ldwang commited on
Commit
1a8c91f
1 Parent(s): 6cf695c

Upload configuration_aquila.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. configuration_aquila.py +80 -16
configuration_aquila.py CHANGED
@@ -1,5 +1,5 @@
1
  # coding=utf-8
2
- # Copyright 2023 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
  #
4
  # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
  # and OPT implementations in this library. It has been modified from its
@@ -17,10 +17,17 @@
17
  # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
  # See the License for the specific language governing permissions and
19
  # limitations under the License.
 
 
20
  """ Aquila model configuration"""
21
 
22
- from transformers import PretrainedConfig
 
 
23
 
 
 
 
24
 
25
 
26
  class AquilaConfig(PretrainedConfig):
@@ -34,7 +41,7 @@ class AquilaConfig(PretrainedConfig):
34
 
35
 
36
  Args:
37
- vocab_size (`int`, *optional*, defaults to 32000):
38
  Vocabulary size of the Aquila model. Defines the number of different tokens that can be represented by the
39
  `inputs_ids` passed when calling [`AquilaModel`]
40
  hidden_size (`int`, *optional*, defaults to 4096):
@@ -42,24 +49,55 @@ class AquilaConfig(PretrainedConfig):
42
  intermediate_size (`int`, *optional*, defaults to 11008):
43
  Dimension of the MLP representations.
44
  num_hidden_layers (`int`, *optional*, defaults to 32):
45
- Number of hidden layers in the Transformer encoder.
46
  num_attention_heads (`int`, *optional*, defaults to 32):
47
- Number of attention heads for each attention layer in the Transformer encoder.
 
 
 
 
 
 
 
 
48
  hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
49
  The non-linear activation function (function or string) in the decoder.
50
- max_position_embeddings (`int`, *optional*, defaults to 2048):
51
- The maximum sequence length that this model might ever be used with. Typically set this to something large
52
- just in case (e.g., 512 or 1024 or 2048).
53
  initializer_range (`float`, *optional*, defaults to 0.02):
54
  The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
55
- rms_norm_eps (`float`, *optional*, defaults to 1e-12):
56
  The epsilon used by the rms normalization layers.
57
  use_cache (`bool`, *optional*, defaults to `True`):
58
  Whether or not the model should return the last key/values attentions (not used by all models). Only
59
  relevant if `config.is_decoder=True`.
60
- tie_word_embeddings(`bool`, *optional*, defaults to `False`):
 
 
 
 
 
 
 
 
 
 
 
61
  Whether to tie weight embeddings
62
- Example:
 
 
 
 
 
 
 
 
 
 
 
 
 
63
 
64
  ```python
65
  >>> from transformers import AquilaModel, AquilaConfig
@@ -73,29 +111,32 @@ class AquilaConfig(PretrainedConfig):
73
  >>> # Accessing the model configuration
74
  >>> configuration = model.config
75
  ```"""
 
76
  model_type = "aquila"
77
  keys_to_ignore_at_inference = ["past_key_values"]
78
 
79
  def __init__(
80
  self,
81
- vocab_size=100008,
82
  hidden_size=4096,
83
  intermediate_size=11008,
84
  num_hidden_layers=32,
85
  num_attention_heads=32,
86
  num_key_value_heads=None,
87
  hidden_act="silu",
88
- max_position_embeddings=2048,
89
  initializer_range=0.02,
90
  rms_norm_eps=1e-6,
91
  use_cache=True,
92
- pad_token_id=0,
93
  bos_token_id=1,
94
  eos_token_id=2,
95
  pretraining_tp=1,
96
  tie_word_embeddings=False,
97
  rope_theta=10000.0,
98
  rope_scaling=None,
 
 
99
  **kwargs,
100
  ):
101
  self.vocab_size = vocab_size
@@ -103,14 +144,13 @@ class AquilaConfig(PretrainedConfig):
103
  self.hidden_size = hidden_size
104
  self.intermediate_size = intermediate_size
105
  self.num_hidden_layers = num_hidden_layers
 
106
 
107
  # for backward compatibility
108
  if num_key_value_heads is None:
109
  num_key_value_heads = num_attention_heads
110
 
111
  self.num_key_value_heads = num_key_value_heads
112
-
113
- self.num_attention_heads = num_attention_heads
114
  self.hidden_act = hidden_act
115
  self.initializer_range = initializer_range
116
  self.rms_norm_eps = rms_norm_eps
@@ -118,6 +158,9 @@ class AquilaConfig(PretrainedConfig):
118
  self.use_cache = use_cache
119
  self.rope_theta = rope_theta
120
  self.rope_scaling = rope_scaling
 
 
 
121
 
122
  super().__init__(
123
  pad_token_id=pad_token_id,
@@ -126,3 +169,24 @@ class AquilaConfig(PretrainedConfig):
126
  tie_word_embeddings=tie_word_embeddings,
127
  **kwargs,
128
  )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
  #
4
  # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
  # and OPT implementations in this library. It has been modified from its
 
17
  # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
  # See the License for the specific language governing permissions and
19
  # limitations under the License.
20
+
21
+ # Most of the source code is adapted from Llama's source code
22
  """ Aquila model configuration"""
23
 
24
+ from transformers.configuration_utils import PretrainedConfig
25
+ from transformers.utils import logging
26
+
27
 
28
+ logger = logging.get_logger(__name__)
29
+
30
+ AQUILA_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
31
 
32
 
33
  class AquilaConfig(PretrainedConfig):
 
41
 
42
 
43
  Args:
44
+ vocab_size (`int`, *optional*, defaults to 143973):
45
  Vocabulary size of the Aquila model. Defines the number of different tokens that can be represented by the
46
  `inputs_ids` passed when calling [`AquilaModel`]
47
  hidden_size (`int`, *optional*, defaults to 4096):
 
49
  intermediate_size (`int`, *optional*, defaults to 11008):
50
  Dimension of the MLP representations.
51
  num_hidden_layers (`int`, *optional*, defaults to 32):
52
+ Number of hidden layers in the Transformer decoder.
53
  num_attention_heads (`int`, *optional*, defaults to 32):
54
+ Number of attention heads for each attention layer in the Transformer decoder.
55
+ num_key_value_heads (`int`, *optional*):
56
+ This is the number of key_value heads that should be used to implement Grouped Query Attention. If
57
+ `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
58
+ `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
59
+ converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
60
+ by meanpooling all the original heads within that group. For more details checkout [this
61
+ paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
62
+ `num_attention_heads`.
63
  hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
64
  The non-linear activation function (function or string) in the decoder.
65
+ max_position_embeddings (`int`, *optional*, defaults to 8192):
66
+ The maximum sequence length that this model might ever be used with.
 
67
  initializer_range (`float`, *optional*, defaults to 0.02):
68
  The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
69
+ rms_norm_eps (`float`, *optional*, defaults to 1e-06):
70
  The epsilon used by the rms normalization layers.
71
  use_cache (`bool`, *optional*, defaults to `True`):
72
  Whether or not the model should return the last key/values attentions (not used by all models). Only
73
  relevant if `config.is_decoder=True`.
74
+ pad_token_id (`int`, *optional*):
75
+ Padding token id.
76
+ bos_token_id (`int`, *optional*, defaults to 1):
77
+ Beginning of stream token id.
78
+ eos_token_id (`int`, *optional*, defaults to 2):
79
+ End of stream token id.
80
+ pretraining_tp (`int`, *optional*, defaults to 1):
81
+ Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
82
+ document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to understand more about it. This value is
83
+ necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
84
+ issue](https://github.com/pytorch/pytorch/issues/76232).
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. See the following thread for more information on how
94
+ these scaling strategies behave:
95
+ https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This is an
96
+ experimental feature, subject to breaking API changes in future versions.
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
 
102
  ```python
103
  >>> from transformers import AquilaModel, AquilaConfig
 
111
  >>> # Accessing the model configuration
112
  >>> configuration = model.config
113
  ```"""
114
+
115
  model_type = "aquila"
116
  keys_to_ignore_at_inference = ["past_key_values"]
117
 
118
  def __init__(
119
  self,
120
+ vocab_size=143973,
121
  hidden_size=4096,
122
  intermediate_size=11008,
123
  num_hidden_layers=32,
124
  num_attention_heads=32,
125
  num_key_value_heads=None,
126
  hidden_act="silu",
127
+ max_position_embeddings=8192,
128
  initializer_range=0.02,
129
  rms_norm_eps=1e-6,
130
  use_cache=True,
131
+ pad_token_id=None,
132
  bos_token_id=1,
133
  eos_token_id=2,
134
  pretraining_tp=1,
135
  tie_word_embeddings=False,
136
  rope_theta=10000.0,
137
  rope_scaling=None,
138
+ attention_bias=False,
139
+ attention_dropout=0.0,
140
  **kwargs,
141
  ):
142
  self.vocab_size = vocab_size
 
144
  self.hidden_size = hidden_size
145
  self.intermediate_size = intermediate_size
146
  self.num_hidden_layers = num_hidden_layers
147
+ self.num_attention_heads = num_attention_heads
148
 
149
  # for backward compatibility
150
  if num_key_value_heads is None:
151
  num_key_value_heads = num_attention_heads
152
 
153
  self.num_key_value_heads = num_key_value_heads
 
 
154
  self.hidden_act = hidden_act
155
  self.initializer_range = initializer_range
156
  self.rms_norm_eps = rms_norm_eps
 
158
  self.use_cache = use_cache
159
  self.rope_theta = rope_theta
160
  self.rope_scaling = rope_scaling
161
+ self._rope_scaling_validation()
162
+ self.attention_bias = attention_bias
163
+ self.attention_dropout = attention_dropout
164
 
165
  super().__init__(
166
  pad_token_id=pad_token_id,
 
169
  tie_word_embeddings=tie_word_embeddings,
170
  **kwargs,
171
  )
172
+
173
+ def _rope_scaling_validation(self):
174
+ """
175
+ Validate the `rope_scaling` configuration.
176
+ """
177
+ if self.rope_scaling is None:
178
+ return
179
+
180
+ if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:
181
+ raise ValueError(
182
+ "`rope_scaling` must be a dictionary with with two fields, `type` and `factor`, "
183
+ f"got {self.rope_scaling}"
184
+ )
185
+ rope_scaling_type = self.rope_scaling.get("type", None)
186
+ rope_scaling_factor = self.rope_scaling.get("factor", None)
187
+ if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]:
188
+ raise ValueError(
189
+ f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"
190
+ )
191
+ if rope_scaling_factor is None or not isinstance(rope_scaling_factor, float) or rope_scaling_factor <= 1.0:
192
+ raise ValueError(f"`rope_scaling`'s factor field must be a float > 1, got {rope_scaling_factor}")