configuration_deepseek_v2.py
10.4 KB · 211 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 DeepseekV2Config(PretrainedConfig):
8 r"""
9 This is the configuration class to store the configuration of a [`DeepseekV2Model`]. 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-V2 with multi-latent attention.
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 102400):
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 [`DeepseekV2Model`]
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_attention_heads (`int`, *optional*, defaults to 32):
30 Number of attention heads for each attention layer in the Transformer decoder.
31 n_shared_experts (`int`, *optional*, defaults to None):
32 Number of shared experts, None means dense model.
33 n_routed_experts (`int`, *optional*, defaults to None):
34 Number of routed experts, None means dense model.
35 routed_scaling_factor (`float`, *optional*, defaults to 1.0):
36 Scaling factor or routed experts.
37 topk_method (`str`, *optional*, defaults to `gready`):
38 Topk method used in routed gate.
39 n_group (`int`, *optional*, defaults to None):
40 Number of groups for routed experts.
41 topk_group (`int`, *optional*, defaults to None):
42 Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
43 num_experts_per_tok (`int`, *optional*, defaults to None):
44 Number of selected experts, None means dense model.
45 moe_layer_freq (`int`, *optional*, defaults to 1):
46 The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
47 first_k_dense_replace (`int`, *optional*, defaults to 0):
48 Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
49 \--k dense layers--/
50 norm_topk_prob (`bool`, *optional*, defaults to False):
51 Whether to normalize the weights of the routed experts.
52 scoring_func (`str`, *optional*, defaults to 'softmax'):
53 Method of computing expert weights.
54 aux_loss_alpha (`float`, *optional*, defaults to 0.001):
55 Auxiliary loss weight coefficient.
56 seq_aux = (`bool`, *optional*, defaults to True):
57 Whether to compute the auxiliary loss for each individual sample.
58 num_key_value_heads (`int`, *optional*):
59 This is the number of key_value heads that should be used to implement Grouped Query Attention. If
60 `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
61 `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
62 converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
63 by meanpooling all the original heads within that group. For more details checkout [this
64 paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
65 `num_attention_heads`.
66 hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
67 The non-linear activation function (function or string) in the decoder.
68 max_position_embeddings (`int`, *optional*, defaults to 2048):
69 The maximum sequence length that this model might ever be used with.
70 initializer_range (`float`, *optional*, defaults to 0.02):
71 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
72 rms_norm_eps (`float`, *optional*, defaults to 1e-06):
73 The epsilon used by the rms normalization layers.
74 use_cache (`bool`, *optional*, defaults to `True`):
75 Whether or not the model should return the last key/values attentions (not used by all models). Only
76 relevant if `config.is_decoder=True`.
77 pad_token_id (`int`, *optional*):
78 Padding token id.
79 bos_token_id (`int`, *optional*, defaults to 1):
80 Beginning of stream token id.
81 eos_token_id (`int`, *optional*, defaults to 2):
82 End of stream token id.
83 pretraining_tp (`int`, *optional*, defaults to 1):
84 Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
85 document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
86 necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
87 issue](https://github.com/pytorch/pytorch/issues/76232).
88 tie_word_embeddings (`bool`, *optional*, defaults to `False`):
89 Whether to tie weight embeddings
90 rope_theta (`float`, *optional*, defaults to 10000.0):
91 The base period of the RoPE embeddings.
92 rope_scaling (`Dict`, *optional*):
93 Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
94 strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
95 `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
96 `max_position_embeddings` to the expected new maximum.
97 attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
98 Whether to use a bias in the query, key, value and output projection layers during self-attention.
99 attention_dropout (`float`, *optional*, defaults to 0.0):
100 The dropout ratio for the attention probabilities.
101 use_mla (`bool`, *optional*, defaults to `True`): Use multi-latent attention or multi-head attention. If True,
102 the model will use multi-latent attention, otherwise, it will use multi-head attention.
103
104 ```python
105 >>> from transformers import DeepseekV2Model, DeepseekV2Config
106
107 >>> # Initializing a Deepseek-V2 style configuration
108 >>> configuration = DeepseekV2Config()
109
110 >>> # Accessing the model configuration
111 >>> configuration = model.config
112 ```"""
113
114 model_type = "deepseek_v2"
115 keys_to_ignore_at_inference = ["past_key_values"]
116
117 def __init__(
118 self,
119 vocab_size=102400,
120 hidden_size=4096,
121 intermediate_size=11008,
122 moe_intermediate_size = 1407,
123 num_hidden_layers=30,
124 num_attention_heads=32,
125 num_key_value_heads=32,
126 n_shared_experts = None,
127 n_routed_experts = None,
128 ep_size = 1,
129 routed_scaling_factor = 1.0,
130 kv_lora_rank = 512,
131 q_lora_rank = 1536,
132 qk_rope_head_dim = 64,
133 v_head_dim = 128,
134 qk_nope_head_dim = 128,
135 topk_method = 'gready',
136 n_group = None,
137 topk_group = None,
138 num_experts_per_tok = None,
139 moe_layer_freq = 1,
140 first_k_dense_replace = 0,
141 norm_topk_prob = False,
142 scoring_func = 'softmax',
143 aux_loss_alpha = 0.001,
144 seq_aux = True,
145 hidden_act="silu",
146 max_position_embeddings=2048,
147 initializer_range=0.02,
148 rms_norm_eps=1e-6,
149 use_cache=True,
150 pad_token_id=None,
151 bos_token_id=100000,
152 eos_token_id=100001,
153 pretraining_tp=1,
154 tie_word_embeddings=False,
155 rope_theta=10000.0,
156 rope_scaling=None,
157 attention_bias=False,
158 attention_dropout=0.0,
159 use_mla=True,
160 **kwargs,
161 ):
162 self.vocab_size = vocab_size
163 self.max_position_embeddings = max_position_embeddings
164 self.hidden_size = hidden_size
165 self.intermediate_size = intermediate_size
166 self.moe_intermediate_size = moe_intermediate_size
167 self.num_hidden_layers = num_hidden_layers
168 self.num_attention_heads = num_attention_heads
169 self.n_shared_experts = n_shared_experts
170 self.n_routed_experts = n_routed_experts
171 self.ep_size = ep_size
172 self.routed_scaling_factor = routed_scaling_factor
173 self.kv_lora_rank = kv_lora_rank
174 self.q_lora_rank = q_lora_rank
175 self.qk_rope_head_dim = qk_rope_head_dim
176 self.v_head_dim = v_head_dim
177 self.qk_nope_head_dim = qk_nope_head_dim
178 self.topk_method = topk_method
179 self.n_group = n_group
180 self.topk_group = topk_group
181 self.num_experts_per_tok = num_experts_per_tok
182 self.moe_layer_freq = moe_layer_freq
183 self.first_k_dense_replace = first_k_dense_replace
184 self.norm_topk_prob = norm_topk_prob
185 self.scoring_func = scoring_func
186 self.aux_loss_alpha = aux_loss_alpha
187 self.seq_aux = seq_aux
188 # for backward compatibility
189 if num_key_value_heads is None:
190 num_key_value_heads = num_attention_heads
191
192 self.num_key_value_heads = num_key_value_heads
193 self.hidden_act = hidden_act
194 self.initializer_range = initializer_range
195 self.rms_norm_eps = float(rms_norm_eps)
196 self.pretraining_tp = pretraining_tp
197 self.use_cache = use_cache
198 self.rope_theta = rope_theta
199 self.rope_scaling = rope_scaling
200 self.attention_bias = attention_bias
201 self.attention_dropout = attention_dropout
202 self.use_mla = use_mla
203
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 )
211