Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2024 weak-kajuma and the HuggingFace Inc. team. All rights reserved.3#4# This code is based on Llama implementations in this library and Microsoft's5# Differential Transformer implementations.6 7# Licensed under the Apache License, Version 2.0 (the "License");8# you may not use this file except in compliance with the License.9# You may obtain a copy of the License at10#11#     http://www.apache.org/licenses/LICENSE-2.012#13# Unless required by applicable law or agreed to in writing, software14# distributed under the License is distributed on an "AS IS" BASIS,15# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.16# See the License for the specific language governing permissions and17# limitations under the License.18"""DiffLlama model configuration"""19 20from ...configuration_utils import PretrainedConfig21from ...modeling_rope_utils import rope_config_validation22 23 24class DiffLlamaConfig(PretrainedConfig):25    r"""26    This is the configuration class to store the configuration of a [`DiffLlamaModel`]. It is used to instantiate an DiffLlama27    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults28    will yield a similar configuration to that of the [kajuma/DiffLlama-0.3B-handcut](https://huggingface.co/kajuma/DiffLlama-0.3B-handcut).29 30    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the31    documentation from [`PretrainedConfig`] for more information.32 33 34    Args:35        vocab_size (`int`, *optional*, defaults to 32000):36            Vocabulary size of the DiffLlama model. Defines the number of different tokens that can be represented by the37            `inputs_ids` passed when calling [`DiffLlamaModel`]38        hidden_size (`int`, *optional*, defaults to 2048):39            Dimension of the hidden representations.40        intermediate_size (`int`, *optional*, defaults to 8192):41            Dimension of the MLP representations.42        num_hidden_layers (`int`, *optional*, defaults to 16):43            Number of hidden layers in the Transformer decoder.44        num_attention_heads (`int`, *optional*, defaults to 32):45            Number of attention heads for each attention layer in the Transformer decoder.46        num_key_value_heads (`int`, *optional*):47            This is the number of key_value heads that should be used to implement Grouped Query Attention. If48            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if49            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When50            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed51            by meanpooling all the original heads within that group. For more details, check out [this52            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to53            `num_attention_heads`.54        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):55            The non-linear activation function (function or string) in the decoder.56        max_position_embeddings (`int`, *optional*, defaults to 2048):57            The maximum sequence length that this model might ever be used with.58        initializer_range (`float`, *optional*, defaults to 0.02):59            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.60        rms_norm_eps (`float`, *optional*, defaults to 1e-05):61            The epsilon used by the rms normalization layers.62        use_cache (`bool`, *optional*, defaults to `True`):63            Whether or not the model should return the last key/values attentions (not used by all models). Only64            relevant if `config.is_decoder=True`.65        pad_token_id (`int`, *optional*):66            Padding token id.67        bos_token_id (`int`, *optional*, defaults to 1):68            Beginning of stream token id.69        eos_token_id (`int`, *optional*, defaults to 2):70            End of stream token id.71        tie_word_embeddings (`bool`, *optional*, defaults to `False`):72            Whether to tie weight embeddings73        rope_theta (`float`, *optional*, defaults to 10000.0):74            The base period of the RoPE embeddings.75        rope_scaling (`Dict`, *optional*):76            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type77            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value78            accordingly.79            Expected contents:80                `rope_type` (`str`):81                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',82                    'diffllama3'], with 'default' being the original RoPE implementation.83                `factor` (`float`, *optional*):84                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In85                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *86                    original maximum pre-trained length.87                `original_max_position_embeddings` (`int`, *optional*):88                    Used with 'dynamic', 'longrope' and 'diffllama3'. The original max position embeddings used during89                    pretraining.90                `attention_factor` (`float`, *optional*):91                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention92                    computation. If unspecified, it defaults to value recommended by the implementation, using the93                    `factor` field to infer the suggested value.94                `beta_fast` (`float`, *optional*):95                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear96                    ramp function. If unspecified, it defaults to 32.97                `beta_slow` (`float`, *optional*):98                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear99                    ramp function. If unspecified, it defaults to 1.100                `short_factor` (`list[float]`, *optional*):101                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<102                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden103                    size divided by the number of attention heads divided by 2104                `long_factor` (`list[float]`, *optional*):105                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<106                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden107                    size divided by the number of attention heads divided by 2108                `low_freq_factor` (`float`, *optional*):109                    Only used with 'diffllama3'. Scaling factor applied to low frequency components of the RoPE110                `high_freq_factor` (`float`, *optional*):111                    Only used with 'diffllama3'. Scaling factor applied to high frequency components of the RoPE112        attention_bias (`bool`, *optional*, defaults to `False`):113            Whether to use a bias in the query, key, value and output projection layers during self-attention.114        attention_dropout (`float`, *optional*, defaults to 0.0):115            The dropout ratio for the attention probabilities.116        lambda_std_dev (`float`, *optional*, defaults to 0.1):117            The standard deviation for initialization of parameter lambda in attention layer.118        head_dim (`int`, *optional*):119            The attention head dimension. If None, it will default to hidden_size // num_heads120 121    ```python122    >>> from transformers import DiffLlamaModel, DiffLlamaConfig123 124    >>> # Initializing a DiffLlama diffllama-7b style configuration125    >>> configuration = DiffLlamaConfig()126 127    >>> # Initializing a model from the diffllama-7b style configuration128    >>> model = DiffLlamaModel(configuration)129 130    >>> # Accessing the model configuration131    >>> configuration = model.config132    ```"""133 134    model_type = "diffllama"135    keys_to_ignore_at_inference = ["past_key_values"]136 137    def __init__(138        self,139        vocab_size=32000,140        hidden_size=2048,141        intermediate_size=8192,142        num_hidden_layers=16,143        num_attention_heads=32,144        num_key_value_heads=None,145        hidden_act="silu",146        max_position_embeddings=2048,147        initializer_range=0.02,148        rms_norm_eps=1e-5,149        use_cache=True,150        pad_token_id=None,151        bos_token_id=1,152        eos_token_id=2,153        tie_word_embeddings=False,154        rope_theta=10000.0,155        rope_scaling=None,156        attention_bias=False,157        attention_dropout=0.0,158        lambda_std_dev=0.1,159        head_dim=None,160        **kwargs,161    ):162        self.vocab_size = vocab_size163        self.max_position_embeddings = max_position_embeddings164        self.hidden_size = hidden_size165        self.intermediate_size = intermediate_size166        self.num_hidden_layers = num_hidden_layers167        self.num_attention_heads = num_attention_heads168 169        # for backward compatibility170        if num_key_value_heads is None:171            num_key_value_heads = num_attention_heads172 173        self.num_key_value_heads = num_key_value_heads174        self.hidden_act = hidden_act175        self.initializer_range = initializer_range176        self.rms_norm_eps = rms_norm_eps177        self.use_cache = use_cache178        self.rope_theta = rope_theta179        self.rope_scaling = rope_scaling180        self.attention_bias = attention_bias181        self.attention_dropout = attention_dropout182        self.lambda_std_dev = lambda_std_dev183        self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads184        # Validate the correctness of rotary position embeddings parameters185        # BC: if there is a 'type' field, copy it it to 'rope_type'.186        if self.rope_scaling is not None and "type" in self.rope_scaling:187            self.rope_scaling["rope_type"] = self.rope_scaling["type"]188        rope_config_validation(self)189 190        super().__init__(191            pad_token_id=pad_token_id,192            bos_token_id=bos_token_id,193            eos_token_id=eos_token_id,194            tie_word_embeddings=tie_word_embeddings,195            **kwargs,196        )197 198 199__all__ = ["DiffLlamaConfig"]200