configuration_mot.py

# coding=utf-8
# Copyright 2024 The OpenAI Team Authors and HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  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.
"""MixtureOfTokens configuration"""

from transformers import PretrainedConfig
from transformers.utils import logging


logger = logging.get_logger(__name__)


class MoTConfig(PretrainedConfig):
    """
    This is the configuration class to store the configuration of a [`MoTModel`]. It is used to
    instantiate a MixtureOfTokens 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 MixtureOfTokens
    [mot](https://huggingface.co/mot) architecture.

    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 50257):
            Vocabulary size of the MixtureOfTokens model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`MoTModel`].
        n_positions (`int`, *optional*, defaults to 1024):
            The maximum sequence length that this model might ever be used with. Typically set this to something large
            just in case (e.g., 512 or 1024 or 2048).
        n_embd (`int`, *optional*, defaults to 768):
            Dimensionality of the embeddings and hidden states.
        n_layer (`int`, *optional*, defaults to 12):
            Number of hidden layers in the Transformer encoder.
        n_head (`int`, *optional*, defaults to 12):
            Number of attention heads for each attention layer in the Transformer encoder.
        n_inner (`int`, *optional*):
            Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
        n_expert (`int`, *optional*, defaults to 32):
            The number of experts.
        group_size (`int`, *optional*, defaults to 32):
            The number of tokens per expert.
        expert_size (`int`, *optional*):
            The dimensionality of an expert. `None` will set it to n_inner / n_head.
        init_scale (`float`, *optional*, defaults to 1.0):
            The scaling factor for the initialization of MoTMLP weights. Inactive when creating through `from_pretrained`.
        activation_function (`str`, *optional*, defaults to `"gelu_new"`):
            Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`.
        resid_pdrop (`float`, *optional*, defaults to 0.1):
            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        embd_pdrop (`float`, *optional*, defaults to 0.1):
            The dropout ratio for the embeddings.
        attn_pdrop (`float`, *optional*, defaults to 0.1):
            The dropout ratio for the attention.
        layer_norm_epsilon (`float`, *optional*, defaults to 1e-05):
            The epsilon to use in the layer normalization layers.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        scale_attn_weights (`bool`, *optional*, defaults to `True`):
            Scale attention weights by dividing by sqrt(hidden_size)..
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models).
        bos_token_id (`int`, *optional*, defaults to 50256):
            Id of the beginning of sentence token in the vocabulary.
        eos_token_id (`int`, *optional*, defaults to 50256):
            Id of the end of sentence token in the vocabulary.
        scale_attn_by_inverse_layer_idx (`bool`, *optional*, defaults to `False`):
            Whether to additionally scale attention weights by `1 / layer_idx + 1`.
        reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`):
            Whether to scale keys (K) prior to computing attention (dot-product) and upcast attention
            dot-product/softmax to float() when training with mixed precision.
        emit_softmax_over_experts (`bool`, *optional*, defaults to `False`):
            Determines the method of redistributing aggregated tokens in the MoT MLP. By default the model uses the merge weights.
            This flag switches it to taking a softmax over the experts.
        use_discrete_routing (`bool`, *optional*, defaults to `False`):
            Discretize the mixing, sending only to the expert with the highest weight. Inference-only.

    Example:

    ```python
    >>> from transformers import MoTConfig, MoTModel

    >>> # Initializing a MoT configuration
    >>> configuration = MoTConfig()

    >>> # Initializing a model (with random weights) from the configuration
    >>> model = MoTModel(configuration)

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

    model_type = "mot"
    keys_to_ignore_at_inference = ["past_key_values"]
    attribute_map = {
        "hidden_size": "n_embd",
        "max_position_embeddings": "n_positions",
        "num_attention_heads": "n_head",
        "num_hidden_layers": "n_layer",
    }

    def __init__(
        self,
        vocab_size=50257,
        n_positions=1024,
        n_embd=768,
        n_layer=12,
        n_head=12,
        n_inner=None,
        n_expert=32,
        group_size=32,
        expert_size=None,
        init_scale=1.0,
        activation_function="gelu_new",
        resid_pdrop=0.1,
        embd_pdrop=0.1,
        attn_pdrop=0.1,
        layer_norm_epsilon=1e-5,
        initializer_range=0.02,
        scale_attn_weights=True,
        use_cache=True,
        bos_token_id=50256,
        eos_token_id=50256,
        scale_attn_by_inverse_layer_idx=False,
        reorder_and_upcast_attn=False,
        emit_softmax_over_experts=False,
        use_discrete_routing=False,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.n_positions = n_positions
        self.n_embd = n_embd
        self.n_layer = n_layer
        self.n_head = n_head
        self.n_inner = n_inner
        self.n_expert = n_expert
        self.group_size = group_size
        self.expert_size = expert_size
        self.init_scale = init_scale
        self.activation_function = activation_function
        self.resid_pdrop = resid_pdrop
        self.embd_pdrop = embd_pdrop
        self.attn_pdrop = attn_pdrop
        self.layer_norm_epsilon = layer_norm_epsilon
        self.initializer_range = initializer_range
        self.scale_attn_weights = scale_attn_weights
        self.use_cache = use_cache
        self.scale_attn_by_inverse_layer_idx = scale_attn_by_inverse_layer_idx
        self.reorder_and_upcast_attn = reorder_and_upcast_attn
        self.emit_softmax_over_experts = emit_softmax_over_experts
        self.use_discrete_routing = use_discrete_routing

        self.bos_token_id = bos_token_id
        self.eos_token_id = eos_token_id

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