configuration_hunyuan.py
13.1 KB · 286 lines · python Raw
1 # Licensed under the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT (the "License");
2 # you may not use this file except in compliance with the License.
3 # You may obtain a copy of the License at
4 #
5 # https://github.com/Tencent-Hunyuan/HunyuanImage-3.0/blob/main/LICENSE
6 #
7 # Unless required by applicable law or agreed to in writing, software
8 # distributed under the License is distributed on an "AS IS" BASIS,
9 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
10 # See the License for the specific language governing permissions and
11 # limitations under the License.
12 # ==============================================================================
13
14 from transformers.configuration_utils import PretrainedConfig
15 from transformers.utils import logging
16 from typing import List, Union
17
18
19 logger = logging.get_logger(__name__)
20
21
22 class HunyuanImage3Config(PretrainedConfig):
23 r"""
24 This is the configuration class to store the configuration of a [`HunyuanImage3Model`]. It is used to instantiate
25 an Hunyuan model according to the specified arguments, defining the model architecture. Instantiating a
26 configuration with the defaults will yield a similar configuration to that of the Hunyuan-7B.
27
28 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
29 documentation from [`PretrainedConfig`] for more information.
30
31
32 Args:
33 vocab_size (`int`, *optional*, defaults to 32000):
34 Vocabulary size of the Hunyuan Image 3 model. Defines the number of different tokens that can be
35 represented by the `inputs_ids` passed when calling [`HunyuanImage3Model`]
36 hidden_size (`int`, *optional*, defaults to 4096):
37 Dimension of the hidden representations.
38 intermediate_size (`int`, *optional*, defaults to 11008):
39 Dimension of the MLP representations or shared MLP representations.
40 moe_intermediate_size (`int` or `List`, *optional*, defaults to 11008):
41 Dimension of the MLP representations in MoE. Use a list if you want a different size per layer.
42 num_hidden_layers (`int`, *optional*, defaults to 32):
43 Number of hidden layers in the Transformer decoder.
44 num_attention_heads (`int`, *optional*, defaults to 32):
45 Number of attention heads for each attention layer in the Transformer decoder.
46 num_key_value_heads (`int`, *optional*):
47 This is the number of key_value heads that should be used to implement Grouped Query Attention. If
48 `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
49 `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
50 converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
51 by meanpooling all the original heads within that group. For more details checkout [this
52 paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
53 `num_attention_heads`.
54 hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
55 The non-linear activation function (function or string) in the decoder.
56 max_position_embeddings (`int`, *optional*, defaults to 2048):
57 The maximum sequence length that this model might ever be used with.
58 initializer_range (`float`, *optional*, defaults to 0.02):
59 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
60 rms_norm_eps (`float`, *optional*, defaults to 1e-06):
61 The epsilon used by the rms normalization layers.
62 use_cache (`bool`, *optional*, defaults to `True`):
63 Whether or not the model should return the last key/values attentions (not used by all models). Only
64 relevant if `config.is_decoder=True`.
65 pad_token_id (`int`, *optional*):
66 Padding token id.
67 bos_token_id (`int`, *optional*, defaults to 1):
68 Beginning of stream token id.
69 eos_token_id (`int`, *optional*, defaults to 2):
70 End of stream token id.
71 pretraining_tp (`int`, *optional*, defaults to 1):
72 Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
73 document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
74 necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
75 issue](https://github.com/pytorch/pytorch/issues/76232).
76 tie_word_embeddings (`bool`, *optional*, defaults to `False`):
77 Whether to tie weight embeddings
78 rope_theta (`float`, *optional*, defaults to 10000.0):
79 The base period of the RoPE embeddings.
80 rope_scaling (`Dict`, *optional*):
81 Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
82 strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
83 `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
84 `max_position_embeddings` to the expected new maximum. See the following thread for more information on how
85 these scaling strategies behave:
86 https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This is an
87 experimental feature, subject to breaking API changes in future versions.
88 attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
89 Whether to use a bias in the query, key, value and output projection layers during self-attention.
90 attention_dropout (`float`, *optional*, defaults to 0.0):
91 The dropout ratio for the attention probabilities.
92 use_qk_norm (`bool`, *optional*, defaults to `False`):
93 Whether query and key in attention use norm
94 use_cla (`bool`, *optional*, defaults to `False`):
95 Whether to use CLA in attention
96 cla_share_factor (`int`, *optional*, defaults to 1):
97 The share factor of CLA
98 num_experts (`int` or `List`, *optional*, defaults to 1):
99 The number of experts for moe. If it is a list, it will be used as the number of experts for each layer.
100 num_shared_expert (`int` or `List`, *optional*, defaults to 1):
101 The number of shared experts for moe. If it is a list, it will be used as the number of shared experts
102 for each layer.
103 moe_topk (`int` or `List`, *optional*, defaults to 1):
104 The topk value for moe. If it is a list, it will be used as the topk value for each layer.
105 capacity_factor (Not used) (`float` or `List`, *optional*, defaults to 1.0):
106 The capacity factor for moe. If it is a list, it will be used as the capacity factor for each layer.
107 moe_layer_num_skipped (`int`, *optional*, defaults to 0):
108 First moe_layer_num_skipped layers do not use MoE.
109 """
110
111 model_type = "Hunyuan"
112 keys_to_ignore_at_inference = ["past_key_values"]
113
114 def __init__(
115 self,
116 vocab_size=290943,
117 hidden_size=4096,
118 intermediate_size: int=11008,
119 moe_intermediate_size: Union[int, List]=None,
120 num_hidden_layers=32,
121 num_attention_heads=32,
122 num_key_value_heads=None,
123 attention_head_dim=None,
124 hidden_act="silu",
125 max_position_embeddings=2048,
126 initializer_range=0.02,
127 rms_norm_eps=1e-5,
128 use_cache=True,
129 pad_token_id=0,
130 bos_token_id=1,
131 eos_token_id=2,
132 eod_token_id=3,
133 im_start_id=4,
134 im_end_id=5,
135 text_start_id=6,
136 text_end_id=7,
137 image_token_id=8,
138 video_start_id=9,
139 video_end_id=10,
140 im_newline_id=11,
141 mask_init_id=12,
142 pretraining_tp=1,
143 tie_word_embeddings=False,
144 rope_theta=10000.0,
145 rope_scaling=None,
146 attention_bias=False,
147 mlp_bias=False,
148 attention_dropout=0.0,
149 use_qk_norm=False,
150 use_rotary_pos_emb=True,
151 use_cla=False,
152 cla_share_factor=1,
153 norm_type="hf_rms",
154 num_experts: Union[int, List] = 1,
155 use_mixed_mlp_moe=False,
156 num_shared_expert: Union[int, List] = 1,
157 moe_topk: Union[int, List] = 1,
158 capacity_factor: int = 1.0,
159 moe_drop_tokens=False,
160 moe_random_routing_dropped_token=False,
161 use_mla=False,
162 kv_lora_rank=512,
163 q_lora_rank=1536,
164 qk_rope_head_dim=64,
165 v_head_dim=128,
166 qk_nope_head_dim=128,
167 moe_layer_num_skipped=0,
168 norm_topk_prob=True,
169 routed_scaling_factor=1.0,
170 group_limited_greedy=False,
171 n_group=None,
172 topk_group=None,
173 add_classification_head=False,
174 class_num=0,
175 pool_type="last",
176 pad_id=-1,
177 # Added
178 moe_impl="eager",
179 vae_downsample_factor=(16, 16), # (h, w)
180 img_proj_type="unet",
181 patch_size=1,
182 patch_embed_hidden_dim=1024,
183 image_base_size=1024,
184 vae=None,
185 vit=None,
186 vit_processor=None,
187 vit_aligner=None,
188 **kwargs,
189 ):
190 self.vocab_size = vocab_size
191 self.max_position_embeddings = max_position_embeddings
192 self.hidden_size = hidden_size
193 self.intermediate_size = intermediate_size
194 self.moe_intermediate_size = moe_intermediate_size
195 self.num_hidden_layers = num_hidden_layers
196 self.num_attention_heads = num_attention_heads
197 self.moe_impl = moe_impl
198 self.num_experts = num_experts
199 self.use_mixed_mlp_moe = use_mixed_mlp_moe
200 self.num_shared_expert = num_shared_expert
201 self.moe_topk = moe_topk
202 self.capacity_factor = capacity_factor
203 self.moe_drop_tokens = moe_drop_tokens
204 self.moe_random_routing_dropped_token = moe_random_routing_dropped_token
205
206 if attention_head_dim is not None:
207 self.attention_head_dim = attention_head_dim
208 else:
209 self.attention_head_dim = self.hidden_size // num_attention_heads
210
211 # for backward compatibility
212 if num_key_value_heads is None:
213 num_key_value_heads = num_attention_heads
214
215 self.num_key_value_heads = num_key_value_heads
216 self.hidden_act = hidden_act
217 self.initializer_range = initializer_range
218 self.rms_norm_eps = rms_norm_eps
219 self.pretraining_tp = pretraining_tp
220 self.use_cache = use_cache
221 self.rope_theta = rope_theta
222 self.rope_scaling = rope_scaling
223 self.attention_bias = attention_bias
224 self.mlp_bias = mlp_bias
225 self.attention_dropout = attention_dropout
226 self.use_qk_norm = use_qk_norm
227 self.use_rotary_pos_emb = use_rotary_pos_emb
228 self.use_cla = use_cla
229 self.cla_share_factor = cla_share_factor
230 self.norm_type = norm_type
231 # MLA args
232 self.use_mla = use_mla
233 self.kv_lora_rank = kv_lora_rank
234 self.q_lora_rank = q_lora_rank
235 self.qk_rope_head_dim = qk_rope_head_dim
236 self.qk_nope_head_dim = qk_nope_head_dim
237 self.v_head_dim = v_head_dim
238
239 # DeepSeek related args
240 self.moe_layer_num_skipped = moe_layer_num_skipped
241 self.norm_topk_prob = norm_topk_prob
242 self.routed_scaling_factor = routed_scaling_factor
243 self.group_limited_greedy = group_limited_greedy
244 self.n_group = n_group
245 self.topk_group = topk_group
246 self.add_classification_head = add_classification_head
247 self.class_num = class_num
248 self.pool_type = pool_type
249 self.pad_id = pad_id
250
251 if self.class_num is not None:
252 self.dense_list = [self.hidden_size, self.class_num]
253
254 # ViT args
255 self.vit = vit
256 self.vit_processor = vit_processor
257 self.vit_aligner = vit_aligner
258
259 # Image Gen args
260 self.vae = vae
261 self.vae_downsample_factor = vae_downsample_factor
262 self.img_proj_type = img_proj_type
263 self.patch_size = patch_size
264 self.patch_embed_hidden_dim = patch_embed_hidden_dim
265 self.image_base_size = image_base_size
266
267 # token id
268 self.eod_token_id = eod_token_id
269 self.im_start_id = im_start_id
270 self.im_end_id = im_end_id
271 self.text_start_id = text_start_id
272 self.text_end_id = text_end_id
273 self.image_token_id = image_token_id
274 self.video_start_id = video_start_id
275 self.video_end_id = video_end_id
276 self.im_newline_id = im_newline_id
277 self.mask_init_id = mask_init_id
278
279 super().__init__(
280 pad_token_id=pad_token_id,
281 bos_token_id=bos_token_id,
282 eos_token_id=eos_token_id,
283 tie_word_embeddings=tie_word_embeddings,
284 **kwargs,
285 )
286