RedHatAI/Kimi-K2-Instruct-quantized.w4a16

1# Copy from https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/configuration_deepseek.py2 3from transformers.configuration_utils import PretrainedConfig4from transformers.utils import logging5 6logger = logging.get_logger(__name__)7 8DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}9class DeepseekV3Config(PretrainedConfig):10    r"""11    This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek12    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the13    defaults will yield a similar configuration to that of the DeepSeek-V3.14 15    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the16    documentation from [`PretrainedConfig`] for more information.17 18 19    Args:20        vocab_size (`int`, *optional*, defaults to 129280):21            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the22            `inputs_ids` passed when calling [`DeepseekV3Model`]23        hidden_size (`int`, *optional*, defaults to 4096):24            Dimension of the hidden representations.25        intermediate_size (`int`, *optional*, defaults to 11008):26            Dimension of the MLP representations.27        moe_intermediate_size (`int`, *optional*, defaults to 1407):28            Dimension of the MoE representations.29        num_hidden_layers (`int`, *optional*, defaults to 32):30            Number of hidden layers in the Transformer decoder.31        num_nextn_predict_layers (`int`, *optional*, defaults to 1):32            Number of nextn predict layers in the DeepSeekV3 Model.33        num_attention_heads (`int`, *optional*, defaults to 32):34            Number of attention heads for each attention layer in the Transformer decoder.35        n_shared_experts (`int`, *optional*, defaults to None):36            Number of shared experts, None means dense model.37        n_routed_experts (`int`, *optional*, defaults to None):38            Number of routed experts, None means dense model.39        routed_scaling_factor (`float`, *optional*, defaults to 1.0):40            Scaling factor or routed experts.41        topk_method (`str`, *optional*, defaults to `gready`):42            Topk method used in routed gate.43        n_group (`int`, *optional*, defaults to None):44            Number of groups for routed experts.45        topk_group (`int`, *optional*, defaults to None):46            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).47        num_experts_per_tok (`int`, *optional*, defaults to None):48            Number of selected experts, None means dense model.49        moe_layer_freq (`int`, *optional*, defaults to 1):50            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.51        first_k_dense_replace (`int`, *optional*, defaults to 0):52            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).53                                                            \--k dense layers--/54        norm_topk_prob (`bool`, *optional*, defaults to False):55            Whether to normalize the weights of the routed experts.56        scoring_func (`str`, *optional*, defaults to 'softmax'):57            Method of computing expert weights.58        aux_loss_alpha (`float`, *optional*, defaults to 0.001):59            Auxiliary loss weight coefficient.60        seq_aux = (`bool`, *optional*, defaults to True):61            Whether to compute the auxiliary loss for each individual sample.62        num_key_value_heads (`int`, *optional*):63            This is the number of key_value heads that should be used to implement Grouped Query Attention. If64            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if65            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When66            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed67            by meanpooling all the original heads within that group. For more details checkout [this68            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to69            `num_attention_heads`.70        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):71            The non-linear activation function (function or string) in the decoder.72        max_position_embeddings (`int`, *optional*, defaults to 2048):73            The maximum sequence length that this model might ever be used with.74        initializer_range (`float`, *optional*, defaults to 0.02):75            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.76        rms_norm_eps (`float`, *optional*, defaults to 1e-06):77            The epsilon used by the rms normalization layers.78        use_cache (`bool`, *optional*, defaults to `True`):79            Whether or not the model should return the last key/values attentions (not used by all models). Only80            relevant if `config.is_decoder=True`.81        pad_token_id (`int`, *optional*):82            Padding token id.83        bos_token_id (`int`, *optional*, defaults to 1):84            Beginning of stream token id.85        eos_token_id (`int`, *optional*, defaults to 2):86            End of stream token id.87        pretraining_tp (`int`, *optional*, defaults to 1):88            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this89            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is90            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this91            issue](https://github.com/pytorch/pytorch/issues/76232).92        tie_word_embeddings (`bool`, *optional*, defaults to `False`):93            Whether to tie weight embeddings94        rope_theta (`float`, *optional*, defaults to 10000.0):95            The base period of the RoPE embeddings.96        rope_scaling (`Dict`, *optional*):97            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling98            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is99            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update100            `max_position_embeddings` to the expected new maximum.101        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):102            Whether to use a bias in the query, key, value and output projection layers during self-attention.103        attention_dropout (`float`, *optional*, defaults to 0.0):104            The dropout ratio for the attention probabilities.105 106    ```python107    >>> from transformers import DeepseekV3Model, DeepseekV3Config108 109    >>> # Initializing a Deepseek-V3 style configuration110    >>> configuration = DeepseekV3Config()111 112    >>> # Accessing the model configuration113    >>> configuration = model.config114    ```"""115 116    model_type = "deepseek_v3"117    keys_to_ignore_at_inference = ["past_key_values"]118 119    def __init__(120        self,121        vocab_size=129280,122        hidden_size=7168,123        intermediate_size=18432,124        moe_intermediate_size = 2048,125        num_hidden_layers=61,126        num_nextn_predict_layers=1,127        num_attention_heads=128,128        num_key_value_heads=128,129        n_shared_experts = 1,130        n_routed_experts = 256,131        ep_size = 1,132        routed_scaling_factor = 2.5,133        kv_lora_rank = 512,134        q_lora_rank = 1536,135        qk_rope_head_dim = 64,136        v_head_dim = 128,137        qk_nope_head_dim = 128,138        topk_method = 'noaux_tc',139        n_group = 8,140        topk_group = 4,141        num_experts_per_tok = 8,142        moe_layer_freq = 1,143        first_k_dense_replace = 3,144        norm_topk_prob = True,145        scoring_func = 'sigmoid',146        aux_loss_alpha = 0.001,147        seq_aux = True,148        hidden_act="silu",149        max_position_embeddings=4096,150        initializer_range=0.02,151        rms_norm_eps=1e-6,152        use_cache=True,153        pad_token_id=None,154        bos_token_id=0,155        eos_token_id=1,156        pretraining_tp=1,157        tie_word_embeddings=False,158        rope_theta=10000.0,159        rope_scaling=None,160        attention_bias=False,161        attention_dropout=0.0,162        **kwargs,163    ):164        self.vocab_size = vocab_size165        self.max_position_embeddings = max_position_embeddings166        self.hidden_size = hidden_size167        self.intermediate_size = intermediate_size168        self.moe_intermediate_size = moe_intermediate_size169        self.num_hidden_layers = num_hidden_layers170        self.num_nextn_predict_layers = num_nextn_predict_layers171        self.num_attention_heads = num_attention_heads172        self.n_shared_experts = n_shared_experts173        self.n_routed_experts = n_routed_experts174        self.ep_size = ep_size175        self.routed_scaling_factor = routed_scaling_factor176        self.kv_lora_rank = kv_lora_rank177        self.q_lora_rank = q_lora_rank178        self.qk_rope_head_dim = qk_rope_head_dim179        self.v_head_dim = v_head_dim180        self.qk_nope_head_dim = qk_nope_head_dim181        self.topk_method = topk_method182        self.n_group = n_group183        self.topk_group = topk_group184        self.num_experts_per_tok = num_experts_per_tok185        self.moe_layer_freq = moe_layer_freq186        self.first_k_dense_replace = first_k_dense_replace187        self.norm_topk_prob = norm_topk_prob188        self.scoring_func = scoring_func189        self.aux_loss_alpha = aux_loss_alpha190        self.seq_aux = seq_aux191        # for backward compatibility192        if num_key_value_heads is None:193            num_key_value_heads = num_attention_heads194 195        self.num_key_value_heads = num_key_value_heads196        self.hidden_act = hidden_act197        self.initializer_range = initializer_range198        self.rms_norm_eps = rms_norm_eps199        self.pretraining_tp = pretraining_tp200        self.use_cache = use_cache201        self.rope_theta = rope_theta202        self.rope_scaling = rope_scaling203        self.attention_bias = attention_bias204        self.attention_dropout = attention_dropout205 206        super().__init__(207            pad_token_id=pad_token_id,208            bos_token_id=bos_token_id,209            eos_token_id=eos_token_id,210            tie_word_embeddings=tie_word_embeddings,211            **kwargs,212        )