import abc import hashlib import json import logging import os from collections.abc import Iterable from typing import TYPE_CHECKING, Any, Optional, TypeVar from tqdm import tqdm from lm_eval import utils if TYPE_CHECKING: from sqlitedict import SqliteDict from lm_eval.api.instance import Instance eval_logger = logging.getLogger(__name__) T = TypeVar("T", bound="LM") class LM(abc.ABC): def __init__(self) -> None: """Defines the interface that should be implemented by all LM subclasses. LMs are assumed to take text (strings) as input and yield strings as output (inputs/outputs should be tokenization-agnostic.) """ # set rank and world size to a single process, by default. self._rank = 0 self._world_size = 1 self.cache_hook: CacheHook = CacheHook(None) @abc.abstractmethod def loglikelihood(self, requests) -> list[tuple[float, bool]]: """Compute log-likelihood of generating a continuation from a context. Downstream tasks should attempt to use loglikelihood instead of other LM calls whenever possible. :param requests: list[Instance] A list of Instance objects, with property `args` which returns a tuple (context, continuation). `context: str` Context string. Implementations of LM must be able to handle an empty context string. `continuation: str` The continuation over which log likelihood will be calculated. If there is a word boundary, the space should be in the continuation. For example, context="hello" continuation=" world" is correct. :return: list[tuple[float, bool]] A list of pairs (logprob, isgreedy) `logprob: float` The log probability of `continuation`. `isgreedy`: Whether `continuation` would be generated by greedy sampling from `context`. """ pass @abc.abstractmethod def loglikelihood_rolling(self, requests) -> list[float]: """Compute full log-likelihood of a string, with no truncation, for perplexity computation - We will use the full max context length of the model. - For inputs that exceed the max context length, we divide the tokenized string into chunks of up to the max context length. - IMPORTANT: Each document's loglikelihood/perplexity is computed *separately*, unlike other implementations which may simply concatenate multiple documents together. - IMPORTANT: We maximize the amount of context for each prediction. Specifically, for inputs that we break into multiple chunks, the last input will still a full-sized context. Example: Input tokens: [ 0 1 2 3 4 5 6 7 8 9 ] Prefix: BOS/EOS Max context length: 4 Resulting input/prediction pairs: INPUT: BOS 0 1 2 PRED: 0 1 2 3 INPUT: 3 4 5 6 PRED: 4 5 6 7 INPUT: 5 6 7 8 PRED: 8 9 Observe that: 1. Each token is predicted exactly once 2. For the last pair, we provide the full context, but only score the last two tokens :param requests: list[Instance] A list of Instance objects with property `args` which returns a tuple (context,). string: str String for which we are computing overall loglikelihood :return: list[tuple[float]] A list of tuples (logprob,) logprob: float The log probability of `context` conditioned on the BOS/EOS token. Can also be overridden for custom cases by `prefix_token_id`. """ pass # TODO: Add an optional max length @abc.abstractmethod def generate_until(self, requests) -> list[str]: """Generate greedily until a stopping sequence :param requests: list[Instance] A list of Instance objects with property `args` which returns a tuple (context, gen_kwargs). context: str Context string gen_kwargs: dict A dictionary of keyword arguments to pass to the generation function e.g. top_k, until, etc. :return: list[str] A list of model generated continuations. continuation: str The generated continuation. """ pass def apply_chat_template( self, chat_history: list[dict[str, str]], add_generation_prompt=True ) -> str: """ Defines how to transform few-shot examples provided as chat history into a format that can be used as input to the LM. :param chat_history: list[dict[str, str]] A list of dictionaries with keys 'role' and 'content'. Values are strings representing the role name and the content of the message, respectively. :param add_generation_prompt: bool Whether to append an assistant gen prefix (for e.g. <|assistant|>) to the assistant messages in the chat history. False if prefilling an assistant message. :return: str A string representing the chat history in a format that can be used as input to the LM. """ raise NotImplementedError( "To use this model with chat templates, please implement the 'apply_chat_template' method for your model type." ) @classmethod def create_from_arg_string( cls: type[T], arg_string: str, additional_config: dict | None = None ) -> T: """ Creates an instance of the LM class using the given argument string and additional config. Parameters: - arg_string: A string containing arguments in the format key1=value1,key2=value2. - additional_config: Optional dictionary containing additional configuration parameters. Returns: - Instance of the LM class. """ additional_config = {} if additional_config is None else additional_config args = utils.simple_parse_args_string(arg_string) args2 = {k: v for k, v in additional_config.items() if v is not None} return cls(**args, **args2) @classmethod def create_from_arg_obj( cls: type[T], arg_dict: dict, additional_config: dict | None = None ) -> T: """ Creates an instance of the LM class using the given arg_obj Parameters: - arg_obj: A dict containing arguments in the format key1=value1,key2=value2. - additional_config: Optional dictionary containing additional configuration parameters. Returns: - Instance of the LM class. """ additional_config = ( {} if additional_config is None else {k: v for k, v in additional_config.items() if v is not None} ) return cls(**arg_dict, **additional_config) @property def rank(self): # used in the case of parallelism. Hardcoded to # ensure no errors arise using API models which do # not support multi-device parallelism nor expect it. return self._rank @property def world_size(self): # used in the case of parallelism. Hardcoded to # ensure no errors arise using API models which do # not support multi-device parallelism nor expect it. return self._world_size @property def tokenizer_name(self) -> str: """Must be defined for LM subclasses which implement Chat Templating. Should return the name of the tokenizer or chat template used. Used only to properly fingerprint caches when requests are being cached with `--cache_requests`, otherwise not used. """ raise NotImplementedError( "To use this model with chat templates, please implement the 'tokenizer_name' property." ) def chat_template(self, chat_template: bool | str = False) -> str | None: """Returns the chat template structure for user/assistant messages if a template is provided. This method is intended to be overridden in a subclass to define a specific chat template format. For models that do not support chat templates, this method returns None by default. """ return "" def set_cache_hook(self, cache_hook: "CacheHook") -> None: self.cache_hook = cache_hook ### SQLite-based caching of LM responses def hash_args(attr: str, args: Iterable[Any]) -> str: dat = json.dumps([attr] + list(args)) return hashlib.sha256(dat.encode("utf-8")).hexdigest() class CacheHook: def __init__(self, cachinglm: Optional["CachingLM"]) -> None: if cachinglm is None: self.dbdict: SqliteDict | None = None return self.dbdict = cachinglm.dbdict def add_partial(self, attr: str, req: Iterable[Any], res: Any) -> None: if self.dbdict is None: return hsh = hash_args(attr, req) self.dbdict[hsh] = res class CachingLM: def __init__(self, lm: LM, cache_db: str) -> None: """LM wrapper that returns cached results if they exist, and uses the underlying LM if not. :param lm: LM Underlying LM :param cache_db: str Path to cache db """ from sqlitedict import SqliteDict self.lm: LM = lm self.cache_db: str = cache_db if os.path.dirname(cache_db): os.makedirs(os.path.dirname(cache_db), exist_ok=True) self.dbdict = SqliteDict(cache_db, autocommit=True) # add hook to lm lm.set_cache_hook(self.get_cache_hook()) def __getattr__(self, attr: str) -> Any: lm_attr = getattr(self.lm, attr) if attr not in ["loglikelihood", "loglikelihood_rolling", "generate_until"]: eval_logger.debug(f"Passing through attribute '{attr}' to underlying LM") return lm_attr def _fn(requests: list["Instance"]) -> list["Instance"]: res = [] remaining_reqs = [] warned = False # figure out which ones are cached and which ones are new eval_logger.info( f"Loading '{attr}' responses from cache '{self.cache_db}' where possible..." ) for req in tqdm(requests, desc="Checking cached requests"): hsh = hash_args(attr, req.args) if attr == "generate_until" and req.args[1].get("do_sample", False): # when we are doing non-greedy generation, don't use the cache # (else every "randomly sampled" generation would be identical for repeats > 1). if not warned: eval_logger.warning( f"Arguments to lm.generate_until() '{req.args[1]}' include non-deterministic sampling. Caching will not be performed for such requests." ) warned = True res.append(None) remaining_reqs.append(req) elif hsh in self.dbdict: ob = self.dbdict[hsh] assert ob is not None res.append(ob) else: res.append(None) remaining_reqs.append(req) eval_logger.info( f"Cached requests: {len(requests) - len(remaining_reqs)}, Requests remaining: {len(remaining_reqs)}" ) # actually run the LM on the requests that do not have cached results rem_res = getattr(self.lm, attr)(remaining_reqs) if remaining_reqs else [] # stick the new ones back into the list and also cache any of the new ones resptr = 0 for req, r in zip(remaining_reqs, rem_res, strict=True): while res[resptr] is not None: resptr += 1 res[resptr] = r # caching hsh = hash_args(attr, req.args) self.dbdict[hsh] = r self.dbdict.commit() return res return _fn def get_cache_hook(self) -> "CacheHook": return CacheHook(self) class TemplateLM(LM): """ A class acting as intermediary between the LM base class and boilerplate often included in other LM subclasses. """ tokenizer = None backend = "causal" @property @abc.abstractmethod def eot_token_id(self) -> int: pass @property def prefix_token_id(self): # it is used as prefix for loglikelihood return self.eot_token_id @abc.abstractmethod def tok_encode( self, string: str, add_special_tokens: bool | None = None, **kwargs ) -> list[int]: """ Tokenize a string using the model's tokenizer and return a list of token IDs. NOTE: This method is expected to handle strings which already contain the BOS token (when add_special_tokens=None). Otherwise, will use add_special_tokens if specified. """ pass @abc.abstractmethod def _loglikelihood_tokens( self, requests: list[tuple[tuple[str, str], list[int], list[int]]], **kwargs ) -> list[tuple[float, bool]]: pass def _encode_pair( self, context: str, continuation: str ) -> tuple[list[int], list[int]]: """ Encode a context-continuation pair into separate token ID lists. This method handles the tokenization of context and continuation strings while preserving proper boundary handling. Trailing spaces in the context are moved to the beginning of the continuation to ensure correct tokenization at the word boundary. For Seq2Seq models (encoder-decoder), context and continuation are encoded separately. For other model types (decoder-only), the full sequence is encoded together to ensure proper tokenization, then split at the context boundary. :param context: str The context string. Can be empty (will be handled by the caller). :param continuation: str The continuation string to be scored. :return: tuple[list[int], list[int]] A tuple of (context_enc, continuation_enc) where: - context_enc: Token IDs for the context - continuation_enc: Token IDs for the continuation Note: This method does NOT handle empty context. The caller should handle empty context (see loglikelihood method). """ assert context, "Context cannot be empty!" n_spaces = len(context) - len(context.rstrip()) if n_spaces > 0: continuation = context[-n_spaces:] + continuation context = context[:-n_spaces] if self.backend == "causal": 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:] else: # for SEQ2SEQ case we need to encode separately context_enc = self.tok_encode(context) continuation_enc = self.tok_encode(continuation, add_special_tokens=False) return context_enc, continuation_enc def loglikelihood( self, requests: list["Instance"], disable_tqdm: bool = False ) -> list[tuple[float, bool]]: """ Compute log-likelihood of generating continuations from contexts. This is the concrete implementation for TemplateLM and its subclasses. It tokenizes context-continuation pairs and delegates scoring to _loglikelihood_tokens. **IMPORTANT**: This method is expected to handle empty context strings. When context is empty (""), it uses the model's prefix_token_id (typically BOS or EOS token) as context. If the continuation already starts with the prefix token, it reuses that token as context instead of duplicating it. :param requests: list[Instance] List of Instance objects with property `args` returning (context, continuation) tuples. :param disable_tqdm: bool Whether to disable the progress bar in _loglikelihood_tokens. :return: list[tuple[float, bool]] List of (log_prob, is_greedy) tuples for each request. Implementation details: - Empty context: Uses prefix_token_id (BOS/EOS) as context - Non-empty context: Uses _encode_pair for proper tokenization - Avoids token duplication when continuation starts with prefix_token_id """ new_reqs = [] for context, continuation in [req.args for req in requests]: if context == "": continuation_enc = self.tok_encode( continuation, add_special_tokens=False ) # BOS or EOS as context: handle when context is empty -> (context + continuation) -> (BOS + continuation context_enc, continuation_enc = ( ([self.prefix_token_id], continuation_enc) if self.prefix_token_id != continuation_enc[0] else (continuation_enc[:1], continuation_enc[1:]) ) # BOS or EOS as context 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, disable_tqdm=disable_tqdm) @abc.abstractmethod def loglikelihood_rolling( self, requests, disable_tqdm: bool = False ) -> list[float]: pass @abc.abstractmethod def generate_until(self, requests, disable_tqdm: bool = False) -> list[str]: pass def chat_template(self, chat_template: bool | str = False) -> str | None: """ Set and get the appropriate chat template for the model. This method sets the tokenizer's chat_template and returns the template string for reproducibility. The template selection logic is adapted from the Transformers library's `apply_chat_template` method in the Tokenizer class. The original implementation can be found at: https://github.com/huggingface/transformers/blob/fc35907f95459d7a6c5281dfadd680b6f7b620e3/src/transformers/tokenization_utils_base.py#L1687 This method ensures that the right template is chosen based on the following: 0. If the model has no 'tokenizer' attribute: assumes that there is only a single possible chat template, handled on the model provider side internally. Returns the empty string. 1. If the model's tokenizer has multiple templates: a. Use the specified template if it exists in the dictionary. b. Use the default template from the list if no specific template is provided. c. Raise an error if no default template exists and no specific template is provided. 2. If the model's tokenizer has a single template or no template: a. Use the tokenizer's chat template if available. b. Fall back to the default chat template if no tokenizer chat template exists. Args: chat_template (Union[bool, str]): Specifies the chat template to use. - If False or None, no template is applied. - If True, the default or only available template is used. - If a string, the template with the matching name is used. Returns: Optional[str]: The selected chat template, or None if no template is applied. """ if self.tokenizer is None: return "" if chat_template is False or chat_template is None: eval_logger.warning( "model.chat_template was called with the chat_template set to False or None. " "Therefore no chat template will be applied. Make sure this is an intended behavior." ) return None # Convert boolean chat_template to None to ensure compatibility with the adapted logic if isinstance(chat_template, bool): chat_template = None using_default_template = False # First, handle the cases when the model has a dict of multiple templates try: template = ( self.tokenizer.chat_template or self.tokenizer.default_chat_template ) except AttributeError: return None if isinstance(template, dict): using_default_dict = self.tokenizer.chat_template is None if chat_template is not None: if chat_template in template: selected_template = template[chat_template] if using_default_dict: using_default_template = True else: raise ValueError( f"The specified chat template '{chat_template}' is not available. " f"Available template names are {sorted(template.keys())}." ) else: # If user didn't pass a chat template, use the default template from the dict if "default" in template: selected_template = template["default"] using_default_template = True else: raise ValueError( "This model has multiple chat templates with no default specified! Please either pass a chat " "template or the name of the template you wish to use to the `chat_template` argument. Available " f"template names are {sorted(template.keys())}." ) # Cases when the model has a single template or no template else: # priority: `chat_template` argument > `tokenizer.chat_template` > `tokenizer.default_chat_template if isinstance(chat_template, str): eval_logger.warning( "Chat template name provided, but the tokenizer's chat template is not a dictionary. " "Using the tokenizer's chat template or the default template instead." ) if self.tokenizer.chat_template is not None: selected_template = self.tokenizer.chat_template else: selected_template = self.tokenizer.default_chat_template using_default_template = True if using_default_template: eval_logger.warning( "No chat template is set for this tokenizer, falling back to a default class-level template. This is " "very error-prone, because models are often trained with templates different from the class default! " "Default chat templates are a legacy feature and will be removed in Transformers v4.43, at which " "point any code depending on them will stop working. We recommend setting a valid chat template before " "then to ensure that this model continues working without issues." ) return selected_template