naver-hyperclovax/HyperCLOVAX-SEED-Omni-8B

1# coding=utf-82# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.3#4# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX5# and OPT implementations in this library. It has been modified from its6# original forms to accommodate minor architectural differences compared7# to GPT-NeoX and OPT used by the Meta AI team that trained the model.8#9# Licensed under the Apache License, Version 2.0 (the "License");10# you may not use this file except in compliance with the License.11# You may obtain a copy of the License at12#13#     http://www.apache.org/licenses/LICENSE-2.014#15# Unless required by applicable law or agreed to in writing, software16# distributed under the License is distributed on an "AS IS" BASIS,17# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.18# See the License for the specific language governing permissions and19# limitations under the License.20"""LLaMA model configuration"""21 22from transformers.configuration_utils import PretrainedConfig23 24# from transformers.modeling_rope_utils import rope_config_validation25# from transformers import PretrainedConfig, rope_config_validation26 27 28class HyperCLOVAXConfig(PretrainedConfig):29    r"""30    This is the configuration class to store the configuration of a [`LlamaModel`]. It is used to instantiate an LLaMA31    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the32    defaults will yield a similar configuration to that of the LLaMA-7B.33 34    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the35    documentation from [`PretrainedConfig`] for more information.36 37 38    Args:39        vocab_size (`int`, *optional*, defaults to 32000):40            Vocabulary size of the LLaMA model. Defines the number of different tokens that can be represented by the41            `inputs_ids` passed when calling [`LlamaModel`]42        hidden_size (`int`, *optional*, defaults to 4096):43            Dimension of the hidden representations.44        intermediate_size (`int`, *optional*, defaults to 11008):45            Dimension of the MLP representations.46        num_hidden_layers (`int`, *optional*, defaults to 32):47            Number of hidden layers in the Transformer decoder.48        num_attention_heads (`int`, *optional*, defaults to 32):49            Number of attention heads for each attention layer in the Transformer decoder.50        num_key_value_heads (`int`, *optional*):51            This is the number of key_value heads that should be used to implement Grouped Query Attention. If52            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if53            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When54            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed55            by meanpooling all the original heads within that group. For more details checkout [this56            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to57            `num_attention_heads`.58        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):59            The non-linear activation function (function or string) in the decoder.60        max_position_embeddings (`int`, *optional*, defaults to 2048):61            The maximum sequence length that this model might ever be used with. Llama 1 supports up to 2048 tokens,62            Llama 2 up to 4096, CodeLlama up to 16384.63        initializer_range (`float`, *optional*, defaults to 0.02):64            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.65        rms_norm_eps (`float`, *optional*, defaults to 1e-06):66            The epsilon used by the rms normalization layers.67        use_cache (`bool`, *optional*, defaults to `True`):68            Whether or not the model should return the last key/values attentions (not used by all models). Only69            relevant if `config.is_decoder=True`.70        pad_token_id (`int`, *optional*):71            Padding token id.72        bos_token_id (`int`, *optional*, defaults to 1):73            Beginning of stream token id.74        eos_token_id (`int`, *optional*, defaults to 2):75            End of stream token id.76        pretraining_tp (`int`, *optional*, defaults to 1):77            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this78            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to79            understand more about it. This value is necessary to ensure exact reproducibility of the pretraining80            results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).81        tie_word_embeddings (`bool`, *optional*, defaults to `False`):82            Whether to tie weight embeddings83        rope_theta (`float`, *optional*, defaults to 10000.0):84            The base period of the RoPE embeddings.85        rope_scaling (`Dict`, *optional*):86            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type87            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value88            accordingly.89            Expected contents:90                `rope_type` (`str`):91                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',92                    'llama3'], with 'default' being the original RoPE implementation.93                `factor` (`float`, *optional*):94                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In95                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *96                    original maximum pre-trained length.97                `original_max_position_embeddings` (`int`, *optional*):98                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during99                    pretraining.100                `attention_factor` (`float`, *optional*):101                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention102                    computation. If unspecified, it defaults to value recommended by the implementation, using the103                    `factor` field to infer the suggested value.104                `beta_fast` (`float`, *optional*):105                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear106                    ramp function. If unspecified, it defaults to 32.107                `beta_slow` (`float`, *optional*):108                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear109                    ramp function. If unspecified, it defaults to 1.110                `short_factor` (`List[float]`, *optional*):111                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<112                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden113                    size divided by the number of attention heads divided by 2114                `long_factor` (`List[float]`, *optional*):115                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<116                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden117                    size divided by the number of attention heads divided by 2118                `low_freq_factor` (`float`, *optional*):119                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE120                `high_freq_factor` (`float`, *optional*):121                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE122        attention_bias (`bool`, *optional*, defaults to `False`):123            Whether to use a bias in the query, key, value and output projection layers during self-attention.124        attention_dropout (`float`, *optional*, defaults to 0.0):125            The dropout ratio for the attention probabilities.126        mlp_bias (`bool`, *optional*, defaults to `False`):127            Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.128        head_dim (`int`, *optional*):129            The attention head dimension. If None, it will default to hidden_size // num_heads130 131    ```python132    >>> from transformers import LlamaModel, LlamaConfig133 134    >>> # Initializing a LLaMA llama-7b style configuration135    >>> configuration = LlamaConfig()136 137    >>> # Initializing a model from the llama-7b style configuration138    >>> model = LlamaModel(configuration)139 140    >>> # Accessing the model configuration141    >>> configuration = model.config142    ```"""143 144    model_type = "hyperclovax"145    keys_to_ignore_at_inference = ["past_key_values"]146 147    def __init__(148        self,149        vocab_size=32000,150        hidden_size=4096,151        intermediate_size=11008,152        num_hidden_layers=32,153        num_attention_heads=32,154        num_key_value_heads=None,155        hidden_act="silu",156        max_position_embeddings=2048,157        initializer_range=0.02,158        rms_norm_eps=1e-6,159        use_cache=True,160        pad_token_id=None,161        bos_token_id=1,162        eos_token_id=2,163        pretraining_tp=1,164        tie_word_embeddings=False,165        rope_theta=10000.0,166        rope_scaling=None,167        attention_bias=False,168        attention_dropout=0.0,169        mlp_bias=False,170        head_dim=None,171        embedding_multiplier=1.0,  # mup172        logits_scaling=1.0,  # mup173        attention_multiplier=1.0,  # mup174        residual_multiplier=1.0,  # mup175        use_post_norm=False,  # post-norm176        auto_map={177            "AutoConfig": "configuration_hyperclovax.HyperCLOVAXConfig",178            "AutoModel": "modeling_hyperclovax.HyperCLOVAXModel",179            "AutoModelForCausalLM": "modeling_hyperclovax.HyperCLOVAXForCausalLM",180        },181        **kwargs,182    ):183        self.vocab_size = vocab_size184        self.max_position_embeddings = max_position_embeddings185        self.hidden_size = hidden_size186        self.intermediate_size = intermediate_size187        self.num_hidden_layers = num_hidden_layers188        self.num_attention_heads = num_attention_heads189 190        # for backward compatibility191        if num_key_value_heads is None:192            num_key_value_heads = num_attention_heads193 194        self.num_key_value_heads = num_key_value_heads195        self.hidden_act = hidden_act196        self.initializer_range = initializer_range197        self.rms_norm_eps = rms_norm_eps198        self.pretraining_tp = pretraining_tp199        self.use_cache = use_cache200        self.rope_theta = rope_theta201        self.rope_scaling = rope_scaling202        self.attention_bias = attention_bias203        self.attention_dropout = attention_dropout204        self.mlp_bias = mlp_bias205        self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads206        # Validate the correctness of rotary position embeddings parameters207        # BC: if there is a 'type' field, copy it it to 'rope_type'.208        if self.rope_scaling is not None and "type" in self.rope_scaling:209            self.rope_scaling["rope_type"] = self.rope_scaling["type"]210        # rope_config_validation(self)211 212        # mup213        self.embedding_multiplier = embedding_multiplier214        self.logits_scaling = logits_scaling215        self.attention_multiplier = attention_multiplier216        self.residual_multiplier = residual_multiplier217 218        # post-norm (dual-norm)219        self.use_post_norm = use_post_norm220 221        super().__init__(222            pad_token_id=pad_token_id,223            bos_token_id=bos_token_id,224            eos_token_id=eos_token_id,225            tie_word_embeddings=tie_word_embeddings,226            auto_map=auto_map,227            **kwargs,228        )229