刚刚，DeepSeek发布全新模型，不是R2

{  "architectures": [    "DeepseekV3ForCausalLM"  ],  "attention_bias": false,  "attention_dropout": 0.0,  "auto_map": {    "AutoConfig": "configuration_deepseek.DeepseekV3Config",    "AutoModel": "modeling_deepseek.DeepseekV3Model",    "AutoModelForCausalLM": "modeling_deepseek.DeepseekV3ForCausalLM"  },  "bos_token_id": 0,  "eos_token_id": 1,  "ep_size": 1,  "first_k_dense_replace": 3,  "hidden_act": "silu",  "hidden_size": 7168,  "initializer_range": 0.02,  "intermediate_size": 18432,  "kv_lora_rank": 512,  "max_position_embeddings": 163840,  "model_type": "deepseek_v3",  "moe_intermediate_size": 2048,  "moe_layer_freq": 1,  "n_group": 8,  "n_routed_experts": 256,  "n_shared_experts": 1,  "norm_topk_prob": true,  "num_attention_heads": 128,  "num_experts_per_tok": 8,  "num_hidden_layers": 61,  "num_key_value_heads": 128,  "num_nextn_predict_layers": 1,  "q_lora_rank": 1536,  "qk_nope_head_dim": 128,  "qk_rope_head_dim": 64,  "quantization_config": {    "activation_scheme": "dynamic",    "fmt": "e4m3",    "quant_method": "fp8",    "weight_block_size": [      128,      128    ]  },  "rms_norm_eps": 1e-06,  "rope_scaling": {    "beta_fast": 32,    "beta_slow": 1,    "factor": 40,    "mscale": 1.0,    "mscale_all_dim": 1.0,    "original_max_position_embeddings": 4096,    "type": "yarn"  },  "rope_theta": 10000,  "routed_scaling_factor": 2.5,  "scoring_func": "sigmoid",  "tie_word_embeddings": false,  "topk_group": 4,  "topk_method": "noaux_tc",  "torch_dtype": "bfloat16",  "transformers_version": "4.46.3",  "use_cache": true,  "v_head_dim": 128,  "vocab_size": 129280}

from transformers.configuration_utils import PretrainedConfigfrom transformers.utils import logging
logger = logging.get_logger(__name__)
DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}class DeepseekV3Config(PretrainedConfig):    r"""    This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the    defaults will yield a similar configuration to that of the DeepSeek-V3.    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 129280):            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the            `inputs_ids` passed when calling [`DeepseekV3Model`]        hidden_size (`int`, *optional*, defaults to 4096):            Dimension of the hidden representations.        intermediate_size (`int`, *optional*, defaults to 11008):            Dimension of the MLP representations.        moe_intermediate_size (`int`, *optional*, defaults to 1407):            Dimension of the MoE representations.        num_hidden_layers (`int`, *optional*, defaults to 32):            Number of hidden layers in the Transformer decoder.        num_nextn_predict_layers (`int`, *optional*, defaults to 1):            Number of nextn predict layers in the DeepSeekV3 Model.        num_attention_heads (`int`, *optional*, defaults to 32):            Number of attention heads for each attention layer in the Transformer decoder.        n_shared_experts (`int`, *optional*, defaults to None):            Number of shared experts, None means dense model.        n_routed_experts (`int`, *optional*, defaults to None):            Number of routed experts, None means dense model.        routed_scaling_factor (`float`, *optional*, defaults to 1.0):            Scaling factor or routed experts.        topk_method (`str`, *optional*, defaults to `gready`):            Topk method used in routed gate.        n_group (`int`, *optional*, defaults to None):            Number of groups for routed experts.        topk_group (`int`, *optional*, defaults to None):            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).        num_experts_per_tok (`int`, *optional*, defaults to None):            Number of selected experts, None means dense model.        moe_layer_freq (`int`, *optional*, defaults to 1):            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.        first_k_dense_replace (`int`, *optional*, defaults to 0):            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).                                                            \--k dense layers--/        norm_topk_prob (`bool`, *optional*, defaults to False):            Whether to normalize the weights of the routed experts.        scoring_func (`str`, *optional*, defaults to 'softmax'):            Method of computing expert weights.        aux_loss_alpha (`float`, *optional*, defaults to 0.001):            Auxiliary loss weight coefficient.        seq_aux = (`bool`, *optional*, defaults to True):            Whether to compute the auxiliary loss for each individual sample.        num_key_value_heads (`int`, *optional*):            This is the number of key_value heads that should be used to implement Grouped Query Attention. If            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed            by meanpooling all the original heads within that group. For more details checkout [this            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to            `num_attention_heads`.        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):            The non-linear activation function (function or string) in the decoder.        max_position_embeddings (`int`, *optional*, defaults to 2048):            The maximum sequence length that this model might ever be used with.        initializer_range (`float`, *optional*, defaults to 0.02):            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.        rms_norm_eps (`float`, *optional*, defaults to 1e-06):            The epsilon used by the rms normalization layers.        use_cache (`bool`, *optional*, defaults to `True`):            Whether or not the model should return the last key/values attentions (not used by all models). Only            relevant if `config.is_decoder=True`.        pad_token_id (`int`, *optional*):            Padding token id.        bos_token_id (`int`, *optional*, defaults to 1):            Beginning of stream token id.        eos_token_id (`int`, *optional*, defaults to 2):            End of stream token id.        tie_word_embeddings (`bool`, *optional*, defaults to `False`):            Whether to tie weight embeddings        rope_theta (`float`, *optional*, defaults to 10000.0):            The base period of the RoPE embeddings.        rope_scaling (`Dict`, *optional*):            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update            `max_position_embeddings` to the expected new maximum.        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):            Whether to use a bias in the query, key, value and output projection layers during self-attention.        attention_dropout (`float`, *optional*, defaults to 0.0):            The dropout ratio for the attention probabilities.    ```python    >>> from transformers import DeepseekV3Model, DeepseekV3Config    >>> # Initializing a Deepseek-V3 style configuration    >>> configuration = DeepseekV3Config()    >>> # Accessing the model configuration    >>> configuration = model.config    ```"""