|
| 1 | +# Copyright 2025 Bytedance-Seed Ltd and the HuggingFace Inc. team. All rights reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | +"""SeedOss model configuration""" |
| 15 | + |
| 16 | +from transformers.configuration_utils import PretrainedConfig |
| 17 | +from transformers.modeling_rope_utils import rope_config_validation |
| 18 | + |
| 19 | + |
| 20 | +class SeedOssConfig(PretrainedConfig): |
| 21 | + r""" |
| 22 | + This is the configuration class to store the configuration of a [`SeedOssModel`]. It is used to instantiate an SeedOss |
| 23 | + model according to the specified arguments, defining the model architecture. Instantiating a configuration with the |
| 24 | + defaults will yield a similar configuration to that of the SeedOss-36B. |
| 25 | + e.g. [ByteDance-Seed/SeedOss-36B](https://huggingface.co/ByteDance-Seed/SeedOss-36B) |
| 26 | +
|
| 27 | + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the |
| 28 | + documentation from [`PretrainedConfig`] for more information. |
| 29 | +
|
| 30 | +
|
| 31 | + Args: |
| 32 | + vocab_size (`int`, *optional*, defaults to 155136): |
| 33 | + Vocabulary size of the SeedOss model. Defines the number of different tokens that can be represented by the |
| 34 | + `inputs_ids` passed when calling [`SeedOssModel`] |
| 35 | + hidden_size (`int`, *optional*, defaults to 4096): |
| 36 | + Dimension of the hidden representations. |
| 37 | + intermediate_size (`int`, *optional*, defaults to 27648): |
| 38 | + Dimension of the MLP representations. |
| 39 | + num_hidden_layers (`int`, *optional*, defaults to 64): |
| 40 | + Number of hidden layers in the Transformer decoder. |
| 41 | + num_attention_heads (`int`, *optional*, defaults to 80): |
| 42 | + Number of attention heads for each attention layer in the Transformer decoder. |
| 43 | + num_key_value_heads (`int`, *optional*, defaults to 8): |
| 44 | + This is the number of key_value heads that should be used to implement Grouped Query Attention. If |
| 45 | + `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if |
| 46 | + `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When |
| 47 | + converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed |
| 48 | + by meanpooling all the original heads within that group. For more details, check out [this |
| 49 | + paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `8`. |
| 50 | + hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): |
| 51 | + The non-linear activation function (function or string) in the decoder. |
| 52 | + max_position_embeddings (`int`, *optional*, defaults to 524288): |
| 53 | + The maximum sequence length that this model might ever be used with. |
| 54 | + initializer_range (`float`, *optional*, defaults to 0.02): |
| 55 | + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. |
| 56 | + rms_norm_eps (`float`, *optional*, defaults to 1e-06): |
| 57 | + The epsilon used by the rms normalization layers. |
| 58 | + use_cache (`bool`, *optional*, defaults to `True`): |
| 59 | + Whether or not the model should return the last key/values attentions (not used by all models). Only |
| 60 | + relevant if `config.is_decoder=True`. |
| 61 | + pad_token_id (`int`, *optional*, defaults to 1): |
| 62 | + Padding token id. |
| 63 | + bos_token_id (`int`, *optional*, defaults to 0): |
| 64 | + Beginning of stream token id. |
| 65 | + eos_token_id (`int`, *optional*, defaults to 2): |
| 66 | + End of stream token id. |
| 67 | + pretraining_tp (`int`, *optional*, defaults to 1): |
| 68 | + Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this |
| 69 | + document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to |
| 70 | + understand more about it. This value is necessary to ensure exact reproducibility of the pretraining |
| 71 | + results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232). |
| 72 | + tie_word_embeddings (`bool`, *optional*, defaults to `False`): |
| 73 | + Whether to tie weight embeddings |
| 74 | + rope_theta (`float`, *optional*, defaults to 10000.0): |
| 75 | + The base period of the RoPE embeddings. |
| 76 | + rope_scaling (`Dict`, *optional*): |
| 77 | + Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type |
| 78 | + and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value |
| 79 | + accordingly. |
| 80 | + Expected contents: |
| 81 | + `rope_type` (`str`): |
| 82 | + The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope', |
| 83 | + 'llama3'], with 'default' being the original RoPE implementation. |
| 84 | + `factor` (`float`, *optional*): |
| 85 | + Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In |
| 86 | + most scaling types, a `factor` of x will enable the model to handle sequences of length x * |
| 87 | + original maximum pre-trained length. |
| 88 | + `original_max_position_embeddings` (`int`, *optional*): |
| 89 | + Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during |
| 90 | + pretraining. |
| 91 | + `attention_factor` (`float`, *optional*): |
| 92 | + Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention |
| 93 | + computation. If unspecified, it defaults to value recommended by the implementation, using the |
| 94 | + `factor` field to infer the suggested value. |
| 95 | + `beta_fast` (`float`, *optional*): |
| 96 | + Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear |
| 97 | + ramp function. If unspecified, it defaults to 32. |
| 98 | + `beta_slow` (`float`, *optional*): |
| 99 | + Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear |
| 100 | + ramp function. If unspecified, it defaults to 1. |
| 101 | + `short_factor` (`list[float]`, *optional*): |
| 102 | + Only used with 'longrope'. The scaling factor to be applied to short contexts (< |
| 103 | + `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden |
| 104 | + size divided by the number of attention heads divided by 2 |
| 105 | + `long_factor` (`list[float]`, *optional*): |
| 106 | + Only used with 'longrope'. The scaling factor to be applied to long contexts (< |
| 107 | + `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden |
| 108 | + size divided by the number of attention heads divided by 2 |
| 109 | + `low_freq_factor` (`float`, *optional*): |
| 110 | + Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE |
| 111 | + `high_freq_factor` (`float`, *optional*): |
| 112 | + Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE |
| 113 | + attention_bias (`bool`, *optional*, defaults to `True`): |
| 114 | + Whether to use a bias in the query, key, value layers during self-attention. |
| 115 | + attention_out_bias (`bool`, *optional*, defaults to `False`): |
| 116 | + Whether to use a bias in the output projection layer during self-attention. |
| 117 | + attention_dropout (`float`, *optional*, defaults to 0.1): |
| 118 | + The dropout ratio for the attention probabilities. |
| 119 | + residual_dropout (`float`, *optional*, defaults to 0.1): |
| 120 | + Residual connection dropout value. |
| 121 | + mlp_bias (`bool`, *optional*, defaults to `False`): |
| 122 | + Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers. |
| 123 | + head_dim (`int`, *optional*, defaults to 128): |
| 124 | + The attention head dimension. |
| 125 | +
|
| 126 | + ```python |
| 127 | + >>> from transformers import SeedOssModel, SeedOssConfig |
| 128 | +
|
| 129 | + >>> # Initializing a SeedOss-36b style configuration |
| 130 | + >>> configuration = SeedOssConfig() |
| 131 | +
|
| 132 | + >>> # Initializing a model from the SeedOss-36b style configuration |
| 133 | + >>> model = SeedOssModel(configuration) |
| 134 | +
|
| 135 | + >>> # Accessing the model configuration |
| 136 | + >>> configuration = model.config |
| 137 | + ```""" |
| 138 | + |
| 139 | + model_type = "seed_oss" |
| 140 | + keys_to_ignore_at_inference = ["past_key_values"] |
| 141 | + # Default tensor parallel plan for base model `SeedOssModel` |
| 142 | + base_model_tp_plan = { |
| 143 | + "layers.*.self_attn.q_proj": "colwise", |
| 144 | + "layers.*.self_attn.k_proj": "colwise", |
| 145 | + "layers.*.self_attn.v_proj": "colwise", |
| 146 | + "layers.*.self_attn.o_proj": "rowwise", |
| 147 | + "layers.*.mlp.gate_proj": "colwise", |
| 148 | + "layers.*.mlp.up_proj": "colwise", |
| 149 | + "layers.*.mlp.down_proj": "rowwise", |
| 150 | + } |
| 151 | + base_model_pp_plan = { |
| 152 | + "embed_tokens": (["input_ids"], ["inputs_embeds"]), |
| 153 | + "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), |
| 154 | + "norm": (["hidden_states"], ["hidden_states"]), |
| 155 | + } |
| 156 | + |
| 157 | + def __init__( |
| 158 | + self, |
| 159 | + vocab_size=155136, |
| 160 | + hidden_size=4096, |
| 161 | + intermediate_size=27648, |
| 162 | + num_hidden_layers=64, |
| 163 | + num_attention_heads=80, |
| 164 | + num_key_value_heads=8, |
| 165 | + hidden_act="silu", |
| 166 | + max_position_embeddings=524288, |
| 167 | + initializer_range=0.02, |
| 168 | + rms_norm_eps=1e-6, |
| 169 | + use_cache=True, |
| 170 | + pad_token_id=1, |
| 171 | + bos_token_id=0, |
| 172 | + eos_token_id=2, |
| 173 | + pretraining_tp=1, |
| 174 | + tie_word_embeddings=False, |
| 175 | + rope_theta=10000.0, |
| 176 | + rope_scaling=None, |
| 177 | + attention_bias=True, |
| 178 | + attention_out_bias=False, |
| 179 | + attention_dropout=0.1, |
| 180 | + residual_dropout=0.1, |
| 181 | + mlp_bias=False, |
| 182 | + head_dim=128, |
| 183 | + **kwargs, |
| 184 | + ): |
| 185 | + self.vocab_size = vocab_size |
| 186 | + self.max_position_embeddings = max_position_embeddings |
| 187 | + self.hidden_size = hidden_size |
| 188 | + self.intermediate_size = intermediate_size |
| 189 | + self.num_hidden_layers = num_hidden_layers |
| 190 | + self.num_attention_heads = num_attention_heads |
| 191 | + # for backward compatibility |
| 192 | + if num_key_value_heads is None: |
| 193 | + num_key_value_heads = num_attention_heads |
| 194 | + |
| 195 | + self.num_key_value_heads = num_key_value_heads |
| 196 | + self.hidden_act = hidden_act |
| 197 | + self.initializer_range = initializer_range |
| 198 | + self.rms_norm_eps = rms_norm_eps |
| 199 | + self.pretraining_tp = pretraining_tp |
| 200 | + self.use_cache = use_cache |
| 201 | + self.rope_theta = rope_theta |
| 202 | + self.rope_scaling = rope_scaling |
| 203 | + self.attention_bias = attention_bias |
| 204 | + self.attention_out_bias = attention_out_bias |
| 205 | + self.attention_dropout = attention_dropout |
| 206 | + self.residual_dropout = residual_dropout |
| 207 | + self.mlp_bias = mlp_bias |
| 208 | + self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads |
| 209 | + # Validate the correctness of rotary position embeddings parameters |
| 210 | + # BC: if there is a 'type' field, copy it it to 'rope_type'. |
| 211 | + if self.rope_scaling is not None and "type" in self.rope_scaling: |
| 212 | + self.rope_scaling["rope_type"] = self.rope_scaling["type"] |
| 213 | + rope_config_validation(self) |
| 214 | + |
| 215 | + super().__init__( |
| 216 | + pad_token_id=pad_token_id, |
| 217 | + bos_token_id=bos_token_id, |
| 218 | + eos_token_id=eos_token_id, |
| 219 | + tie_word_embeddings=tie_word_embeddings, |
| 220 | + **kwargs, |
| 221 | + ) |
| 222 | + |
| 223 | + |
| 224 | +__all__ = ["SeedOssConfig"] |
0 commit comments