amd/DeepSeek-R1-MXFP4

1from transformers.configuration_utils import PretrainedConfig2from transformers.utils import logging3 4logger = logging.get_logger(__name__)5 6DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}7class DeepseekV3Config(PretrainedConfig):8    r"""9    This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek10    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the11    defaults will yield a similar configuration to that of the DeepSeek-V3.12 13    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the14    documentation from [`PretrainedConfig`] for more information.15 16 17    Args:18        vocab_size (`int`, *optional*, defaults to 129280):19            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the20            `inputs_ids` passed when calling [`DeepseekV3Model`]21        hidden_size (`int`, *optional*, defaults to 4096):22            Dimension of the hidden representations.23        intermediate_size (`int`, *optional*, defaults to 11008):24            Dimension of the MLP representations.25        moe_intermediate_size (`int`, *optional*, defaults to 1407):26            Dimension of the MoE representations.27        num_hidden_layers (`int`, *optional*, defaults to 32):28            Number of hidden layers in the Transformer decoder.29        num_nextn_predict_layers (`int`, *optional*, defaults to 1):30            Number of nextn predict layers in the DeepSeekV3 Model.31        num_attention_heads (`int`, *optional*, defaults to 32):32            Number of attention heads for each attention layer in the Transformer decoder.33        n_shared_experts (`int`, *optional*, defaults to None):34            Number of shared experts, None means dense model.35        n_routed_experts (`int`, *optional*, defaults to None):36            Number of routed experts, None means dense model.37        routed_scaling_factor (`float`, *optional*, defaults to 1.0):38            Scaling factor or routed experts.39        topk_method (`str`, *optional*, defaults to `gready`):40            Topk method used in routed gate.41        n_group (`int`, *optional*, defaults to None):42            Number of groups for routed experts.43        topk_group (`int`, *optional*, defaults to None):44            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).45        num_experts_per_tok (`int`, *optional*, defaults to None):46            Number of selected experts, None means dense model.47        moe_layer_freq (`int`, *optional*, defaults to 1):48            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.49        first_k_dense_replace (`int`, *optional*, defaults to 0):50            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).51                                                            \--k dense layers--/52        norm_topk_prob (`bool`, *optional*, defaults to False):53            Whether to normalize the weights of the routed experts.54        scoring_func (`str`, *optional*, defaults to 'softmax'):55            Method of computing expert weights.56        aux_loss_alpha (`float`, *optional*, defaults to 0.001):57            Auxiliary loss weight coefficient.58        seq_aux = (`bool`, *optional*, defaults to True):59            Whether to compute the auxiliary loss for each individual sample.60        num_key_value_heads (`int`, *optional*):61            This is the number of key_value heads that should be used to implement Grouped Query Attention. If62            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if63            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When64            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed65            by meanpooling all the original heads within that group. For more details checkout [this66            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to67            `num_attention_heads`.68        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):69            The non-linear activation function (function or string) in the decoder.70        max_position_embeddings (`int`, *optional*, defaults to 2048):71            The maximum sequence length that this model might ever be used with.72        initializer_range (`float`, *optional*, defaults to 0.02):73            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.74        rms_norm_eps (`float`, *optional*, defaults to 1e-06):75            The epsilon used by the rms normalization layers.76        use_cache (`bool`, *optional*, defaults to `True`):77            Whether or not the model should return the last key/values attentions (not used by all models). Only78            relevant if `config.is_decoder=True`.79        pad_token_id (`int`, *optional*):80            Padding token id.81        bos_token_id (`int`, *optional*, defaults to 1):82            Beginning of stream token id.83        eos_token_id (`int`, *optional*, defaults to 2):84            End of stream token id.85        pretraining_tp (`int`, *optional*, defaults to 1):86            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this87            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is88            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this89            issue](https://github.com/pytorch/pytorch/issues/76232).90        tie_word_embeddings (`bool`, *optional*, defaults to `False`):91            Whether to tie weight embeddings92        rope_theta (`float`, *optional*, defaults to 10000.0):93            The base period of the RoPE embeddings.94        rope_scaling (`Dict`, *optional*):95            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling96            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is97            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update98            `max_position_embeddings` to the expected new maximum.99        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):100            Whether to use a bias in the query, key, value and output projection layers during self-attention.101        attention_dropout (`float`, *optional*, defaults to 0.0):102            The dropout ratio for the attention probabilities.103 104    ```python105    >>> from transformers import DeepseekV3Model, DeepseekV3Config106 107    >>> # Initializing a Deepseek-V3 style configuration108    >>> configuration = DeepseekV3Config()109 110    >>> # Accessing the model configuration111    >>> configuration = model.config112    ```"""113 114    model_type = "deepseek_v3"115    keys_to_ignore_at_inference = ["past_key_values"]116 117    def __init__(118        self,119        vocab_size=129280,120        hidden_size=7168,121        intermediate_size=18432,122        moe_intermediate_size = 2048,123        num_hidden_layers=61,124        num_nextn_predict_layers=1,125        num_attention_heads=128,126        num_key_value_heads=128,127        n_shared_experts = 1,128        n_routed_experts = 256,129        ep_size = 1,130        routed_scaling_factor = 2.5,131        kv_lora_rank = 512,132        q_lora_rank = 1536,133        qk_rope_head_dim = 64,134        v_head_dim = 128,135        qk_nope_head_dim = 128,136        topk_method = 'noaux_tc',137        n_group = 8,138        topk_group = 4,139        num_experts_per_tok = 8,140        moe_layer_freq = 1,141        first_k_dense_replace = 3,142        norm_topk_prob = True,143        scoring_func = 'sigmoid',144        aux_loss_alpha = 0.001,145        seq_aux = True,146        hidden_act="silu",147        max_position_embeddings=4096,148        initializer_range=0.02,149        rms_norm_eps=1e-6,150        use_cache=True,151        pad_token_id=None,152        bos_token_id=0,153        eos_token_id=1,154        pretraining_tp=1,155        tie_word_embeddings=False,156        rope_theta=10000.0,157        rope_scaling=None,158        attention_bias=False,159        attention_dropout=0.0,160        **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.moe_intermediate_size = moe_intermediate_size167        self.num_hidden_layers = num_hidden_layers168        self.num_nextn_predict_layers = num_nextn_predict_layers169        self.num_attention_heads = num_attention_heads170        self.n_shared_experts = n_shared_experts171        self.n_routed_experts = n_routed_experts172        self.ep_size = ep_size173        self.routed_scaling_factor = routed_scaling_factor174        self.kv_lora_rank = kv_lora_rank175        self.q_lora_rank = q_lora_rank176        self.qk_rope_head_dim = qk_rope_head_dim177        self.v_head_dim = v_head_dim178        self.qk_nope_head_dim = qk_nope_head_dim179        self.topk_method = topk_method180        self.n_group = n_group181        self.topk_group = topk_group182        self.num_experts_per_tok = num_experts_per_tok183        self.moe_layer_freq = moe_layer_freq184        self.first_k_dense_replace = first_k_dense_replace185        self.norm_topk_prob = norm_topk_prob186        self.scoring_func = scoring_func187        self.aux_loss_alpha = aux_loss_alpha188        self.seq_aux = seq_aux189        # for backward compatibility190        if num_key_value_heads is None:191            num_key_value_heads = num_attention_heads192 193        self.num_key_value_heads = num_key_value_heads194        self.hidden_act = hidden_act195        self.initializer_range = initializer_range196        self.rms_norm_eps = rms_norm_eps197        self.pretraining_tp = pretraining_tp198        self.use_cache = use_cache199        self.rope_theta = rope_theta200        self.rope_scaling = rope_scaling201        self.attention_bias = attention_bias202        self.attention_dropout = attention_dropout203 204        super().__init__(205            pad_token_id=pad_token_id,206            bos_token_id=bos_token_id,207            eos_token_id=eos_token_id,208            tie_word_embeddings=tie_word_embeddings,209            **kwargs,210        )