tiny-random/longcat-flash

1 2"""LongcatFlash model configuration"""3 4from transformers.configuration_utils import PretrainedConfig5from transformers.modeling_rope_utils import rope_config_validation6 7 8LONGCAT_PRETRAINED_CONFIG_ARCHIVE_MAP = {}9 10 11class LongcatFlashConfig(PretrainedConfig):12    r"""13    This is the configuration class to store the configuration of a [`LongcatFlashModel`]. It is used to instantiate an LongcatFlash14    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the15    defaults will yield a similar configuration to that of the LongcatFlash.16    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the17    documentation from [`PretrainedConfig`] for more information.18 19 20    Args:21        vocab_size (`int`, *optional*, defaults to 131072):22            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the23            `inputs_ids` passed when calling [`LongcatFlashModel`]24        hidden_size (`int`, *optional*, defaults to 7168):25            Dimension of the hidden representations.26        ffn_hidden_size (`int`, *optional*, defaults to 18432):27            Dimension of the MLP representations.28        expert_ffn_hidden_size (`int`, *optional*, defaults to 2048):29            Dimension of the MoE representations.30        num_layers (`int`, *optional*, defaults to 61):31            Number of hidden layers in the Transformer decoder.32        num_attention_heads (`int`, *optional*, defaults to 128):33            Number of attention heads for each attention layer in the Transformer decoder.34        num_key_value_heads (`int`, *optional*, defaults to 128):35            This is the number of key_value heads that should be used to implement Grouped Query Attention. If36            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if37            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When38            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed39            by meanpooling all the original heads within that group. For more details checkout [this40            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to41            `num_attention_heads`.42        n_routed_experts (`int`, *optional*, defaults to 256):43            Number of routed experts.44        routed_scaling_factor (`float`, *optional*, defaults to 2.5):45            Scaling factor or routed experts.46        kv_lora_rank (`int`, *optional*, defaults to 512):47            Rank of the LoRA matrices for key and value projections.48        q_lora_rank (`int`, *optional*, defaults to 1536):49            Rank of the LoRA matrices for query projections.50        qk_rope_head_dim (`int`, *optional*, defaults to 64):51            Dimension of the query/key heads that use rotary position embeddings.52        v_head_dim (`int`, *optional*, defaults to 128):53            Dimension of the value heads.54        qk_nope_head_dim (`int`, *optional*, defaults to 128):55            Dimension of the query/key heads that don't use rotary position embeddings.56        norm_topk_prob (`bool`, *optional*, defaults to `True`):57            Whether to normalize the weights of the routed experts.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 4096):61            The maximum sequence length that this model might ever be used with.62        rms_norm_eps (`float`, *optional*, defaults to 1e-06):63            The epsilon used by the rms normalization layers.64        use_cache (`bool`, *optional*, defaults to `True`):65            Whether or not the model should return the last key/values attentions (not used by all models). Only66            relevant if `config.is_decoder=True`.67        pad_token_id (`int`, *optional*):68            Padding token id.69        bos_token_id (`int`, *optional*, defaults to 0):70            Beginning of stream token id.71        eos_token_id (`int`, *optional*, defaults to 1):72            End of stream token id.73        tie_word_embeddings (`bool`, *optional*, defaults to `False`):74            Whether to tie weight embeddings75        rope_theta (`float`, *optional*, defaults to 10000.0):76            The base period of the RoPE embeddings.77        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):78            Whether to use a bias in the query, key, value and output projection layers during self-attention.79        attention_dropout (`float`, *optional*, defaults to 0.0):80            The dropout ratio for the attention probabilities.81        attention_method (`str`, *optional*, defaults to `"MLA"`):82            The attention method to use.83        initializer_range (`float`, *optional*, defaults to 0.006):84            The initializer range for the model.85        router_bias (`bool`, *optional*, defaults to `False`):86            Whether to use a bias in the router.87        zero_expert_num (`int`, *optional*, defaults to `None`):88            The number of zero experts to use.89        zero_expert_type (`str`, *optional*, defaults to `None`):90            The type of zero expert to use.91 92    ```python93    >>> from transformers import LongcatFlashModel, LongcatFlashConfig94 95    >>> # Initializing a LongcatFlash style configuration96    >>> configuration = LongcatFlashConfig()97 98    >>> # Accessing the model configuration99    >>> configuration = model.config100    ```"""101 102    model_type = "longcat_flash"103    keys_to_ignore_at_inference = ["past_key_values"]104    base_model_tp_plan = {105        "layers.*.self_attn.k_proj": "colwise",106        "layers.*.self_attn.v_proj": "colwise",107        "layers.*.self_attn.o_proj": "rowwise",108        "layers.*.mlp.experts.*.gate_proj": "local_colwise",109        "layers.*.mlp.experts.*.up_proj": "local_colwise",110        "layers.*.mlp.experts.*.down_proj": "local_rowwise",111        "layers.*.mlps.*.gate_proj": "local_colwise",112        "layers.*.mlps.*.up_proj": "local_colwise",113        "layers.*.mlps.*.down_proj": "local_rowwise",114    }115    base_model_pp_plan = {116        "embed_tokens": (["input_ids"], ["inputs_embeds"]),117        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),118        "norm": (["hidden_states"], ["hidden_states"]),119    }120 121    def __init__(122        self,123        vocab_size=131072,124        hidden_size=7168,125        ffn_hidden_size=18432,126        expert_ffn_hidden_size=2048,127        num_layers=61,128        num_attention_heads=128,129        num_key_value_heads=None,130        n_routed_experts=256,131        routed_scaling_factor=1,132        kv_lora_rank=512,133        q_lora_rank=1536,134        qk_rope_head_dim=64,135        v_head_dim=128,136        qk_nope_head_dim=128,137        mla_scale_q_lora=True,138        mla_scale_kv_lora=True,139        moe_topk=8,140        norm_topk_prob=False,141        hidden_act="silu",142        max_position_embeddings=4096,143        rms_norm_eps=1e-6,144        use_cache=True,145        pad_token_id=None,146        bos_token_id=0,147        eos_token_id=1,148        tie_word_embeddings=False,149        rope_theta=10000.0,150        attention_bias=False,151        attention_dropout=0.0,152        attention_method='MLA',153        initializer_range=0.006,154        router_bias=False,155        zero_expert_num=None,156        zero_expert_type=None,157        **kwargs,158    ):159        self.vocab_size = vocab_size160        self.max_position_embeddings = max_position_embeddings161        self.hidden_size = hidden_size162        self.ffn_hidden_size = ffn_hidden_size163        self.expert_ffn_hidden_size = expert_ffn_hidden_size164        self.num_layers = num_layers165        self.num_attention_heads = num_attention_heads166        self.n_routed_experts = n_routed_experts167        self.routed_scaling_factor = routed_scaling_factor168        self.kv_lora_rank = kv_lora_rank169        self.q_lora_rank = q_lora_rank170        self.qk_rope_head_dim = qk_rope_head_dim171        self.v_head_dim = v_head_dim172        self.qk_nope_head_dim = qk_nope_head_dim173        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim174        self.moe_topk = moe_topk175        self.norm_topk_prob = norm_topk_prob176        self.mla_scale_q_lora = mla_scale_q_lora177        self.mla_scale_kv_lora = mla_scale_kv_lora178        self.attention_method = attention_method179        self.initializer_range = initializer_range180        self.router_bias = router_bias181        self.zero_expert_num = zero_expert_num182        self.zero_expert_type = zero_expert_type183 184        if self.attention_method == "MLA":185            self.head_dim = qk_rope_head_dim186        else:187            ValueError('attention_method should be one of ["MLA"]')188 189 190        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.rms_norm_eps = rms_norm_eps196        self.use_cache = use_cache197        self.rope_theta = rope_theta198        self.attention_bias = attention_bias199        self.attention_dropout = attention_dropout200 201        rope_config_validation(self)202 203        super().__init__(204            pad_token_id=pad_token_id,205            bos_token_id=bos_token_id,206            eos_token_id=eos_token_id,207            tie_word_embeddings=tie_word_embeddings,208            **kwargs,209        )210 211    @property212    def num_hidden_layers(self):213        return self.num_layers214 215 216__all__ = ["LongcatFlashConfig"]217