Aluode/PerceptionLabPortable
0
1# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨2# This file was automatically generated from src/transformers/models/t5gemma/modular_t5gemma.py.3# Do NOT edit this file manually as any edits will be overwritten by the generation of4# the file from the modular. If any change should be done, please apply the change to the5# modular_t5gemma.py file directly. One of our CI enforces this.6# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨7# coding=utf-88# Copyright 2025 Google Inc. HuggingFace Inc. team. All rights reserved.9#10#11# Licensed under the Apache License, Version 2.0 (the "License");12# you may not use this file except in compliance with the License.13# You may obtain a copy of the License at14#15# http://www.apache.org/licenses/LICENSE-2.016#17# Unless required by applicable law or agreed to in writing, software18# distributed under the License is distributed on an "AS IS" BASIS,19# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.20# See the License for the specific language governing permissions and21# limitations under the License.22from typing import Any, Optional, Union23 24from ...configuration_utils import PretrainedConfig, layer_type_validation25 26 27class T5GemmaModuleConfig(PretrainedConfig):28 r"""29 This is the configuration class to store the configuration of a [`T5GemmaModuleModel`]. It is used to instantiate an T5GemmaModule30 model according to the specified arguments, defining the model architecture. Instantiating a configuration with the31 defaults will yield a similar configuration to that of the T5GemmaModule-7B.32 e.g. [google/t5_gemma_module-7b](https://huggingface.co/google/t5_gemma_module-7b)33 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the34 documentation from [`PretrainedConfig`] for more information.35 Args:36 vocab_size (`int`, *optional*, defaults to 256000):37 Vocabulary size of the T5GemmaModule model. Defines the number of different tokens that can be represented by the38 `inputs_ids` passed when calling [`T5GemmaModuleModel`]39 hidden_size (`int`, *optional*, defaults to 2304):40 Dimension of the hidden representations.41 intermediate_size (`int`, *optional*, defaults to 9216):42 Dimension of the MLP representations.43 num_hidden_layers (`int`, *optional*, defaults to 26):44 Number of hidden layers in the Transformer decoder.45 num_attention_heads (`int`, *optional*, defaults to 8):46 Number of attention heads for each attention layer in the Transformer decoder.47 num_key_value_heads (`int`, *optional*, defaults to 4):48 This is the number of key_value heads that should be used to implement Grouped Query Attention. If49 `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if50 `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When51 converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed52 by meanpooling all the original heads within that group. For more details, check out [this53 paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to54 `num_attention_heads`.55 head_dim (`int`, *optional*, defaults to 256):56 The attention head dimension.57 hidden_activation (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):58 The non-linear activation function (function or string) in the decoder. Will default to `"gelu_pytorch_tanh"`59 if not specified. `"gelu_pytorch_tanh"` uses an approximation of the `"gelu"` activation function.60 max_position_embeddings (`int`, *optional*, defaults to 8192):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). Only68 relevant if `config.is_decoder=True`.69 pad_token_id (`int`, *optional*, defaults to 0):70 Padding token id.71 eos_token_id (`int`, *optional*, defaults to 1):72 End of stream token id.73 bos_token_id (`int`, *optional*, defaults to 2):74 Beginning of stream token id.75 tie_word_embeddings (`bool`, *optional*, defaults to `True`):76 Whether to tie weight embeddings77 rope_theta (`float`, *optional*, defaults to 10000.0):78 The base period of the RoPE embeddings.79 attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):80 Whether to use a bias in the query, key, value and output projection layers during self-attention.81 attention_dropout (`float`, *optional*, defaults to 0.0):82 The dropout ratio for the attention probabilities.83 query_pre_attn_scalar (`float`, *optional*, defaults to 256):84 scaling factor used on the attention scores85 sliding_window (`int`, *optional*, defaults to 4096):86 in T5GemmaModule, every other layer uses sliding window attention. This is the size of the sliding window.87 layer_types (`list`, *optional*):88 Attention pattern for each layer.89 final_logit_softcapping (`float`, *optional*, defaults to 30.0):90 scaling factor when applying tanh softcapping on the logits.91 attn_logit_softcapping (`float`, *optional*, defaults to 50.0):92 scaling factor when applying tanh softcapping on the attention scores.93 94 ```python95 >>> from transformers import T5GemmaModuleModel, T5GemmaModuleConfig96 >>> # Initializing a T5GemmaModule t5_gemma_module-7b style configuration97 >>> configuration = T5GemmaModuleConfig()98 >>> # Initializing a model from the t5_gemma_module-7b style configuration99 >>> model = T5GemmaModuleModel(configuration)100 >>> # Accessing the model configuration101 >>> configuration = model.config102 ```"""103 104 model_type = "t5_gemma_module"105 keys_to_ignore_at_inference = ["past_key_values"]106 base_model_tp_plan = {107 "layers.*.self_attn.q_proj": "colwise",108 "layers.*.self_attn.k_proj": "colwise",109 "layers.*.self_attn.v_proj": "colwise",110 "layers.*.self_attn.o_proj": "rowwise",111 "layers.*.mlp.gate_proj": "colwise",112 "layers.*.mlp.up_proj": "colwise",113 "layers.*.mlp.down_proj": "rowwise",114 }115 base_model_pp_plan = {116 "embed_tokens": (["input_ids"], ["inputs_embeds"]),117 "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),118 "norm": (["hidden_states"], ["hidden_states"]),119 }120 121 def __init__(122 self,123 vocab_size=256000,124 hidden_size=2304,125 intermediate_size=9216,126 num_hidden_layers=26,127 num_attention_heads=8,128 num_key_value_heads=4,129 head_dim=256,130 hidden_activation="gelu_pytorch_tanh",131 max_position_embeddings=8192,132 initializer_range=0.02,133 rms_norm_eps=1e-6,134 use_cache=True,135 pad_token_id=0,136 eos_token_id=1,137 bos_token_id=2,138 tie_word_embeddings=True,139 rope_theta=10000.0,140 attention_bias=False,141 attention_dropout=0.0,142 query_pre_attn_scalar=256,143 sliding_window=4096,144 layer_types=None,145 final_logit_softcapping=30.0,146 attn_logit_softcapping=50.0,147 **kwargs,148 ):149 super().__init__(150 pad_token_id=pad_token_id,151 bos_token_id=bos_token_id,152 eos_token_id=eos_token_id,153 tie_word_embeddings=tie_word_embeddings,154 **kwargs,155 )156 self.vocab_size = vocab_size157 self.max_position_embeddings = max_position_embeddings158 self.hidden_size = hidden_size159 self.intermediate_size = intermediate_size160 self.num_hidden_layers = num_hidden_layers161 self.num_attention_heads = num_attention_heads162 self.head_dim = head_dim163 self.num_key_value_heads = num_key_value_heads164 self.initializer_range = initializer_range165 self.rms_norm_eps = rms_norm_eps166 self.use_cache = use_cache167 self.rope_theta = rope_theta168 self.attention_bias = attention_bias169 self.attention_dropout = attention_dropout170 self.hidden_activation = hidden_activation171 self.query_pre_attn_scalar = query_pre_attn_scalar172 self.sliding_window = sliding_window173 self.final_logit_softcapping = final_logit_softcapping174 self.attn_logit_softcapping = attn_logit_softcapping175 self.layer_types = layer_types176 177 if self.layer_types is None:178 self.layer_types = [179 "sliding_attention" if bool((i + 1) % 2) else "full_attention" for i in range(self.num_hidden_layers)180 ]181 layer_type_validation(self.layer_types, self.num_hidden_layers)182 183 184class T5GemmaConfig(PretrainedConfig):185 r"""186 This is the configuration class to store the configuration of a [`T5GemmaModel`]. It is used to instantiate an T5Gemma187 model according to the specified arguments, defining the model architecture. Instantiating a configuration with the188 defaults will yield a similar configuration to a hypothetical balanced Gemma2 encoder-decoder model.189 e.g. [google/t5gemma-2b-2b-prefixlm-it](https://huggingface.co/google/t5gemma-2b-2b-prefixlm-it)190 ```python191 >>> from transformers import T5GemmaConfig, T5GemmaModel192 >>> t5gemma_config = T5GemmaConfig.from_pretrained("google/t5gemma-2b-2b-prefixlm-it")193 >>> model = T5GemmaModel(t5gemma_config)194 ```195 Configuration objects inherit from [PretrainedConfig] and can be used to control the model outputs. Read the196 documentation from [PretrainedConfig] for more information.197 Args:198 encoder (`Union[T5GemmaModuleConfig, dict]`, optional, *optional*):199 Configuration for the encoder.200 decoder (`Union[T5GemmaModuleConfig, dict]`, optional, *optional*):201 Configuration for the decoder.202 is_encoder_decoder (bool, optional, *optional*, defaults to `True`):203 Whether the model is used as an encoder/decoder or not.204 dropout_rate (`float`, *optional*, defaults to 0.0):205 The ratio for all dropout layers (following T5).206 classifier_dropout_rate (`float`, *optional*, defaults to 0.0):207 The dropout ratio for classifier (following T5).208 attention_dropout (`float`, *optional*, defaults to 0.0):209 The dropout ratio for attention.210 tie_word_embeddings (`bool`, *optional*, defaults to `True`):211 Whether tie input and output embeddings.212 vocab_size (`int`, *optional*, defaults to 256000):213 Vocabulary size of the T5Gemma model (the same as Gemma 2).214 kwargs (additional keyword arguments, optional, *optional*):215 Will be passed to the PretrainedConfig base class.216 """217 218 model_type = "t5gemma"219 keys_to_ignore_at_inference = ["past_key_values"]220 base_model_tp_plan = {221 # encoder222 "encoder.layers.*.self_attn.q_proj": "colwise",223 "encoder.layers.*.self_attn.k_proj": "colwise",224 "encoder.layers.*.self_attn.v_proj": "colwise",225 "encoder.layers.*.self_attn.o_proj": "rowwise",226 "encoder.layers.*.mlp.gate_proj": "colwise",227 "encoder.layers.*.mlp.up_proj": "colwise",228 "encoder.layers.*.mlp.down_proj": "rowwise",229 # decoder230 "decoder.layers.*.self_attn.q_proj": "colwise",231 "decoder.layers.*.self_attn.k_proj": "colwise",232 "decoder.layers.*.self_attn.v_proj": "colwise",233 "decoder.layers.*.self_attn.o_proj": "rowwise",234 "decoder.layers.*.cross_attn.q_proj": "colwise",235 "decoder.layers.*.cross_attn.k_proj": "colwise",236 "decoder.layers.*.cross_attn.v_proj": "colwise",237 "decoder.layers.*.cross_attn.o_proj": "rowwise",238 "decoder.layers.*.mlp.gate_proj": "colwise",239 "decoder.layers.*.mlp.up_proj": "colwise",240 "decoder.layers.*.mlp.down_proj": "rowwise",241 }242 base_model_pp_plan = {243 # encoder244 "encoder.embed_tokens": (["input_ids"], ["inputs_embeds"]),245 "encoder.layers": (["hidden_states", "attention_mask"], ["hidden_states"]),246 "encoder.norm": (["hidden_states"], ["hidden_states"]),247 # decoder248 "decoder.embed_tokens": (["input_ids"], ["inputs_embeds"]),249 "decoder.layers": (["hidden_states", "attention_mask"], ["hidden_states"]),250 "decoder.norm": (["hidden_states"], ["hidden_states"]),251 }252 253 def __init__(254 self,255 encoder: Optional[Union[T5GemmaModuleConfig, dict[Any, Any]]] = None,256 decoder: Optional[Union[T5GemmaModuleConfig, dict[Any, Any]]] = None,257 is_encoder_decoder: bool = True,258 dropout_rate: float = 0.0,259 classifier_dropout_rate: float = 0.0,260 attention_dropout: float = 0.0,261 tie_word_embeddings: bool = True,262 vocab_size: int = 256000,263 **kwargs,264 ):265 if isinstance(encoder, dict):266 encoder = T5GemmaModuleConfig(**encoder)267 elif encoder is None:268 encoder = T5GemmaModuleConfig()269 else:270 assert isinstance(encoder, T5GemmaModuleConfig), f"{type(encoder)} is not supported."271 272 if isinstance(decoder, dict):273 decoder = T5GemmaModuleConfig(**decoder)274 elif decoder is None:275 decoder = encoder276 else:277 assert isinstance(decoder, T5GemmaModuleConfig), f"{type(decoder)} is not supported."278 279 encoder = T5GemmaModuleConfig(**encoder.to_dict())280 decoder = T5GemmaModuleConfig(**decoder.to_dict())281 282 encoder.is_decoder = False283 encoder.dropout_rate = dropout_rate284 encoder.attention_dropout = attention_dropout285 self.encoder = encoder286 287 decoder.is_decoder = True288 decoder.use_cache = True289 decoder.dropout_rate = dropout_rate290 decoder.attention_dropout = attention_dropout291 decoder.cross_attention_hidden_size = encoder.hidden_size292 self.decoder = decoder293 294 for special_token_key in ["bos_token_id", "pad_token_id", "eos_token_id"]:295 if special_token_key not in kwargs:296 kwargs[special_token_key] = getattr(decoder, special_token_key)297 298 super().__init__(**kwargs)299 300 self.is_encoder_decoder = is_encoder_decoder301 self.use_cache = kwargs.get("use_cache", decoder.use_cache)302 self.initializer_range = kwargs.get("initializer_range", decoder.initializer_range)303 self.dropout_rate = dropout_rate304 self.attention_dropout = attention_dropout305 self.classifier_dropout_rate = classifier_dropout_rate306 self.tie_word_embeddings = tie_word_embeddings307 308 # Used in pipeline generation.309 self.vocab_size = vocab_size310 311 def __setattr__(self, key, value):312 shared_attr_with_submodules = [313 "output_hidden_states",314 "output_attentions",315 "_attn_implementation",316 "dropout_rate",317 "attention_dropout",318 "vocab_size",319 ]320 321 if key in shared_attr_with_submodules:322 setattr(self.encoder, key, value)323 setattr(self.decoder, key, value)324 super().__setattr__(key, value)325 326 327__all__ = ["T5GemmaConfig", "T5GemmaModuleConfig"]328