configuration_qwen2.py
6.9 KB · 149 lines · python Raw
1 # coding=utf-8
2 # Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
3 #
4 # Licensed under the Apache License, Version 2.0 (the "License");
5 # you may not use this file except in compliance with the License.
6 # You may obtain a copy of the License at
7 #
8 # http://www.apache.org/licenses/LICENSE-2.0
9 #
10 # Unless required by applicable law or agreed to in writing, software
11 # distributed under the License is distributed on an "AS IS" BASIS,
12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 # See the License for the specific language governing permissions and
14 # limitations under the License.
15 """ Qwen2 model configuration"""
16
17 from transformers.configuration_utils import PretrainedConfig
18 from transformers.utils import logging
19
20
21 logger = logging.get_logger(__name__)
22
23 QWEN2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
24 "Qwen/Qwen2-7B-beta": "https://huggingface.co/Qwen/Qwen2-7B-beta/resolve/main/config.json",
25 }
26
27
28 class Qwen2Config(PretrainedConfig):
29 r"""
30 This is the configuration class to store the configuration of a [`Qwen2Model`]. It is used to instantiate a
31 Qwen2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
32 with the defaults will yield a similar configuration to that of
33 Qwen2-7B-beta [Qwen/Qwen2-7B-beta](https://huggingface.co/Qwen/Qwen2-7B-beta).
34
35 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
36 documentation from [`PretrainedConfig`] for more information.
37
38
39 Args:
40 vocab_size (`int`, *optional*, defaults to 151936):
41 Vocabulary size of the Qwen2 model. Defines the number of different tokens that can be represented by the
42 `inputs_ids` passed when calling [`Qwen2Model`]
43 hidden_size (`int`, *optional*, defaults to 4096):
44 Dimension of the hidden representations.
45 intermediate_size (`int`, *optional*, defaults to 22016):
46 Dimension of the MLP representations.
47 num_hidden_layers (`int`, *optional*, defaults to 32):
48 Number of hidden layers in the Transformer encoder.
49 num_attention_heads (`int`, *optional*, defaults to 32):
50 Number of attention heads for each attention layer in the Transformer encoder.
51 num_key_value_heads (`int`, *optional*, defaults to 32):
52 This is the number of key_value heads that should be used to implement Grouped Query Attention. If
53 `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
54 `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
55 converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
56 by meanpooling all the original heads within that group. For more details checkout [this
57 paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
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 32768):
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 tie_word_embeddings (`bool`, *optional*, defaults to `False`):
70 Whether the model's input and output word embeddings should be tied.
71 rope_theta (`float`, *optional*, defaults to 10000.0):
72 The base period of the RoPE embeddings.
73 use_sliding_window (`bool`, *optional*, defaults to `False`):
74 Whether to use sliding window attention.
75 sliding_window (`int`, *optional*, defaults to 4096):
76 Sliding window attention (SWA) window size. If not specified, will default to `4096`.
77 max_window_layers (`int`, *optional*, defaults to 28):
78 The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
79 attention_dropout (`float`, *optional*, defaults to 0.0):
80 The dropout ratio for the attention probabilities.
81
82 ```python
83 >>> from transformers import Qwen2Model, Qwen2Config
84
85 >>> # Initializing a Qwen2 style configuration
86 >>> configuration = Qwen2Config()
87
88 >>> # Initializing a model from the Qwen2-7B style configuration
89 >>> model = Qwen2Model(configuration)
90
91 >>> # Accessing the model configuration
92 >>> configuration = model.config
93 ```"""
94
95 model_type = "qwen2"
96 keys_to_ignore_at_inference = ["past_key_values"]
97
98 def __init__(
99 self,
100 vocab_size=151936,
101 hidden_size=4096,
102 intermediate_size=22016,
103 num_hidden_layers=32,
104 num_attention_heads=32,
105 num_key_value_heads=32,
106 hidden_act="silu",
107 max_position_embeddings=32768,
108 initializer_range=0.02,
109 rms_norm_eps=1e-6,
110 use_cache=True,
111 tie_word_embeddings=False,
112 rope_theta=10000.0,
113 use_sliding_window=False,
114 sliding_window=4096,
115 max_window_layers=28,
116 attention_dropout=0.0,
117 **kwargs,
118 ):
119 self.vocab_size = vocab_size
120 self.max_position_embeddings = max_position_embeddings
121 self.hidden_size = hidden_size
122 self.intermediate_size = intermediate_size
123 self.num_hidden_layers = num_hidden_layers
124 self.num_attention_heads = num_attention_heads
125 self.use_sliding_window = use_sliding_window
126 self.sliding_window = sliding_window
127 self.max_window_layers = max_window_layers
128
129 # for backward compatibility
130 if num_key_value_heads is None:
131 num_key_value_heads = num_attention_heads
132
133 self.num_key_value_heads = num_key_value_heads
134 self.hidden_act = hidden_act
135 self.initializer_range = initializer_range
136 self.rms_norm_eps = rms_norm_eps
137 self.use_cache = use_cache
138 self.rope_theta = rope_theta
139 self.attention_dropout = attention_dropout
140 if kwargs.get('attn_implementation', None) is None:
141 self.attn_implementation = kwargs['attn_implementation'] = 'flash_attention_2'
142 else:
143 self.attn_implementation = kwargs['attn_implementation']
144
145 super().__init__(
146 tie_word_embeddings=tie_word_embeddings,
147 **kwargs,
148 )
149