Crystalcareai
commited on
Commit
•
9a24831
1
Parent(s):
7a9d077
Upload configuration_deepseek.py
Browse files- configuration_deepseek.py +200 -0
configuration_deepseek.py
ADDED
@@ -0,0 +1,200 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 GemmoeConfig(PretrainedConfig):
|
8 |
+
r"""
|
9 |
+
This is the configuration class to store the configuration of a [`DeepseekModel`]. 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-7B.
|
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 [`DeepseekModel`]
|
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 |
+
num_experts_per_tok (`int`, *optional*, defaults to None):
|
36 |
+
Number of selected experts, None means dense model.
|
37 |
+
moe_layer_freq (`int`, *optional*, defaults to 1):
|
38 |
+
The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
|
39 |
+
first_k_dense_replace (`int`, *optional*, defaults to 0):
|
40 |
+
Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
|
41 |
+
\--k dense layers--/
|
42 |
+
norm_topk_prob (`bool`, *optional*, defaults to False):
|
43 |
+
Whether to normalize the weights of the routed experts.
|
44 |
+
scoring_func (`str`, *optional*, defaults to 'softmax'):
|
45 |
+
Method of computing expert weights.
|
46 |
+
aux_loss_alpha (`float`, *optional*, defaults to 0.001):
|
47 |
+
Auxiliary loss weight coefficient.
|
48 |
+
seq_aux = (`bool`, *optional*, defaults to True):
|
49 |
+
Whether to compute the auxiliary loss for each individual sample.
|
50 |
+
num_key_value_heads (`int`, *optional*):
|
51 |
+
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
52 |
+
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
53 |
+
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
|
54 |
+
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
|
55 |
+
by meanpooling all the original heads within that group. For more details checkout [this
|
56 |
+
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
|
57 |
+
`num_attention_heads`.
|
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 2048):
|
61 |
+
The maximum sequence length that this model might ever be used with.
|
62 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
63 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
64 |
+
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
|
65 |
+
The epsilon used by the rms normalization layers.
|
66 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
67 |
+
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
68 |
+
relevant if `config.is_decoder=True`.
|
69 |
+
pad_token_id (`int`, *optional*):
|
70 |
+
Padding token id.
|
71 |
+
bos_token_id (`int`, *optional*, defaults to 1):
|
72 |
+
Beginning of stream token id.
|
73 |
+
eos_token_id (`int`, *optional*, defaults to 2):
|
74 |
+
End of stream token id.
|
75 |
+
pretraining_tp (`int`, *optional*, defaults to 1):
|
76 |
+
Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
|
77 |
+
document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
|
78 |
+
necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
|
79 |
+
issue](https://github.com/pytorch/pytorch/issues/76232).
|
80 |
+
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
81 |
+
Whether to tie weight embeddings
|
82 |
+
rope_theta (`float`, *optional*, defaults to 10000.0):
|
83 |
+
The base period of the RoPE embeddings.
|
84 |
+
rope_scaling (`Dict`, *optional*):
|
85 |
+
Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
|
86 |
+
strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
|
87 |
+
`{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
|
88 |
+
`max_position_embeddings` to the expected new maximum.
|
89 |
+
attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
|
90 |
+
Whether to use a bias in the query, key, value and output projection layers during self-attention.
|
91 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
92 |
+
The dropout ratio for the attention probabilities.
|
93 |
+
|
94 |
+
```python
|
95 |
+
>>> from transformers import DeepseekModel, DeepseekConfig
|
96 |
+
|
97 |
+
>>> # Initializing a Deepseek deepseek-7b style configuration
|
98 |
+
>>> configuration = DeepseekConfig()
|
99 |
+
|
100 |
+
>>> # Accessing the model configuration
|
101 |
+
>>> configuration = model.config
|
102 |
+
```"""
|
103 |
+
|
104 |
+
model_type = "gemmoe"
|
105 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
106 |
+
|
107 |
+
def __init__(
|
108 |
+
self,
|
109 |
+
vocab_size=256000,
|
110 |
+
hidden_size=3072,
|
111 |
+
intermediate_size=24576,
|
112 |
+
moe_intermediate_size = 1407,
|
113 |
+
num_hidden_layers=28,
|
114 |
+
num_attention_heads=16,
|
115 |
+
num_key_value_heads=16,
|
116 |
+
n_shared_experts = None,
|
117 |
+
n_routed_experts = None,
|
118 |
+
num_experts_per_tok = None,
|
119 |
+
moe_layer_freq = 1,
|
120 |
+
first_k_dense_replace = 0,
|
121 |
+
norm_topk_prob = False,
|
122 |
+
scoring_func = 'softmax',
|
123 |
+
aux_loss_alpha = 0.001,
|
124 |
+
seq_aux = True,
|
125 |
+
hidden_act="gelu",
|
126 |
+
max_position_embeddings=2048,
|
127 |
+
initializer_range=0.02,
|
128 |
+
rms_norm_eps=1e-6,
|
129 |
+
use_cache=True,
|
130 |
+
pad_token_id=0,
|
131 |
+
bos_token_id=2,
|
132 |
+
eos_token_id=1,
|
133 |
+
pretraining_tp=1,
|
134 |
+
tie_word_embeddings=True,
|
135 |
+
rope_theta=10000.0,
|
136 |
+
rope_scaling=None,
|
137 |
+
attention_bias=False,
|
138 |
+
attention_dropout=0.0,
|
139 |
+
**kwargs,
|
140 |
+
):
|
141 |
+
self.vocab_size = vocab_size
|
142 |
+
self.max_position_embeddings = max_position_embeddings
|
143 |
+
self.hidden_size = hidden_size
|
144 |
+
self.intermediate_size = intermediate_size
|
145 |
+
self.moe_intermediate_size = moe_intermediate_size
|
146 |
+
self.num_hidden_layers = num_hidden_layers
|
147 |
+
self.num_attention_heads = num_attention_heads
|
148 |
+
self.n_shared_experts = n_shared_experts
|
149 |
+
self.n_routed_experts = n_routed_experts
|
150 |
+
self.num_experts_per_tok = num_experts_per_tok
|
151 |
+
self.moe_layer_freq = moe_layer_freq
|
152 |
+
self.first_k_dense_replace = first_k_dense_replace
|
153 |
+
self.norm_topk_prob = norm_topk_prob
|
154 |
+
self.scoring_func = scoring_func
|
155 |
+
self.aux_loss_alpha = aux_loss_alpha
|
156 |
+
self.seq_aux = seq_aux
|
157 |
+
# for backward compatibility
|
158 |
+
if num_key_value_heads is None:
|
159 |
+
num_key_value_heads = num_attention_heads
|
160 |
+
|
161 |
+
self.num_key_value_heads = num_key_value_heads
|
162 |
+
self.hidden_act = hidden_act
|
163 |
+
self.initializer_range = initializer_range
|
164 |
+
self.rms_norm_eps = rms_norm_eps
|
165 |
+
self.pretraining_tp = pretraining_tp
|
166 |
+
self.use_cache = use_cache
|
167 |
+
self.rope_theta = rope_theta
|
168 |
+
self.rope_scaling = rope_scaling
|
169 |
+
self._rope_scaling_validation()
|
170 |
+
self.attention_bias = attention_bias
|
171 |
+
self.attention_dropout = attention_dropout
|
172 |
+
|
173 |
+
super().__init__(
|
174 |
+
pad_token_id=pad_token_id,
|
175 |
+
bos_token_id=bos_token_id,
|
176 |
+
eos_token_id=eos_token_id,
|
177 |
+
tie_word_embeddings=tie_word_embeddings,
|
178 |
+
**kwargs,
|
179 |
+
)
|
180 |
+
|
181 |
+
def _rope_scaling_validation(self):
|
182 |
+
"""
|
183 |
+
Validate the `rope_scaling` configuration.
|
184 |
+
"""
|
185 |
+
if self.rope_scaling is None:
|
186 |
+
return
|
187 |
+
|
188 |
+
if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:
|
189 |
+
raise ValueError(
|
190 |
+
"`rope_scaling` must be a dictionary with with two fields, `type` and `factor`, "
|
191 |
+
f"got {self.rope_scaling}"
|
192 |
+
)
|
193 |
+
rope_scaling_type = self.rope_scaling.get("type", None)
|
194 |
+
rope_scaling_factor = self.rope_scaling.get("factor", None)
|
195 |
+
if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]:
|
196 |
+
raise ValueError(
|
197 |
+
f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"
|
198 |
+
)
|
199 |
+
if rope_scaling_factor is None or not isinstance(rope_scaling_factor, float) or rope_scaling_factor <= 1.0:
|
200 |
+
raise ValueError(f"`rope_scaling`'s factor field must be a float > 1, got {rope_scaling_factor}")
|