distillation.py

from __future__ import annotations

import logging
import os
import re
from typing import cast

import numpy as np
from huggingface_hub.hf_api import model_info
from transformers import AutoModel, AutoTokenizer
from transformers.modeling_utils import PreTrainedModel
from transformers.tokenization_utils_fast import PreTrainedTokenizerFast

from model2vec.distill.inference import PCADimType, PoolingMode, create_embeddings, post_process_embeddings
from model2vec.distill.utils import select_optimal_device
from model2vec.model import StaticModel
from model2vec.quantization import DType, quantize_embeddings
from model2vec.tokenizer import clean_and_create_vocabulary, replace_vocabulary, turn_tokens_into_ids
from model2vec.vocabulary_quantization import quantize_vocabulary

logger = logging.getLogger(__name__)


def distill_from_model(
    model: PreTrainedModel,
    tokenizer: PreTrainedTokenizerFast,
    vocabulary: list[str] | None = None,
    device: str | None = None,
    pca_dims: PCADimType = 256,
    sif_coefficient: float | None = 1e-4,
    token_remove_pattern: str | None = r"\[unused\d+\]",
    quantize_to: DType | str = DType.Float16,
    vocabulary_quantization: int | None = None,
    pooling: PoolingMode | str = PoolingMode.MEAN,
) -> StaticModel:
    """
    Distill a staticmodel from a sentence transformer.

    This function creates a set of embeddings from a sentence transformer. It does this by doing either
    a forward pass for all subword tokens in the tokenizer, or by doing a forward pass for all tokens in a passed vocabulary.

    If you pass through a vocabulary, we create a custom word tokenizer for that vocabulary.
    If you don't pass a vocabulary, we use the model's tokenizer directly.

    :param model: The model to use.
    :param tokenizer: The tokenizer to use.
    :param vocabulary: The vocabulary to use. If this is None, we use the model's vocabulary.
    :param device: The device to use.
    :param pca_dims: The number of components to use for PCA.
        If this is None, we don't apply PCA.
        If this is 'auto', we don't reduce dimensionality, but still apply PCA.
    :param sif_coefficient: The SIF coefficient to use. If this is None, no weighting is applied.
        Should be a value > 0 and < 1.0. A value of 1e-4 is a good default.
    :param token_remove_pattern: If this is set to a string, we compile this into a regex. Any tokens that conform to this regex pattern will be removed from the vocabulary.
        If the pattern is so general that it removes all tokens, we throw an error. If the pattern can't be compiled into a valid regex, we also throw an error.
    :param quantize_to: The data type to quantize to. Can be any of the DType enum members or their string equivalents.
    :param vocabulary_quantization: The number of clusters to use for vocabulary quantization. If this is None, no quantization is performed.
    :param pooling: The pooling mode to use for creating embeddings. Can be one of:
        'mean' (default): mean over all tokens. Robust and works well in most cases.
        'last': use the last token's hidden state (often the [EOS] token). Common for decoder-style models.
        'first': use the first token's hidden state ([CLS] token in BERT-style models).
        'pooler': use the pooler output (if available). This is often a non-linear projection of the [CLS] token.
    :return: A StaticModel.
    :raises: ValueError if the vocabulary is empty after preprocessing.

    """
    quantize_to = DType(quantize_to)
    backend_tokenizer = tokenizer.backend_tokenizer
    sif_coefficient, token_remove_regex = _validate_parameters(sif_coefficient, token_remove_pattern)

    if vocabulary is None:
        vocabulary = []

    device = select_optimal_device(device)

    n_tokens_before = len(vocabulary)
    # Clean the vocabulary by removing duplicate tokens and tokens that are in the internal vocabulary.
    all_tokens, backend_tokenizer = clean_and_create_vocabulary(
        tokenizer, vocabulary, token_remove_regex=token_remove_regex
    )
    n_tokens_after = len([token for token in all_tokens if not token.is_internal])
    if n_tokens_before:
        logger.info(
            f"Adding {n_tokens_after} tokens to the vocabulary. Removed {n_tokens_before - n_tokens_after} tokens during preprocessing."
        )

    if not all_tokens:
        raise ValueError("The vocabulary is empty after preprocessing. Please check your token_remove_pattern.")

    unk_token = cast(str | None, tokenizer.special_tokens_map.get("unk_token"))
    pad_token = cast(str | None, tokenizer.special_tokens_map.get("pad_token"))

    # Weird if to satsify mypy
    if pad_token is None:
        if unk_token is not None:
            pad_token = unk_token
            logger.warning(
                "The pad token is not set. Setting it to the unk token. This is a workaround for models that don't have a pad token."
            )
        else:
            pad_token = unk_token or all_tokens[0].form
            logger.warning(
                "The pad token is not set. Setting it to the first token in the vocabulary. This is a workaround for models that don't have a pad token."
            )

    # Replace the vocabulary in the tokenizer with the new vocabulary.
    backend_tokenizer = replace_vocabulary(backend_tokenizer, all_tokens, unk_token=unk_token, pad_token=pad_token)
    logger.info(f"Creating embeddings for {len(all_tokens)} tokens")
    # Convert tokens to IDs
    token_ids = turn_tokens_into_ids(all_tokens, tokenizer, unk_token)

    # Create the embeddings
    embeddings = create_embeddings(
        tokenized=token_ids,
        model=model,
        device=device,
        pad_token_id=tokenizer.get_vocab()[pad_token],
        pooling=pooling,
    )

    if vocabulary_quantization is not None:
        _, weights = post_process_embeddings(np.asarray(embeddings), None, sif_coefficient=sif_coefficient)
        embeddings, token_mapping, weights = quantize_vocabulary(
            n_clusters=vocabulary_quantization, weights=weights, embeddings=np.asarray(embeddings)
        )
        embeddings, _ = post_process_embeddings(embeddings, pca_dims, sif_coefficient=sif_coefficient)
    else:
        # Post-process the embeddings.
        embeddings, weights = post_process_embeddings(np.asarray(embeddings), pca_dims, sif_coefficient=sif_coefficient)
        token_mapping = None
    # Quantize the embeddings.
    embeddings = quantize_embeddings(embeddings, quantize_to)

    model_name = getattr(model, "name_or_path", "")

    config = {
        "model_type": "model2vec",
        "architectures": ["StaticModel"],
        "tokenizer_name": model_name,
        "apply_pca": pca_dims,
        "sif_coefficient": sif_coefficient,
        "hidden_dim": embeddings.shape[1],
        "seq_length": 1000000,  # Set this to a high value since we don't have a sequence length limit.
        "normalize": True,
        "pooling": pooling,
    }

    if os.path.exists(model_name):
        # Using a local model. Get the model name from the path.
        model_name = os.path.basename(model_name)
        language = None
    else:
        # Get the language from the model card.
        try:
            info = model_info(model_name)
            language = info.cardData.get("language", None) if info.cardData is not None else None
        except Exception as e:
            # NOTE: bare except because there's many reasons this can fail.
            logger.warning(f"Couldn't get the model info from the Hugging Face Hub: {e}. Setting language to None.")
            language = None

    return StaticModel(
        vectors=embeddings,
        weights=weights,
        token_mapping=token_mapping,
        tokenizer=backend_tokenizer,
        config=config,
        base_model_name=model_name,
        language=language,
        normalize=True,
    )


def _validate_parameters(
    sif_coefficient: float | None,
    token_remove_pattern: str | None,
) -> tuple[float | None, re.Pattern | None]:
    """
    Validate the parameters passed to the distillation function.

    :param sif_coefficient: The SIF coefficient to use. If this is None, no weighting is applied.
        Should be a value >= 0 and < 1.0. A value of 1e-4 is a good default.
    :param token_remove_pattern: If this is set to a string, we compile this into a regex. Any tokens that conform to this regex pattern will be removed from the vocabulary.
    :return: The SIF coefficient to use.
    :raises: ValueError if the regex can't be compiled.

    """
    if sif_coefficient is not None:
        if not 0 < sif_coefficient < 1.0:
            raise ValueError("SIF coefficient must be a value > 0 and < 1.0.")

    token_remove_regex: re.Pattern | None = None
    if token_remove_pattern is not None:
        try:
            token_remove_regex = re.compile(token_remove_pattern)
        except re.error as e:
            raise ValueError(f"Couldn't compile the regex pattern: {e}")

    return sif_coefficient, token_remove_regex


def distill(
    model_name: str,
    vocabulary: list[str] | None = None,
    device: str | None = None,
    pca_dims: PCADimType = 256,
    sif_coefficient: float | None = 1e-4,
    token_remove_pattern: str | None = r"\[unused\d+\]",
    trust_remote_code: bool = False,
    quantize_to: DType | str = DType.Float16,
    vocabulary_quantization: int | None = None,
    pooling: PoolingMode | str = PoolingMode.MEAN,
) -> StaticModel:
    """
    Distill a staticmodel from a sentence transformer.

    This function creates a set of embeddings from a sentence transformer. It does this by doing either
    a forward pass for all subword tokens in the tokenizer, or by doing a forward pass for all tokens in a passed vocabulary.

    If you pass through a vocabulary, we create a custom word tokenizer for that vocabulary.
    If you don't pass a vocabulary, we use the model's tokenizer directly.

    :param model_name: The model name to use. Any sentencetransformer compatible model works.
    :param vocabulary: The vocabulary to use. If this is None, we use the model's vocabulary.
    :param device: The device to use.
    :param pca_dims: The number of components to use for PCA.
        If this is None, we don't apply PCA.
        If this is 'auto', we don't reduce dimenionality, but still apply PCA.
    :param sif_coefficient: The SIF coefficient to use. If this is None, no weighting is applied.
        Should be a value >= 0 and < 1.0. A value of 1e-4 is a good default.
    :param token_remove_pattern: If this is set to a string, we compile this into a regex. Any tokens that conform to this regex pattern will be removed from the vocabulary.
    :param trust_remote_code: Whether to trust the remote code. If this is False, we will only load components coming from `transformers`. If this is True, we will load all components.
    :param quantize_to: The data type to quantize to. Can be any of the DType enum members or their string equivalents.
    :param vocabulary_quantization: The number of clusters to use for vocabulary quantization. If this is None, no quantization is performed.
    :param pooling: The pooling mode to use for creating embeddings. Can be one of:
        'mean' (default): mean over all tokens. Robust and works well in most cases.
        'last': use the last token's hidden state (often the [EOS] token). Common for decoder-style models.
        'first': use the first token's hidden state ([CLS] token in BERT-style models).
        'pooler': use the pooler output (if available). This is often a non-linear projection of the [CLS] token.
    :return: A StaticModel

    """
    model: PreTrainedModel = AutoModel.from_pretrained(model_name, trust_remote_code=trust_remote_code)
    tokenizer = cast(
        PreTrainedTokenizerFast,
        AutoTokenizer.from_pretrained(model_name, trust_remote_code=trust_remote_code, use_fast=True),
    )

    return distill_from_model(
        model=model,
        tokenizer=tokenizer,
        vocabulary=vocabulary,
        device=device,
        pca_dims=pca_dims,
        token_remove_pattern=token_remove_pattern,
        sif_coefficient=sif_coefficient,
        quantize_to=quantize_to,
        vocabulary_quantization=vocabulary_quantization,
        pooling=pooling,
    )