configuration_deepseek.py
9.7 KB · 199 lines · python Raw
1 from transformers.configuration_utils import PretrainedConfig
2 from transformers.utils import logging
3
4 logger = logging.get_logger(__name__)
5
6 DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
7 class DeepseekV3Config(PretrainedConfig):
8 r"""
9 This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek
10 model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
11 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 the
14 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 the
20 `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. If
62 `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
63 `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
64 converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
65 by meanpooling all the original heads within that group. For more details checkout [this
66 paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
67 `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). Only
78 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 tie_word_embeddings (`bool`, *optional*, defaults to `False`):
86 Whether to tie weight embeddings
87 rope_theta (`float`, *optional*, defaults to 10000.0):
88 The base period of the RoPE embeddings.
89 rope_scaling (`Dict`, *optional*):
90 Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
91 strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
92 `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
93 `max_position_embeddings` to the expected new maximum.
94 attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
95 Whether to use a bias in the query, key, value and output projection layers during self-attention.
96 attention_dropout (`float`, *optional*, defaults to 0.0):
97 The dropout ratio for the attention probabilities.
98
99 ```python
100 >>> from transformers import DeepseekV3Model, DeepseekV3Config
101
102 >>> # Initializing a Deepseek-V3 style configuration
103 >>> configuration = DeepseekV3Config()
104
105 >>> # Accessing the model configuration
106 >>> configuration = model.config
107 ```"""
108
109 model_type = "deepseek_v3"
110 keys_to_ignore_at_inference = ["past_key_values"]
111
112 def __init__(
113 self,
114 vocab_size=129280,
115 hidden_size=7168,
116 intermediate_size=18432,
117 moe_intermediate_size = 2048,
118 num_hidden_layers=61,
119 num_nextn_predict_layers=1,
120 num_attention_heads=128,
121 num_key_value_heads=128,
122 n_shared_experts = 1,
123 n_routed_experts = 256,
124 ep_size = 1,
125 routed_scaling_factor = 2.5,
126 kv_lora_rank = 512,
127 q_lora_rank = 1536,
128 qk_rope_head_dim = 64,
129 v_head_dim = 128,
130 qk_nope_head_dim = 128,
131 topk_method = 'noaux_tc',
132 n_group = 8,
133 topk_group = 4,
134 num_experts_per_tok = 8,
135 moe_layer_freq = 1,
136 first_k_dense_replace = 3,
137 norm_topk_prob = True,
138 scoring_func = 'sigmoid',
139 hidden_act="silu",
140 max_position_embeddings=4096,
141 initializer_range=0.02,
142 rms_norm_eps=1e-6,
143 use_cache=True,
144 pad_token_id=None,
145 bos_token_id=0,
146 eos_token_id=1,
147 tie_word_embeddings=False,
148 rope_theta=10000.0,
149 rope_scaling=None,
150 attention_bias=False,
151 attention_dropout=0.0,
152 **kwargs,
153 ):
154 self.vocab_size = vocab_size
155 self.max_position_embeddings = max_position_embeddings
156 self.hidden_size = hidden_size
157 self.intermediate_size = intermediate_size
158 self.moe_intermediate_size = moe_intermediate_size
159 self.num_hidden_layers = num_hidden_layers
160 self.num_nextn_predict_layers = num_nextn_predict_layers
161 self.num_attention_heads = num_attention_heads
162 self.n_shared_experts = n_shared_experts
163 self.n_routed_experts = n_routed_experts
164 self.ep_size = ep_size
165 self.routed_scaling_factor = routed_scaling_factor
166 self.kv_lora_rank = kv_lora_rank
167 self.q_lora_rank = q_lora_rank
168 self.qk_rope_head_dim = qk_rope_head_dim
169 self.v_head_dim = v_head_dim
170 self.qk_nope_head_dim = qk_nope_head_dim
171 self.topk_method = topk_method
172 self.n_group = n_group
173 self.topk_group = topk_group
174 self.num_experts_per_tok = num_experts_per_tok
175 self.moe_layer_freq = moe_layer_freq
176 self.first_k_dense_replace = first_k_dense_replace
177 self.norm_topk_prob = norm_topk_prob
178 self.scoring_func = scoring_func
179 # for backward compatibility
180 if num_key_value_heads is None:
181 num_key_value_heads = num_attention_heads
182
183 self.num_key_value_heads = num_key_value_heads
184 self.hidden_act = hidden_act
185 self.initializer_range = initializer_range
186 self.rms_norm_eps = rms_norm_eps
187 self.use_cache = use_cache
188 self.rope_theta = rope_theta
189 self.rope_scaling = rope_scaling
190 self.attention_bias = attention_bias
191 self.attention_dropout = attention_dropout
192
193 super().__init__(
194 pad_token_id=pad_token_id,
195 bos_token_id=bos_token_id,
196 eos_token_id=eos_token_id,
197 tie_word_embeddings=tie_word_embeddings,
198 **kwargs,
199 )