configuration_mixtral.py

# coding=utf-8
# Copyright 2023 Mixtral AI and the HuggingFace Inc. team. All rights reserved.
#
# 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.
""" Mixtral model configuration"""

import copy
from typing import Any, Dict

from transformers import __version__
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging

logger = logging.get_logger(__name__)

MIXTRAL_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    "mistral-ai/Mixtral-8x7B": "https://huggingface.co/mistral-ai/Mixtral-8x7B/resolve/main/config.json",
}


def recursive_diff_dict(dict_a, dict_b, config_obj=None):
    """
    Helper function to recursively take the diff between two nested dictionaries. The resulting diff only contains the
    values from `dict_a` that are different from values in `dict_b`.
    """
    diff = {}
    default = config_obj.__class__().to_dict() if config_obj is not None else {}
    for key, value in dict_a.items():
        obj_value = getattr(config_obj, str(key), None)
        if (
            isinstance(obj_value, PretrainedConfig)
            and key in dict_b
            and isinstance(dict_b[key], dict)
        ):
            diff_value = recursive_diff_dict(value, dict_b[key], config_obj=obj_value)
            if len(diff_value) > 0:
                diff[key] = diff_value
        elif (
            key not in dict_b
            or value != dict_b[key]
            or key not in default
            or value != default[key]
        ):
            diff[key] = value
    return diff


class MixtralConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`MixtralModel`]. It is used to instantiate an
    Mixtral 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 Mixtral-7B-v0.1 or Mixtral-7B-Instruct-v0.1.

    [mixtralai/Mixtral-8x7B](https://huggingface.co/mixtralai/Mixtral-8x7B)
    [mixtralai/Mixtral-7B-Instruct-v0.1](https://huggingface.co/mixtralai/Mixtral-7B-Instruct-v0.1)

    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 Mixtral model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`MixtralModel`]
        hidden_size (`int`, *optional*, defaults to 4096):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 14336):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 32):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 32):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_key_value_heads (`int`, *optional*, defaults to 8):
            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 `8`.
        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 `4096*32`):
            The maximum sequence length that this model might ever be used with. Mixtral's sliding window attention
            allows sequence of up to 4096*32 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-05):
            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*):
            The id of the padding token.
        bos_token_id (`int`, *optional*, defaults to 1):
            The id of the "beginning-of-sequence" token.
        eos_token_id (`int`, *optional*, defaults to 2):
            The id of the "end-of-sequence" token.
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether the model's input and output word embeddings should be tied.
        rope_theta (`float`, *optional*, defaults to 1000000.0):
            The base period of the RoPE embeddings.
        sliding_window (`int`, *optional*, defaults to 4096):
            Sliding window attention window size. If not specified, will default to `4096`.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        num_experts_per_tok (`int`, *optional*, defaults to 2):
            The number of experts to root per-token, can be also interpreted as the `top-p` routing
            parameter
        num_local_experts (`int`, *optional*, defaults to 8):
            Number of experts per Sparse MLP layer.
        output_router_logits (`bool`, *optional*, defaults to `False`):
            Whether or not the router logits should be returned by the model. Enabeling this will also
            allow the model to output the auxiliary loss. See [here]() for more details
        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
            The aux loss factor for the total loss.

    ```python
    >>> from transformers import MixtralModel, MixtralConfig

    >>> # Initializing a Mixtral 7B style configuration
    >>> configuration = MixtralConfig()

    >>> # Initializing a model from the Mixtral 7B style configuration
    >>> model = MixtralModel(configuration)

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

    model_type = "mixtral"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=32000,
        hidden_size=4096,
        intermediate_size=14336,
        intermediate_size_residual=None,  # 🔍
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=8,
        hidden_act="silu",
        max_position_embeddings=4096 * 32,
        initializer_range=0.02,
        rms_norm_eps=1e-5,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=1,
        eos_token_id=2,
        tie_word_embeddings=False,
        rope_theta=1e6,
        sliding_window=4096,
        attention_dropout=0.0,
        num_experts_per_tok=2,
        num_local_experts=8,
        scale_factor: float = 1.0,  # 🔍
        output_router_logits=False,
        router_aux_loss_coef=0.001,
        moe_type: str = "modulelist",  # 🔍
        num_moe_contract_layers: int = 0,  # 🔍 the number of layers that are not converted into MoE at each side of the model
        use_attn_moe: bool = False,  # 🔍
        top_k_attn: int = None,  # 🔍
        attn_experts: int = None, 
        scale_factor_attn: float = None,  # 🔍
        use_layer_wise_balance: bool = False,  # ✨ whether to fix the balance loss bug for Mixtral
        add_rescale_bias: bool = False,  # 🔍 whether to add bias to the AttentionMoE `o_proj` & MoE `down_proj` for distribution alignment
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.intermediate_size_residual = intermediate_size_residual  # 🔍
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.sliding_window = sliding_window

        # 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.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout

        self.num_experts_per_tok = num_experts_per_tok
        self.num_local_experts = num_local_experts
        self.scale_factor = scale_factor  # 🔍
        self.output_router_logits = output_router_logits
        self.router_aux_loss_coef = router_aux_loss_coef
        self.moe_type = moe_type  # 🔍
        self.num_moe_contract_layers = num_moe_contract_layers  # 🔍

        # 🔍 for Attention MoE
        self.use_attn_moe = use_attn_moe
        self.top_k_attn = top_k_attn
        self.scale_factor_attn = scale_factor_attn
        self.attn_experts = attn_experts

        # ✨ For balance loss bugfix
        self.use_layer_wise_balance = use_layer_wise_balance

        # 🔍 for distribution alignment
        self.add_rescale_bias = add_rescale_bias

        # Attention implementation to use, if relevant.
        self._attn_implementation_internal = kwargs.pop("attn_implementation", None)

        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,
        )

    @property
    def _attn_implementation(self):
        # This property is made private for now (as it cannot be changed and a PreTrainedModel.use_attn_implementation method needs to be implemented.)
        if hasattr(self, "_attn_implementation_internal"):
            if self._attn_implementation_internal is None:
                # `config.attn_implementation` should never be None, for backward compatibility.
                return "flash_attention_2"
                # return "eager"
            else:
                return self._attn_implementation_internal
        else:
            return "flash_attention_2"
            # return "eager"
            


    @_attn_implementation.setter
    def _attn_implementation(self, value):
        self._attn_implementation_internal = value

    def to_dict(self) -> Dict[str, Any]:
        """
        Serializes this instance to a Python dictionary.

        Returns:
            `Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance.
        """
        output = copy.deepcopy(self.__dict__)
        if hasattr(self.__class__, "model_type"):
            output["model_type"] = self.__class__.model_type
        if "_auto_class" in output:
            del output["_auto_class"]
        if "_commit_hash" in output:
            del output["_commit_hash"]
        if "_attn_implementation_internal" in output:
            del output["_attn_implementation_internal"]

        # Transformers version when serializing the model
        output["transformers_version"] = __version__

        for key, value in output.items():
            # Deal with nested configs like CLIP
            if isinstance(value, PretrainedConfig):
                value = value.to_dict()
                del value["transformers_version"]

            output[key] = value

        if hasattr(self, "quantization_config"):
            output["quantization_config"] = (
                self.quantization_config.to_dict()
                if not isinstance(self.quantization_config, dict)
                else self.quantization_config
            )

            # pop the `_pre_quantization_dtype` as torch.dtypes are not serializable.
            _ = output.pop("_pre_quantization_dtype", None)

        self.dict_torch_dtype_to_str(output)

        return output

    def to_diff_dict(self) -> Dict[str, Any]:
        """
        Removes all attributes from config which correspond to the default config attributes for better readability and
        serializes to a Python dictionary.

        Returns:
            `Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance,
        """
        config_dict = self.to_dict()

        # get the default config dict
        default_config_dict = PretrainedConfig().to_dict()

        # get class specific config dict
        class_config_dict = (
            self.__class__().to_dict() if not self.is_composition else {}
        )

        serializable_config_dict = {}

        # only serialize values that differ from the default config
        for key, value in config_dict.items():
            if (
                isinstance(getattr(self, key, None), PretrainedConfig)
                and key in class_config_dict
                and isinstance(class_config_dict[key], dict)
            ):
                # For nested configs we need to clean the diff recursively
                diff = recursive_diff_dict(
                    value, class_config_dict[key], config_obj=getattr(self, key, None)
                )
                if "model_type" in value:
                    # Needs to be set even if it's not in the diff
                    diff["model_type"] = value["model_type"]
                if len(diff) > 0:
                    serializable_config_dict[key] = diff
            elif (
                key not in default_config_dict
                or key == "transformers_version"
                or value != default_config_dict[key]
                or (key in class_config_dict and value != class_config_dict[key])
            ):
                serializable_config_dict[key] = value

        if hasattr(self, "quantization_config"):
            serializable_config_dict["quantization_config"] = (
                self.quantization_config.to_dict()
                if not isinstance(self.quantization_config, dict)
                else self.quantization_config
            )

            # pop the `_pre_quantization_dtype` as torch.dtypes are not serializable.
            _ = serializable_config_dict.pop("_pre_quantization_dtype", None)

        self.dict_torch_dtype_to_str(serializable_config_dict)

        if "_attn_implementation_internal" in serializable_config_dict:
            del serializable_config_dict["_attn_implementation_internal"]

        return serializable_config_dict