configuration_deepseek.py
| 1 | # Copy from https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/configuration_deepseek.py |
| 2 | |
| 3 | from transformers.configuration_utils import PretrainedConfig |
| 4 | from transformers.utils import logging |
| 5 | |
| 6 | logger = logging.get_logger(__name__) |
| 7 | |
| 8 | DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {} |
| 9 | |
| 10 | |
| 11 | class DeepseekV3Config(PretrainedConfig): |
| 12 | r""" |
| 13 | This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek |
| 14 | model according to the specified arguments, defining the model architecture. Instantiating a configuration with the |
| 15 | defaults will yield a similar configuration to that of the DeepSeek-V3. |
| 16 | |
| 17 | Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the |
| 18 | documentation from [`PretrainedConfig`] for more information. |
| 19 | |
| 20 | |
| 21 | Args: |
| 22 | vocab_size (`int`, *optional*, defaults to 129280): |
| 23 | Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the |
| 24 | `inputs_ids` passed when calling [`DeepseekV3Model`] |
| 25 | hidden_size (`int`, *optional*, defaults to 4096): |
| 26 | Dimension of the hidden representations. |
| 27 | intermediate_size (`int`, *optional*, defaults to 11008): |
| 28 | Dimension of the MLP representations. |
| 29 | moe_intermediate_size (`int`, *optional*, defaults to 1407): |
| 30 | Dimension of the MoE representations. |
| 31 | num_hidden_layers (`int`, *optional*, defaults to 32): |
| 32 | Number of hidden layers in the Transformer decoder. |
| 33 | num_nextn_predict_layers (`int`, *optional*, defaults to 1): |
| 34 | Number of nextn predict layers in the DeepSeekV3 Model. |
| 35 | num_attention_heads (`int`, *optional*, defaults to 32): |
| 36 | Number of attention heads for each attention layer in the Transformer decoder. |
| 37 | n_shared_experts (`int`, *optional*, defaults to None): |
| 38 | Number of shared experts, None means dense model. |
| 39 | n_routed_experts (`int`, *optional*, defaults to None): |
| 40 | Number of routed experts, None means dense model. |
| 41 | routed_scaling_factor (`float`, *optional*, defaults to 1.0): |
| 42 | Scaling factor or routed experts. |
| 43 | topk_method (`str`, *optional*, defaults to `gready`): |
| 44 | Topk method used in routed gate. |
| 45 | n_group (`int`, *optional*, defaults to None): |
| 46 | Number of groups for routed experts. |
| 47 | topk_group (`int`, *optional*, defaults to None): |
| 48 | Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups). |
| 49 | num_experts_per_tok (`int`, *optional*, defaults to None): |
| 50 | Number of selected experts, None means dense model. |
| 51 | moe_layer_freq (`int`, *optional*, defaults to 1): |
| 52 | The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers. |
| 53 | first_k_dense_replace (`int`, *optional*, defaults to 0): |
| 54 | Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head). |
| 55 | \--k dense layers--/ |
| 56 | norm_topk_prob (`bool`, *optional*, defaults to False): |
| 57 | Whether to normalize the weights of the routed experts. |
| 58 | scoring_func (`str`, *optional*, defaults to 'softmax'): |
| 59 | Method of computing expert weights. |
| 60 | aux_loss_alpha (`float`, *optional*, defaults to 0.001): |
| 61 | Auxiliary loss weight coefficient. |
| 62 | seq_aux = (`bool`, *optional*, defaults to True): |
| 63 | Whether to compute the auxiliary loss for each individual sample. |
| 64 | num_key_value_heads (`int`, *optional*): |
| 65 | This is the number of key_value heads that should be used to implement Grouped Query Attention. If |
| 66 | `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if |
| 67 | `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When |
| 68 | converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed |
| 69 | by meanpooling all the original heads within that group. For more details checkout [this |
| 70 | paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to |
| 71 | `num_attention_heads`. |
| 72 | hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): |
| 73 | The non-linear activation function (function or string) in the decoder. |
| 74 | max_position_embeddings (`int`, *optional*, defaults to 2048): |
| 75 | The maximum sequence length that this model might ever be used with. |
| 76 | initializer_range (`float`, *optional*, defaults to 0.02): |
| 77 | The standard deviation of the truncated_normal_initializer for initializing all weight matrices. |
| 78 | rms_norm_eps (`float`, *optional*, defaults to 1e-06): |
| 79 | The epsilon used by the rms normalization layers. |
| 80 | use_cache (`bool`, *optional*, defaults to `True`): |
| 81 | Whether or not the model should return the last key/values attentions (not used by all models). Only |
| 82 | relevant if `config.is_decoder=True`. |
| 83 | pad_token_id (`int`, *optional*): |
| 84 | Padding token id. |
| 85 | bos_token_id (`int`, *optional*, defaults to 1): |
| 86 | Beginning of stream token id. |
| 87 | eos_token_id (`int`, *optional*, defaults to 2): |
| 88 | End of stream token id. |
| 89 | pretraining_tp (`int`, *optional*, defaults to 1): |
| 90 | Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this |
| 91 | document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is |
| 92 | necessary to ensure exact reproducibility of the pretraining results. Please refer to [this |
| 93 | issue](https://github.com/pytorch/pytorch/issues/76232). |
| 94 | tie_word_embeddings (`bool`, *optional*, defaults to `False`): |
| 95 | Whether to tie weight embeddings |
| 96 | rope_theta (`float`, *optional*, defaults to 10000.0): |
| 97 | The base period of the RoPE embeddings. |
| 98 | rope_scaling (`Dict`, *optional*): |
| 99 | Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling |
| 100 | strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is |
| 101 | `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update |
| 102 | `max_position_embeddings` to the expected new maximum. |
| 103 | attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): |
| 104 | Whether to use a bias in the query, key, value and output projection layers during self-attention. |
| 105 | attention_dropout (`float`, *optional*, defaults to 0.0): |
| 106 | The dropout ratio for the attention probabilities. |
| 107 | |
| 108 | ```python |
| 109 | >>> from transformers import DeepseekV3Model, DeepseekV3Config |
| 110 | |
| 111 | >>> # Initializing a Deepseek-V3 style configuration |
| 112 | >>> configuration = DeepseekV3Config() |
| 113 | |
| 114 | >>> # Accessing the model configuration |
| 115 | >>> configuration = model.config |
| 116 | ```""" |
| 117 | |
| 118 | model_type = "deepseek_v3" |
| 119 | keys_to_ignore_at_inference = ["past_key_values"] |
| 120 | |
| 121 | def __init__( |
| 122 | self, |
| 123 | vocab_size=129280, |
| 124 | hidden_size=7168, |
| 125 | intermediate_size=18432, |
| 126 | moe_intermediate_size=2048, |
| 127 | num_hidden_layers=61, |
| 128 | num_nextn_predict_layers=1, |
| 129 | num_attention_heads=128, |
| 130 | num_key_value_heads=128, |
| 131 | n_shared_experts=1, |
| 132 | n_routed_experts=256, |
| 133 | ep_size=1, |
| 134 | routed_scaling_factor=2.5, |
| 135 | kv_lora_rank=512, |
| 136 | q_lora_rank=1536, |
| 137 | qk_rope_head_dim=64, |
| 138 | v_head_dim=128, |
| 139 | qk_nope_head_dim=128, |
| 140 | topk_method='noaux_tc', |
| 141 | n_group=8, |
| 142 | topk_group=4, |
| 143 | num_experts_per_tok=8, |
| 144 | moe_layer_freq=1, |
| 145 | first_k_dense_replace=3, |
| 146 | norm_topk_prob=True, |
| 147 | scoring_func='sigmoid', |
| 148 | aux_loss_alpha=0.001, |
| 149 | seq_aux=True, |
| 150 | hidden_act="silu", |
| 151 | max_position_embeddings=4096, |
| 152 | initializer_range=0.02, |
| 153 | rms_norm_eps=1e-6, |
| 154 | use_cache=True, |
| 155 | pad_token_id=None, |
| 156 | bos_token_id=0, |
| 157 | eos_token_id=1, |
| 158 | pretraining_tp=1, |
| 159 | tie_word_embeddings=False, |
| 160 | rope_theta=10000.0, |
| 161 | rope_scaling=None, |
| 162 | attention_bias=False, |
| 163 | attention_dropout=0.0, |
| 164 | **kwargs, |
| 165 | ): |
| 166 | self.vocab_size = vocab_size |
| 167 | self.max_position_embeddings = max_position_embeddings |
| 168 | self.hidden_size = hidden_size |
| 169 | self.intermediate_size = intermediate_size |
| 170 | self.moe_intermediate_size = moe_intermediate_size |
| 171 | self.num_hidden_layers = num_hidden_layers |
| 172 | self.num_nextn_predict_layers = num_nextn_predict_layers |
| 173 | self.num_attention_heads = num_attention_heads |
| 174 | self.n_shared_experts = n_shared_experts |
| 175 | self.n_routed_experts = n_routed_experts |
| 176 | self.ep_size = ep_size |
| 177 | self.routed_scaling_factor = routed_scaling_factor |
| 178 | self.kv_lora_rank = kv_lora_rank |
| 179 | self.q_lora_rank = q_lora_rank |
| 180 | self.qk_rope_head_dim = qk_rope_head_dim |
| 181 | self.v_head_dim = v_head_dim |
| 182 | self.qk_nope_head_dim = qk_nope_head_dim |
| 183 | self.topk_method = topk_method |
| 184 | self.n_group = n_group |
| 185 | self.topk_group = topk_group |
| 186 | self.num_experts_per_tok = num_experts_per_tok |
| 187 | self.moe_layer_freq = moe_layer_freq |
| 188 | self.first_k_dense_replace = first_k_dense_replace |
| 189 | self.norm_topk_prob = norm_topk_prob |
| 190 | self.scoring_func = scoring_func |
| 191 | self.aux_loss_alpha = aux_loss_alpha |
| 192 | self.seq_aux = seq_aux |
| 193 | # for backward compatibility |
| 194 | if num_key_value_heads is None: |
| 195 | num_key_value_heads = num_attention_heads |
| 196 | |
| 197 | self.num_key_value_heads = num_key_value_heads |
| 198 | self.hidden_act = hidden_act |
| 199 | self.initializer_range = initializer_range |
| 200 | self.rms_norm_eps = rms_norm_eps |
| 201 | self.pretraining_tp = pretraining_tp |
| 202 | self.use_cache = use_cache |
| 203 | self.rope_theta = rope_theta |
| 204 | self.rope_scaling = rope_scaling |
| 205 | self.attention_bias = attention_bias |
| 206 | self.attention_dropout = attention_dropout |
| 207 | |
| 208 | super().__init__( |
| 209 | pad_token_id=pad_token_id, |
| 210 | bos_token_id=bos_token_id, |
| 211 | eos_token_id=eos_token_id, |
| 212 | tie_word_embeddings=tie_word_embeddings, |
| 213 | **kwargs, |
| 214 | ) |
| 215 | |