configuration_celerity.py

# coding=utf-8
# Copyright 2023 The OpenAI Team Authors and HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
# Copyright 2023 Cerebras Systems.
#
# 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.
""" Celerity configuration"""

from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging
from typing import Literal

logger = logging.get_logger(__name__)

class CelerityConfig(PretrainedConfig):
    """
    This is the configuration class to store the configuration of a [`CelerityModel`]. It is used to instantiate a Celerity
    model according to the specified arguments, defining the model 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 128256):
            Vocabulary size of the Celerity model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`CelerityModel`].
        n_positions (`int`, *optional*, defaults to 8192):
            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 640):
            Dimensionality of the embeddings and hidden states.
        n_layer (`int`, *optional*, defaults to 13):
            Number of hidden layers in the Transformer encoder.
        n_head (`int`, *optional*, defaults to 10):
            Number of attention heads for each attention layer in the Transformer encoder.
        n_inner (`int`, *optional*, defaults to None):
            Dimensionality of the inner feed-forward layers. `None` will set it to 8 times n_embd
        activation_function (`str`, *optional*, defaults to `"squared_relu"`):
            Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new", "swiglu", "squared_relu"]`.
        resid_pdrop (`float`, *optional*, defaults to 0.0):
            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        embd_pdrop (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the embeddings.
        attn_pdrop (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention.
        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
            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 `128000`):
            BOS token ID.
        eos_token_id (`int`, *optional*, defaults to `128001`):
            EOS token ID.
        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.
        position_embedding_type (`str`, *optional*, defaults to `"alibi"`):
            Positional embedding can be either `"alibi"` or `"learned"`.
        embeddings_scale (`float`, *optional*, defaults to 1.0):
            Scaler multiplier to scale token and position embeddings.
        output_logits_scale (`float`, *optional*, defaults to 1.0):
            Scaler multiplier to scale output logits.
        mhsa_residual_scale (`float`, *optional*, defaults to 1.0):
            Scaler multiplier to scale the multi-head self-attention residual stream.
        mlp_residual_scale (`float`, *optional*, defaults to 1.0):
            Scaler multiplier to scale the MLP residual stream.
        scale_qk_dot_by_d (`bool`, *optional*, defaults to `False`):
            Scale attention weights by dividing by hidden_size instead of sqrt(hidden_size). Need to set
            scale_attn_weights to `True` as well.
    Example:
    ```python
    >>> from transformers import CelerityConfig, CelerityModel
    >>> # Initializing a Celerity configuration
    >>> configuration = CelerityConfig()
    >>> # Initializing a model (with random weights) from the configuration
    >>> model = CelerityModel(configuration)
    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "celerity"
    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",
    }
    attention_mask_type: Literal["causal", "full", "block_causal"] = "causal"

    def __init__(
        self,
        vocab_size=128256,
        n_positions=8192,
        n_embd=640,
        n_layer=13,
        n_head=10,
        n_inner=5120,
        activation_function="squared_relu",
        resid_pdrop=0.0,
        embd_pdrop=0.0,
        attn_pdrop=0.0,
        layer_norm_epsilon=1e-5,
        initializer_range=0.02,
        scale_attn_weights=True,
        use_cache=True,
        bos_token_id=128000,
        eos_token_id=128001,
        scale_attn_by_inverse_layer_idx=False,
        reorder_and_upcast_attn=False,
        position_embedding_type="alibi",
        embeddings_scale=1.0,
        output_logits_scale=1.0,
        mhsa_residual_scale=1.0,
        mlp_residual_scale=1.0,
        scale_qk_dot_by_d=True,
        ### BEGIN DIFFUSION PARAMETERS
        num_extra_input_vocab_tokens=0,
        attention_mask_type: Literal["causal", "full", "block_causal"] = "full",
        ### END DIFFUSION PARAMETERS
        **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.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.bos_token_id = bos_token_id
        self.eos_token_id = eos_token_id
        self.num_extra_input_vocab_tokens = num_extra_input_vocab_tokens

        self.position_embedding_type = position_embedding_type
        self.embeddings_scale = embeddings_scale
        self.output_logits_scale = output_logits_scale
        self.mhsa_residual_scale = mhsa_residual_scale
        self.mlp_residual_scale = mlp_residual_scale
        self.scale_qk_dot_by_d = scale_qk_dot_by_d

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