self.model, **kwargs
                )

                for resp, (context, args_) in zip(response.choices, chunk):
                    s = resp.message.content

                    if until is not None:
                        for term in until:
                            if len(term) > 0:
                                s = s.split(term)[0]

                    res[key].append(s)

                    self.cache_hook.add_partial(
                        "generate_until", (context, {"until": until}), 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)

    def loglikelihood(self, requests):
        raise NotImplementedError("No support for logits.")

    def loglikelihood_rolling(self, requests):
        raise NotImplementedError("No support for logits.")