Update configuration_doge.py
Browse files- configuration_doge.py +52 -38
configuration_doge.py
CHANGED
@@ -5,10 +5,9 @@
|
|
5 |
# modular_doge.py file directly. One of our CI enforces this.
|
6 |
# π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨
|
7 |
# coding=utf-8
|
8 |
-
# Copyright
|
9 |
#
|
10 |
-
#
|
11 |
-
# The Doge family of small language models is trained by Jingze Shi.
|
12 |
#
|
13 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
14 |
# you may not use this file except in compliance with the License.
|
@@ -28,22 +27,20 @@ from transformers.modeling_rope_utils import rope_config_validation
|
|
28 |
class DogeConfig(PretrainedConfig):
|
29 |
r"""
|
30 |
This is the configuration class to store the configuration of a [`DogeModel`]. It is used to instantiate an Doge
|
31 |
-
model according to the specified arguments, defining the model architecture like [SmallDoge/Doge-
|
32 |
|
33 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
34 |
documentation from [`PretrainedConfig`] for more information.
|
35 |
|
36 |
Args:
|
37 |
vocab_size (`int`, *optional*, defaults to 32768):
|
38 |
-
Vocabulary size of the
|
39 |
hidden_size (`int`, *optional*, defaults to 1024):
|
40 |
Dimension of the hidden representations.
|
41 |
intermediate_size (`int`, *optional*, defaults to 2048):
|
42 |
Dimension of the MLP representations.
|
43 |
num_hidden_layers (`int`, *optional*, defaults to 32):
|
44 |
Number of hidden layers in the Transformer decoder.
|
45 |
-
hidden_bias (`bool`, *optional*, defaults to `False`):
|
46 |
-
Whether to use bias in the hidden layers.
|
47 |
hidden_dropout (`float`, *optional*, defaults to 0.0):
|
48 |
Dropout probability for each sequence transformation and state transformation module.
|
49 |
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
@@ -55,14 +52,8 @@ class DogeConfig(PretrainedConfig):
|
|
55 |
use_cache (`bool`, *optional*, defaults to `True`):
|
56 |
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
57 |
relevant if `config.is_decoder=True`.
|
58 |
-
bos_token_id (`int`, *optional*, defaults to 0):
|
59 |
-
Beginning of stream token id.
|
60 |
-
eos_token_id (`int`, *optional*, defaults to 1):
|
61 |
-
End of stream token id.
|
62 |
-
pad_token_id (`int`, *optional*, defaults to 2):
|
63 |
-
Padding token id.
|
64 |
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
65 |
-
Whether
|
66 |
max_position_embeddings (`int`, *optional*, defaults to 2048):
|
67 |
The maximum sequence length that this model might ever be used with.
|
68 |
rope_theta (`float`, *optional*, defaults to 10000.0):
|
@@ -109,18 +100,29 @@ class DogeConfig(PretrainedConfig):
|
|
109 |
When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group.
|
110 |
For more details checkout [this paper](https://arxiv.org/pdf/2305.13245.pdf).
|
111 |
If it is not specified, will default to `num_attention_heads`.
|
|
|
|
|
112 |
attention_dropout (`float`, *optional*, defaults to 0.0):
|
113 |
The dropout ratio for the attention probabilities.
|
|
|
|
|
|
|
|
|
114 |
keep_window_size (`int`, *optional*, defaults to 2048):
|
115 |
The window size of tokens that are not dynamically masked, and dynamic masking is only performed when the sequence length exceeds this value.
|
116 |
-
dynamic_mask_ratio (`float`, *optional*, defaults to 0.0):
|
117 |
-
The ratio to control the proportion of the dynamic mask filled with the minimum value. For more details checkout [this paper](https://arxiv.org/pdf/2412.11834).
|
118 |
is_moe (`bool`, *optional*, defaults to `False`):
|
119 |
-
Whether to use the Cross Domain Mixture of Experts, if `True`, the MoE will inherit the MLP to initialize.
|
120 |
-
num_experts (`int`, *optional*, defaults to
|
121 |
-
Number of
|
122 |
-
num_experts_per_tok (`int`, *optional*, defaults to
|
123 |
Number of selected experts to route per-token.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
|
125 |
```python
|
126 |
>>> from transformers import DogeConfig, DogeModel
|
@@ -144,12 +146,22 @@ class DogeConfig(PretrainedConfig):
|
|
144 |
"layers.*.self_attn.v_proj": "colwise",
|
145 |
"layers.*.self_attn.dt_proj": "rowwise",
|
146 |
"layers.*.self_attn.o_proj": "rowwise",
|
147 |
-
"layers.*.
|
148 |
-
"layers.*.
|
149 |
-
"layers.*.
|
150 |
-
"layers.*.
|
151 |
-
"
|
152 |
-
"layers.*.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
}
|
154 |
|
155 |
def __init__(
|
@@ -158,27 +170,28 @@ class DogeConfig(PretrainedConfig):
|
|
158 |
hidden_size=1024,
|
159 |
intermediate_size=2048,
|
160 |
num_hidden_layers=32,
|
161 |
-
hidden_bias=False,
|
162 |
hidden_dropout=0.0,
|
163 |
hidden_act="silu",
|
164 |
initializer_range=0.02,
|
165 |
rms_norm_eps=1e-06,
|
166 |
use_cache=True,
|
167 |
-
bos_token_id=0,
|
168 |
-
eos_token_id=1,
|
169 |
-
pad_token_id=2,
|
170 |
tie_word_embeddings=False,
|
171 |
max_position_embeddings=2048,
|
172 |
rope_theta=10000.0,
|
173 |
rope_scaling=None,
|
174 |
num_attention_heads=8,
|
175 |
num_key_value_heads=None,
|
|
|
176 |
attention_dropout=0.0,
|
|
|
|
|
177 |
keep_window_size=2048,
|
178 |
-
dynamic_mask_ratio=0.0,
|
179 |
is_moe=False,
|
180 |
-
num_experts=
|
181 |
-
num_experts_per_tok=
|
|
|
|
|
|
|
182 |
**kwargs,
|
183 |
):
|
184 |
self.vocab_size = vocab_size
|
@@ -186,7 +199,6 @@ class DogeConfig(PretrainedConfig):
|
|
186 |
self.intermediate_size = intermediate_size
|
187 |
self.num_hidden_layers = num_hidden_layers
|
188 |
|
189 |
-
self.hidden_bias = hidden_bias
|
190 |
self.hidden_dropout = hidden_dropout
|
191 |
self.hidden_act = hidden_act
|
192 |
self.initializer_range = initializer_range
|
@@ -198,12 +210,17 @@ class DogeConfig(PretrainedConfig):
|
|
198 |
self.rope_scaling = rope_scaling
|
199 |
self.num_attention_heads = num_attention_heads
|
200 |
self.num_key_value_heads = num_key_value_heads
|
|
|
201 |
self.attention_dropout = attention_dropout
|
|
|
|
|
202 |
self.keep_window_size = keep_window_size
|
203 |
-
self.dynamic_mask_ratio = dynamic_mask_ratio
|
204 |
self.is_moe = is_moe
|
205 |
self.num_experts = num_experts
|
206 |
self.num_experts_per_tok = num_experts_per_tok
|
|
|
|
|
|
|
207 |
|
208 |
# Validate the correctness of rotary position embeddings parameters
|
209 |
# BC: if there is a 'type' field, copy it it to 'rope_type'.
|
@@ -216,9 +233,6 @@ class DogeConfig(PretrainedConfig):
|
|
216 |
self.num_key_value_heads = num_attention_heads
|
217 |
|
218 |
super().__init__(
|
219 |
-
bos_token_id=bos_token_id,
|
220 |
-
eos_token_id=eos_token_id,
|
221 |
-
pad_token_id=pad_token_id,
|
222 |
tie_word_embeddings=tie_word_embeddings,
|
223 |
**kwargs,
|
224 |
)
|
|
|
5 |
# modular_doge.py file directly. One of our CI enforces this.
|
6 |
# π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨π¨
|
7 |
# coding=utf-8
|
8 |
+
# Copyright 2025 Jingze Shi and the HuggingFace Inc. team. All rights reserved.
|
9 |
#
|
10 |
+
# The Doge family of small language models is trained by SmallDoge Team.
|
|
|
11 |
#
|
12 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
13 |
# you may not use this file except in compliance with the License.
|
|
|
27 |
class DogeConfig(PretrainedConfig):
|
28 |
r"""
|
29 |
This is the configuration class to store the configuration of a [`DogeModel`]. It is used to instantiate an Doge
|
30 |
+
model according to the specified arguments, defining the model architecture like [SmallDoge/Doge-320M](https://huggingface.co/SmallDoge/Doge-320M).
|
31 |
|
32 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
33 |
documentation from [`PretrainedConfig`] for more information.
|
34 |
|
35 |
Args:
|
36 |
vocab_size (`int`, *optional*, defaults to 32768):
|
37 |
+
Vocabulary size of the Doge2 model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`DogeModel`]
|
38 |
hidden_size (`int`, *optional*, defaults to 1024):
|
39 |
Dimension of the hidden representations.
|
40 |
intermediate_size (`int`, *optional*, defaults to 2048):
|
41 |
Dimension of the MLP representations.
|
42 |
num_hidden_layers (`int`, *optional*, defaults to 32):
|
43 |
Number of hidden layers in the Transformer decoder.
|
|
|
|
|
44 |
hidden_dropout (`float`, *optional*, defaults to 0.0):
|
45 |
Dropout probability for each sequence transformation and state transformation module.
|
46 |
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
|
|
52 |
use_cache (`bool`, *optional*, defaults to `True`):
|
53 |
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
54 |
relevant if `config.is_decoder=True`.
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
56 |
+
Whether the model's input and output word embeddings should be tied.
|
57 |
max_position_embeddings (`int`, *optional*, defaults to 2048):
|
58 |
The maximum sequence length that this model might ever be used with.
|
59 |
rope_theta (`float`, *optional*, defaults to 10000.0):
|
|
|
100 |
When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group.
|
101 |
For more details checkout [this paper](https://arxiv.org/pdf/2305.13245.pdf).
|
102 |
If it is not specified, will default to `num_attention_heads`.
|
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 |
+
mlp_bias (`bool`, *optional*, defaults to `False`):
|
108 |
+
Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
|
109 |
+
sliding_window (`int`, *optional*):
|
110 |
+
Sliding window attention window size. If not specified, will default to `None`.
|
111 |
keep_window_size (`int`, *optional*, defaults to 2048):
|
112 |
The window size of tokens that are not dynamically masked, and dynamic masking is only performed when the sequence length exceeds this value.
|
|
|
|
|
113 |
is_moe (`bool`, *optional*, defaults to `False`):
|
114 |
+
Whether to use the Cross Domain Mixture of Experts, if `True`, the MoE will inherit the MLP to initialize.
|
115 |
+
num_experts (`int`, *optional*, defaults to 16384):
|
116 |
+
Number of routed experts in the model. This is only used when `is_moe=True`.
|
117 |
+
num_experts_per_tok (`int`, *optional*, defaults to 64):
|
118 |
Number of selected experts to route per-token.
|
119 |
+
norm_topk_prob (`bool`, *optional*, defaults to `False`):
|
120 |
+
Whether to normalize the topk probabilities.
|
121 |
+
output_router_logits (`bool`, *optional*, defaults to `False`):
|
122 |
+
Whether or not the router logits should be returned by the model. Enabling this will also
|
123 |
+
allow the model to output the auxiliary loss, including load balancing loss and router z-loss.
|
124 |
+
router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
|
125 |
+
The aux loss factor for the total loss.
|
126 |
|
127 |
```python
|
128 |
>>> from transformers import DogeConfig, DogeModel
|
|
|
146 |
"layers.*.self_attn.v_proj": "colwise",
|
147 |
"layers.*.self_attn.dt_proj": "rowwise",
|
148 |
"layers.*.self_attn.o_proj": "rowwise",
|
149 |
+
"layers.*.input_layernorm.weight": "sequence_parallel",
|
150 |
+
"layers.*.input_residual.weight": "sequence_parallel",
|
151 |
+
"layers.*.post_attention_layernorm.weight": "sequence_parallel",
|
152 |
+
"layers.*.post_attention_residual.weight": "sequence_parallel",
|
153 |
+
"norm.weight": "sequence_parallel",
|
154 |
+
"layers.*.mlp.gate_proj": "colwise",
|
155 |
+
"layers.*.mlp.up_proj": "colwise",
|
156 |
+
"layers.*.mlp.down_proj": "rowwise",
|
157 |
+
"layers.*.mlp.router_gate": "colwise_rep",
|
158 |
+
"layers.*.mlp.down_embed": "rowwise_rep",
|
159 |
+
"layers.*.mlp.up_embed": "rowwise_rep",
|
160 |
+
}
|
161 |
+
base_model_pp_plan = {
|
162 |
+
"embed_tokens": (["input_ids"], ["inputs_embeds"]),
|
163 |
+
"layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
|
164 |
+
"norm": (["hidden_states"], ["hidden_states"]),
|
165 |
}
|
166 |
|
167 |
def __init__(
|
|
|
170 |
hidden_size=1024,
|
171 |
intermediate_size=2048,
|
172 |
num_hidden_layers=32,
|
|
|
173 |
hidden_dropout=0.0,
|
174 |
hidden_act="silu",
|
175 |
initializer_range=0.02,
|
176 |
rms_norm_eps=1e-06,
|
177 |
use_cache=True,
|
|
|
|
|
|
|
178 |
tie_word_embeddings=False,
|
179 |
max_position_embeddings=2048,
|
180 |
rope_theta=10000.0,
|
181 |
rope_scaling=None,
|
182 |
num_attention_heads=8,
|
183 |
num_key_value_heads=None,
|
184 |
+
attention_bias=False,
|
185 |
attention_dropout=0.0,
|
186 |
+
mlp_bias=False,
|
187 |
+
sliding_window=None,
|
188 |
keep_window_size=2048,
|
|
|
189 |
is_moe=False,
|
190 |
+
num_experts=16384,
|
191 |
+
num_experts_per_tok=64,
|
192 |
+
norm_topk_prob=False,
|
193 |
+
output_router_logits=False,
|
194 |
+
router_aux_loss_coef=0.001,
|
195 |
**kwargs,
|
196 |
):
|
197 |
self.vocab_size = vocab_size
|
|
|
199 |
self.intermediate_size = intermediate_size
|
200 |
self.num_hidden_layers = num_hidden_layers
|
201 |
|
|
|
202 |
self.hidden_dropout = hidden_dropout
|
203 |
self.hidden_act = hidden_act
|
204 |
self.initializer_range = initializer_range
|
|
|
210 |
self.rope_scaling = rope_scaling
|
211 |
self.num_attention_heads = num_attention_heads
|
212 |
self.num_key_value_heads = num_key_value_heads
|
213 |
+
self.attention_bias = attention_bias
|
214 |
self.attention_dropout = attention_dropout
|
215 |
+
self.mlp_bias = mlp_bias
|
216 |
+
self.sliding_window = sliding_window
|
217 |
self.keep_window_size = keep_window_size
|
|
|
218 |
self.is_moe = is_moe
|
219 |
self.num_experts = num_experts
|
220 |
self.num_experts_per_tok = num_experts_per_tok
|
221 |
+
self.norm_topk_prob = norm_topk_prob
|
222 |
+
self.output_router_logits = output_router_logits
|
223 |
+
self.router_aux_loss_coef = router_aux_loss_coef
|
224 |
|
225 |
# Validate the correctness of rotary position embeddings parameters
|
226 |
# BC: if there is a 'type' field, copy it it to 'rope_type'.
|
|
|
233 |
self.num_key_value_heads = num_attention_heads
|
234 |
|
235 |
super().__init__(
|
|
|
|
|
|
|
236 |
tie_word_embeddings=tie_word_embeddings,
|
237 |
**kwargs,
|
238 |
)
|