configuration_apriel.py

# coding=utf-8
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Apriel model configuration"""

import math
from typing import Optional, Tuple

from transformers.configuration_utils import PretrainedConfig
from transformers.utils import is_torch_available, logging

logger = logging.get_logger(__name__)

if is_torch_available():
    import torch

def _compute_default_rope_parameters(
    config: Optional[PretrainedConfig] = None,
    device: Optional["torch.device"] = None,
    seq_len: Optional[int] = None,
    **rope_kwargs,
) -> Tuple["torch.Tensor", float]:
    """
    Computes the inverse frequencies according to the original RoPE implementation
    Args:
        config ([`~transformers.PretrainedConfig`]):
            The model configuration.
        device (`torch.device`):
            The device to use for initialization of the inverse frequencies.
        seq_len (`int`, *optional*):
            The current sequence length. Unused for this type of RoPE.
        rope_kwargs (`Dict`, *optional*):
            BC compatibility with the previous RoPE class instantiation, will be removed in v4.45.
    Returns:
        Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
        post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
    """
    if config is not None and len(rope_kwargs) > 0:
        raise ValueError(
            "Unexpected arguments: `**rope_kwargs` and `config` are mutually exclusive in "
            f"`_compute_default_rope_parameters`, got `rope_kwargs`={rope_kwargs} and `config`={config}"
        )
    if len(rope_kwargs) > 0:
        base = rope_kwargs["base"]
        dim = rope_kwargs["dim"]
    elif config is not None:
        base = config.rope_theta
        partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
        head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        dim = int(head_dim * partial_rotary_factor)

    attention_factor = 1.0  # Unused in this type of RoPE

    # Compute the inverse frequencies
    inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).float().to(device) / dim))
    return inv_freq, attention_factor

def _compute_yarn_parameters(
    config: PretrainedConfig, device: "torch.device", seq_len: Optional[int] = None, **rope_kwargs
) -> Tuple["torch.Tensor", float]:
    """
    Computes the inverse frequencies with NTK scaling. Please refer to the
    [original paper](https://arxiv.org/abs/2309.00071)
    Args:
        config ([`~transformers.PretrainedConfig`]):
            The model configuration.
        device (`torch.device`):
            The device to use for initialization of the inverse frequencies.
        seq_len (`int`, *optional*):
            The current sequence length. Unused for this type of RoPE.
        rope_kwargs (`Dict`, *optional*):
            BC compatibility with the previous RoPE class instantiation, will be removed in v4.45.
    Returns:
        Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
        post-processing scaling factor applied to the computed cos/sin.
    """
    # No need to keep BC with yarn, unreleased when this new pattern was created.
    if len(rope_kwargs) > 0:
        raise ValueError(
            f"Unexpected arguments: `**rope_kwargs` should be unset in `_compute_yarn_parameters`, got {rope_kwargs}"
        )

    base = config.rope_theta
    partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
    head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
    dim = int(head_dim * partial_rotary_factor)
        
    # Apriel: Use original max_position_embeddings instead of max_position_embeddings
    max_position_embeddings = config.rope_scaling.get("original_max_position_embeddings", config.max_position_embeddings)
    factor = config.rope_scaling["factor"]

    # Sets the attention factor as suggested in the paper
    attention_factor = config.rope_scaling.get("attention_factor")
    if attention_factor is None:
        attention_factor = 0.1 * math.log(factor) + 1.0

    # Optional config options
    # beta_fast/beta_slow: as suggested in the paper, default to 32/1 (correspondingly)
    beta_fast = config.rope_scaling.get("beta_fast") or 32
    beta_slow = config.rope_scaling.get("beta_slow") or 1

    # Compute the inverse frequencies
    def find_correction_dim(num_rotations, dim, base, max_position_embeddings):
        """Inverse dimension formula to find the dimension based on the number of rotations"""
        return (dim * math.log(max_position_embeddings / (num_rotations * 2 * math.pi))) / (2 * math.log(base))

    def find_correction_range(low_rot, high_rot, dim, base, max_position_embeddings):
        """Find dimension range bounds based on rotations"""
        low = math.floor(find_correction_dim(low_rot, dim, base, max_position_embeddings))
        high = math.ceil(find_correction_dim(high_rot, dim, base, max_position_embeddings))
        return max(low, 0), min(high, dim - 1)

    def linear_ramp_factor(min, max, dim):
        if min == max:
            max += 0.001  # Prevent singularity

        linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
        ramp_func = torch.clamp(linear_func, 0, 1)
        return ramp_func

    # Note on variable naming: "interpolation" comes from the original technique, where we interpolate the position IDs
    # to expand the possible context length. In other words, interpolation = apply scaling factor.
    pos_freqs = base ** (torch.arange(0, dim, 2).float().to(device) / dim)
    inv_freq_extrapolation = 1.0 / pos_freqs
    inv_freq_interpolation = 1.0 / (factor * pos_freqs)

    low, high = find_correction_range(beta_fast, beta_slow, dim, base, max_position_embeddings)

    # Get n-dimensional rotational scaling corrected for extrapolation
    inv_freq_extrapolation_factor = 1 - linear_ramp_factor(low, high, dim // 2).float().to(device)
    inv_freq = (
        inv_freq_interpolation * (1 - inv_freq_extrapolation_factor)
        + inv_freq_extrapolation * inv_freq_extrapolation_factor
    )

    return inv_freq, attention_factor

def _check_received_keys(
    rope_type: str,
    received_keys: set,
    required_keys: set,
    optional_keys: Optional[set] = None,
    ignore_keys: Optional[set] = None,
):
    
    """Compare the received keys in `config.rope_scaling` against the expected and optional keys"""
    # BC: "rope_type" was originally "type" -- let's check for "rope_type" when "type" is present
    if "type" in received_keys:
        received_keys -= {"type"}
        required_keys.add("rope_type")

    # Some models need to store model-specific keys, and we don't want to throw warning at them
    if ignore_keys is not None:
        received_keys -= ignore_keys

    missing_keys = required_keys - received_keys
    if missing_keys:
        raise KeyError(f"Missing required keys in `rope_scaling` for 'rope_type'='{rope_type}': {missing_keys}")

    if optional_keys is not None:
        unused_keys = received_keys - required_keys - optional_keys
    else:
        unused_keys = received_keys - required_keys
    if unused_keys:
        logger.warning(f"Unrecognized keys in `rope_scaling` for 'rope_type'='{rope_type}': {unused_keys}")


def _validate_default_rope_parameters(config: PretrainedConfig, ignore_keys: Optional[set] = None):
    rope_scaling = config.rope_scaling
    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))  # BC: "rope_type" was originally "type"
    required_keys = {"rope_type"}
    received_keys = set(rope_scaling.keys())
    _check_received_keys(rope_type, received_keys, required_keys, ignore_keys=ignore_keys)

def _validate_yarn_parameters(config: PretrainedConfig, ignore_keys: Optional[set] = None):
    rope_scaling = config.rope_scaling
    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))  # BC: "rope_type" was originally "type"
    required_keys = {"rope_type", "factor", "original_max_position_embeddings"}
    optional_keys = {"attention_factor", "beta_fast", "beta_slow"}
    received_keys = set(rope_scaling.keys())
    _check_received_keys(rope_type, received_keys, required_keys, optional_keys, ignore_keys=ignore_keys)

    factor = rope_scaling["factor"]
    if factor is None or not isinstance(factor, float) or factor < 1.0:
        logger.warning(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")

    attention_factor = rope_scaling.get("attention_factor")
    if attention_factor is not None and (not isinstance(attention_factor, float) or attention_factor < 0):
        logger.warning(
            f"`rope_scaling`'s attention_factor field must be a float greater than 0, got {attention_factor}"
        )
    beta_fast = rope_scaling.get("beta_fast")
    if beta_fast is not None and not isinstance(beta_fast, float):
        logger.warning(f"`rope_scaling`'s beta_fast field must be a float, got {beta_fast}")
    beta_slow = rope_scaling.get("beta_slow")
    if beta_slow is not None and not isinstance(beta_slow, float):
        logger.warning(f"`rope_scaling`'s beta_slow field must be a float, got {beta_slow}")

    if (beta_fast or 32) < (beta_slow or 1):
        logger.warning(
            f"`rope_scaling`'s beta_fast field must be greater than beta_slow, got beta_fast={beta_fast} "
            f"(defaults to 32 if None) and beta_slow={beta_slow} (defaults to 1 if None)"
        )
# This maps the "rope_type" string field in rope config to the corresponding function to compute the RoPE parameters
# from the model config. You can append new {'rope_type': callable} pairs to this dictionary to enable custom RoPE
# parameterizations, as long as the callable has the same signature.
ROPE_INIT_FUNCTIONS = {
    "default": _compute_default_rope_parameters,
    "yarn": _compute_yarn_parameters,
}

# Like `ROPE_INIT_FUNCTIONS`, this validation function mapping can be dynamically updated for custom RoPE types.
ROPE_VALIDATION_FUNCTIONS = {
    "default": _validate_default_rope_parameters,
    "yarn": _validate_yarn_parameters,
}

def rope_config_validation(config: PretrainedConfig, ignore_keys: Optional[set] = None):
    """
    Validate the RoPE config arguments, given a `PretrainedConfig` object
    """
    rope_scaling = getattr(config, "rope_scaling", None)  # not a default parameter in `PretrainedConfig`
    if rope_scaling is None:
        return

    # BC: "rope_type" was originally "type"
    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", "default"))
    validation_fn = ROPE_VALIDATION_FUNCTIONS.get(rope_type)
    if validation_fn is not None:
        validation_fn(config, ignore_keys=ignore_keys)
    else:
        logger.warning(
            f"Missing validation function mapping in `ROPE_VALIDATION_FUNCTIONS` for 'rope_type'='{rope_type}'"
        )

class AprielConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`AprielModel`]. It is used to instantiate an Apriel
    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
    defaults will yield a similar configuration to that of the Apriel-5B-Base.

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.


    Args:
        vocab_size (`int`, *optional*, defaults to 32000):
            Vocabulary size of the Apriel model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`AprielModel`]
        hidden_size (`int`, *optional*, defaults to 4096):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 11008):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 32):
            Number of hidden layers in the Transformer decoder.
        num_attention_heads (`int`, *optional*, defaults to 32):
            Number of attention heads for each attention layer in the Transformer decoder.
        num_key_value_heads (`int`, *optional*):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
            by meanpooling all the original heads within that group. For more details checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
            `num_attention_heads`.
        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to 2048):
            The maximum sequence length that this model might ever be used with. Apriel-5B-Base supports up to 16384 tokens.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the rms normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        pad_token_id (`int`, *optional*):
            Padding token id.
        bos_token_id (`int`, *optional*, defaults to 1):
            Beginning of stream token id.
        eos_token_id (`int`, *optional*, defaults to 2):
            End of stream token id.
        pretraining_tp (`int`, *optional*, defaults to 1):
            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to
            understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
            results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether to tie weight embeddings
        rope_theta (`float`, *optional*, defaults to 10000.0):
            The base period of the RoPE embeddings.
        rope_scaling (`Dict`, *optional*):
            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
            accordingly.
            Expected contents:
                `rope_type` (`str`):
                    The sub-variant of RoPE to use. Can be one of ['default', 'yarn'], with 'default' being the original RoPE implementation.
                `factor` (`float`, *optional*):
                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
                    original maximum pre-trained length.
                `original_max_position_embeddings` (`int`, *optional*):
                    Used with 'yarn', 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
                    pretraining.
                `attention_factor` (`float`, *optional*):
                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
                    computation. If unspecified, it defaults to value recommended by the implementation, using the
                    `factor` field to infer the suggested value.
                `beta_fast` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
                    ramp function. If unspecified, it defaults to 32.
                `beta_slow` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
                    ramp function. If unspecified, it defaults to 1.
                `short_factor` (`List[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `long_factor` (`List[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `low_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
                `high_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
        attention_bias (`bool`, *optional*, defaults to `False`):
            Whether to use a bias in the query, key, value and output projection layers during self-attention.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        mlp_bias (`bool`, *optional*, defaults to `False`):
            Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
        head_dim (`int`, *optional*):
            The attention head dimension. If None, it will default to hidden_size // num_attention_heads

    ```python
    >>> from transformers import AprielModel, AprielConfig

    >>> # Initializing an Apriel Apriel-5B-Base style configuration
    >>> configuration = AprielConfig()

    >>> # Initializing a model from the Apriel-5B-Base style configuration
    >>> model = AprielModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "apriel"
    keys_to_ignore_at_inference = ["past_key_values"]
    # Default tensor parallel plan for base model `AprielModel`
    base_model_tp_plan = {
        "layers.*.self_attn.q_proj": "colwise",
        "layers.*.self_attn.k_proj": "colwise",
        "layers.*.self_attn.v_proj": "colwise",
        "layers.*.self_attn.o_proj": "rowwise",
        "layers.*.mlp.gate_proj": "colwise",
        "layers.*.mlp.up_proj": "colwise",
        "layers.*.mlp.down_proj": "rowwise",
    }
    base_model_pp_plan = {
        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
        "norm": (["hidden_states"], ["hidden_states"]),
    }

    def __init__(
        self,
        vocab_size=32000,
        hidden_size=4096,
        intermediate_size=11008,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=None,
        hidden_act="silu",
        max_position_embeddings=2048,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=1,
        eos_token_id=2,
        pretraining_tp=1,
        tie_word_embeddings=False,
        rope_theta=10000.0,
        rope_scaling=None,
        attention_bias=False,
        attention_dropout=0.0,
        mlp_bias=False,
        head_dim=None,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads

        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.pretraining_tp = pretraining_tp
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.rope_scaling = rope_scaling
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout
        self.mlp_bias = mlp_bias
        self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads
        # Validate the correctness of rotary position embeddings parameters
        # BC: if there is a 'type' field, copy it it to 'rope_type'.
        if self.rope_scaling is not None and "type" in self.rope_scaling:
            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
        rope_config_validation(self)

        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )


__all__ = ["AprielConfig"]