# Copyright 2025 Google Inc. HuggingFace Inc. team. All rights reserved. # # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy from collections.abc import Callable from typing import Any, Optional import torch import torch.nn as nn from ... import initialization as init from ...cache_utils import DynamicCache, EncoderDecoderCache, StaticCache from ...configuration_utils import PreTrainedConfig, layer_type_validation from ...generation import GenerationConfig, GenerationMixin, GenerationMode from ...masking_utils import create_bidirectional_mask from ...modeling_flash_attention_utils import FlashAttentionKwargs from ...modeling_outputs import ( BaseModelOutput, BaseModelOutputWithPastAndCrossAttentions, BaseModelOutputWithPooling, Seq2SeqLMOutput, Seq2SeqModelOutput, SequenceClassifierOutput, TokenClassifierOutput, ) from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, RopeParameters from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from ...processing_utils import Unpack from ...utils import ( TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_compilable_check, ) from ...utils.generic import merge_with_config_defaults from ...utils.output_capturing import OutputRecorder, capture_outputs from ..auto import AutoModel from ..gemma3.configuration_gemma3 import Gemma3Config, Gemma3TextConfig from ..gemma3.modeling_gemma3 import ( Gemma3Attention, Gemma3MLP, Gemma3MultiModalProjector, Gemma3PreTrainedModel, Gemma3RMSNorm, Gemma3RotaryEmbedding, Gemma3TextScaledWordEmbedding, apply_rotary_pos_emb, create_causal_mask, create_sliding_window_causal_mask, eager_attention_forward, ) from ..siglip import SiglipVisionConfig from ..t5gemma.modeling_t5gemma import ( T5GemmaClassificationHead, T5GemmaEncoderLayer, T5GemmaLMHead, bidirectional_mask_function, ) logger = logging.get_logger(__name__) class T5Gemma2TextConfig(Gemma3TextConfig, PreTrainedConfig): r""" This is the configuration class to store the configuration of a [`T5Gemma2TextModel`]. It is used to instantiate the encoder's text model portion of the T5Gemma2 Model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the T5Gemma2Text-7B. e.g. [google/t5gemma2_text-7b](https://huggingface.co/google/t5gemma2_text-7b) Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PreTrainedConfig`] for more information. Args: vocab_size (`int`, *optional*, defaults to 262208): Vocabulary size of the T5Gemma2Text model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`T5Gemma2TextModel`] hidden_size (`int`, *optional*, defaults to 2304): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 9216): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 26): Number of hidden layers in the Transformer decoder. num_attention_heads (`int`, *optional*, defaults to 8): Number of attention heads for each attention layer in the Transformer decoder. num_key_value_heads (`int`, *optional*, defaults to 4): This is the number of key_value heads that should be used to implement Grouped Query Attention. If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details, check out [this paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `num_attention_heads`. head_dim (`int`, *optional*, defaults to 256): The attention head dimension. hidden_activation (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`): The non-linear activation function (function or string) in the decoder. Will default to `"gelu_pytorch_tanh"` if not specified. `"gelu_pytorch_tanh"` uses an approximation of the `"gelu"` activation function. max_position_embeddings (`int`, *optional*, defaults to 131072): The maximum sequence length that this model might ever be used with. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. rms_norm_eps (`float`, *optional*, defaults to 1e-06): The epsilon used by the rms normalization layers. use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if `config.is_decoder=True`. pad_token_id (`int`, *optional*, defaults to 0): Padding token id. eos_token_id (`int`, *optional*, defaults to 1): End of stream token id. bos_token_id (`int`, *optional*, defaults to 2): Beginning of stream token id. attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during self-attention. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. query_pre_attn_scalar (`float`, *optional*, defaults to 256): Scaling factor used on the attention scores sliding_window (`int`, *optional*, defaults to 4096): In T5Gemma2Text, every other layer uses sliding window attention. This is the size of the sliding window. layer_types (`list`, *optional*): Attention pattern for each layer. final_logit_softcapping (`float`, *optional*): Scaling factor when applying tanh softcapping on the logits. attn_logit_softcapping (`float`, *optional*): Scaling factor when applying tanh softcapping on the attention scores. rope_parameters (`RopeParameters`, *optional*): Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE with longer `max_position_embeddings`. """ model_type = "t5gemma2_text" def __init__( self, vocab_size: int | None = 262_208, hidden_size: int | None = 2304, intermediate_size: int | None = 9216, num_hidden_layers: int | None = 26, num_attention_heads: int | None = 8, num_key_value_heads: int | None = 4, head_dim: int | None = 256, hidden_activation: str | None = "gelu_pytorch_tanh", max_position_embeddings: int | None = 131_072, initializer_range: float | None = 0.02, rms_norm_eps: int | None = 1e-6, use_cache: bool | None = True, pad_token_id: int | None = 0, eos_token_id: int | None = 1, bos_token_id: int | None = 2, attention_bias: bool | None = False, attention_dropout: float | None = 0.0, query_pre_attn_scalar: int | None = 256, sliding_window: int | None = 4096, layer_types: list[str] | None = None, final_logit_softcapping: float | None = None, attn_logit_softcapping: float | None = None, rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None, **kwargs, ): self.pad_token_id = pad_token_id self.bos_token_id = bos_token_id self.eos_token_id = eos_token_id self.vocab_size = vocab_size self.max_position_embeddings = max_position_embeddings self.hidden_size = hidden_size self.intermediate_size = intermediate_size self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads self.head_dim = head_dim self.num_key_value_heads = num_key_value_heads self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.attention_bias = attention_bias self.attention_dropout = attention_dropout self.hidden_activation = hidden_activation self.query_pre_attn_scalar = query_pre_attn_scalar self.sliding_window = sliding_window self.final_logit_softcapping = final_logit_softcapping self.attn_logit_softcapping = attn_logit_softcapping self.layer_types = layer_types # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub self._sliding_window_pattern = kwargs.get("sliding_window_pattern", 6) if self.layer_types is None: self.layer_types = [ "sliding_attention" if bool((i + 1) % self._sliding_window_pattern) else "full_attention" for i in range(self.num_hidden_layers) ] layer_type_validation(self.layer_types, self.num_hidden_layers) self.rope_parameters = rope_parameters PreTrainedConfig.__init__(**kwargs) class T5Gemma2EncoderConfig(Gemma3Config): model_type = "t5gemma2_encoder" sub_configs = { "text_config": T5Gemma2TextConfig, "vision_config": SiglipVisionConfig, } class T5Gemma2DecoderConfig(Gemma3TextConfig, PreTrainedConfig): r""" This is the configuration class to store the configuration of a [`T5Gemma2DecoderModel`]. It is used to instantiate the decoder text model portion of the T5Gemma2 Model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the T5Gemma2Decoder-7B. e.g. [google/t5gemma2_text-7b](https://huggingface.co/google/t5gemma2_text-7b) Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PreTrainedConfig`] for more information. Args: vocab_size (`int`, *optional*, defaults to 262208): Vocabulary size of the T5Gemma2Decoder model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`T5Gemma2DecoderModel`] hidden_size (`int`, *optional*, defaults to 2304): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 9216): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 26): Number of hidden layers in the Transformer decoder. num_attention_heads (`int`, *optional*, defaults to 8): Number of attention heads for each attention layer in the Transformer decoder. num_key_value_heads (`int`, *optional*, defaults to 4): This is the number of key_value heads that should be used to implement Grouped Query Attention. If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details, check out [this paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `num_attention_heads`. head_dim (`int`, *optional*, defaults to 256): The attention head dimension. hidden_activation (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`): The non-linear activation function (function or string) in the decoder. Will default to `"gelu_pytorch_tanh"` if not specified. `"gelu_pytorch_tanh"` uses an approximation of the `"gelu"` activation function. max_position_embeddings (`int`, *optional*, defaults to 131072): The maximum sequence length that this model might ever be used with. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. rms_norm_eps (`float`, *optional*, defaults to 1e-06): The epsilon used by the rms normalization layers. use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if `config.is_decoder=True`. pad_token_id (`int`, *optional*, defaults to 0): Padding token id. eos_token_id (`int`, *optional*, defaults to 1): End of stream token id. bos_token_id (`int`, *optional*, defaults to 2): Beginning of stream token id. attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during self-attention. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. query_pre_attn_scalar (`float`, *optional*, defaults to 256): Scaling factor used on the attention scores sliding_window (`int`, *optional*, defaults to 4096): In T5Gemma2Decoder, every other layer uses sliding window attention. This is the size of the sliding window. layer_types (`list`, *optional*): Attention pattern for each layer. final_logit_softcapping (`float`, *optional*): Scaling factor when applying tanh softcapping on the logits. attn_logit_softcapping (`float`, *optional*): Scaling factor when applying tanh softcapping on the attention scores. rope_parameters (`RopeParameters`, *optional*): Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE with longer `max_position_embeddings`. """ model_type = "t5gemma2_decoder" def __init__( self, vocab_size: int | None = 262_208, hidden_size: int | None = 2304, intermediate_size: int | None = 9216, num_hidden_layers: int | None = 26, num_attention_heads: int | None = 8, num_key_value_heads: int | None = 4, head_dim: int | None = 256, hidden_activation: str | None = "gelu_pytorch_tanh", max_position_embeddings: int | None = 131_072, initializer_range: float | None = 0.02, rms_norm_eps: int | None = 1e-6, use_cache: bool | None = True, pad_token_id: int | None = 0, eos_token_id: int | None = 1, bos_token_id: int | None = 2, attention_bias: bool | None = False, attention_dropout: float | None = 0.0, query_pre_attn_scalar: int | None = 256, sliding_window: int | None = 4096, layer_types: list[str] | None = None, final_logit_softcapping: float | None = None, attn_logit_softcapping: float | None = None, rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None, **kwargs, ): self.pad_token_id = pad_token_id self.bos_token_id = bos_token_id self.eos_token_id = eos_token_id self.vocab_size = vocab_size self.max_position_embeddings = max_position_embeddings self.hidden_size = hidden_size self.intermediate_size = intermediate_size self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads self.head_dim = head_dim self.num_key_value_heads = num_key_value_heads self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.attention_bias = attention_bias self.attention_dropout = attention_dropout self.hidden_activation = hidden_activation self.query_pre_attn_scalar = query_pre_attn_scalar self.sliding_window = sliding_window self.final_logit_softcapping = final_logit_softcapping self.attn_logit_softcapping = attn_logit_softcapping self.layer_types = layer_types # BC -> the pattern used to be a simple int, and it's still present in configs on the Hub self._sliding_window_pattern = kwargs.get("sliding_window_pattern", 6) if self.layer_types is None: self.layer_types = [ "sliding_attention" if bool((i + 1) % self._sliding_window_pattern) else "full_attention" for i in range(self.num_hidden_layers) ] layer_type_validation(self.layer_types, self.num_hidden_layers) self.rope_parameters = rope_parameters PreTrainedConfig.__init__(**kwargs) class T5Gemma2Config(PreTrainedConfig): r""" This is the configuration class to store the configuration of a [`T5Gemma2Model`]. It is used to instantiate an T5Gemma2 model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to a hypothetical balanced Gemma3 encoder-decoder model. e.g. [google/t5gemma-2-270m-270m](https://huggingface.co/google/t5gemma-2-270m-270m) Configuration objects inherit from [PreTrainedConfig] and can be used to control the model outputs. Read the documentation from [PreTrainedConfig] for more information. Args: encoder (`Union[T5Gemma2EncoderConfig, dict]`, optional, *optional*): Configuration for the encoder. decoder (`Union[T5Gemma2DecoderConfig, dict]`, optional, *optional*): Configuration for the decoder. is_encoder_decoder (bool, optional, *optional*, defaults to `True`): Whether the model is used as an encoder/decoder or not. dropout_rate (`float`, *optional*, defaults to 0.0): The ratio for all dropout layers (following T5). attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for attention. classifier_dropout_rate (`float`, *optional*, defaults to 0.0): The dropout ratio for classifier (following T5). initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. image_token_index (`int`, *optional*, defaults to 256001): The image token index to encode the image prompt. Defaults to 256001, which is right after the eoi_token_index. Note this is different from Gemma 3. tie_word_embeddings (`bool`, *optional*, defaults to `True`): Whether to tie weight embeddings ```python >>> from transformers import T5Gemma2Config, T5Gemma2Model >>> t5gemma2_config = T5Gemma2Config.from_pretrained("google/t5gemma-270m-270m") >>> model = T5Gemma2Model(t5gemma2_config) ``` """ model_type = "t5gemma2" keys_to_ignore_at_inference = ["past_key_values"] sub_configs = { "encoder": T5Gemma2EncoderConfig, "decoder": T5Gemma2DecoderConfig, } attribute_map = { "image_token_id": "image_token_index", "eoi_token_id": "eoi_token_index", } def __init__( self, encoder: T5Gemma2EncoderConfig | dict[str, Any] | None = None, decoder: T5Gemma2DecoderConfig | dict[str, Any] | None = None, is_encoder_decoder: bool = True, dropout_rate: float = 0.0, attention_dropout: float = 0.0, classifier_dropout_rate: float = 0.0, initializer_range: float = 0.02, image_token_index: int = 256_001, tie_word_embeddings: bool | None = True, **kwargs, ): if isinstance(encoder, dict): encoder = T5Gemma2EncoderConfig(**encoder) elif encoder is None: encoder = T5Gemma2EncoderConfig() logger.info("encoder is None, using default T5Gemma2EncoderConfig encoder config.") else: if not isinstance(encoder, T5Gemma2EncoderConfig): raise ValueError(f"{type(encoder)} is not supported.") if isinstance(decoder, dict): decoder = T5Gemma2DecoderConfig(**decoder) elif decoder is None: decoder = T5Gemma2DecoderConfig() logger.info("decoder is None, using default T5Gemma2DecoderConfig decoder config.") else: if not isinstance(decoder, T5Gemma2DecoderConfig): raise ValueError(f"{type(decoder)} is not supported.") if encoder.text_config.hidden_size != decoder.hidden_size: raise ValueError( "Imbalanced encoder-decoder is not supported in T5Gemma2: " f"encoder ({encoder.text_config.hidden_size}) vs decoder ({decoder.hidden_size})." ) if not is_encoder_decoder: raise ValueError("T5Gemma2Model only support encoder-decoder modeling.") if encoder.text_config.vocab_size != decoder.vocab_size: raise ValueError( "Imbalanced encoder-decoder vocabulary size is not supported in T5Gemma2: " f"encoder ({encoder.text_config.vocab_size}) vs decoder ({decoder.vocab_size})." ) # Encoder. encoder.text_config.dropout_rate = dropout_rate encoder.text_config.attention_dropout = attention_dropout encoder.vision_config.attention_dropout = attention_dropout encoder.image_token_index = image_token_index self.encoder = encoder # Decoder. decoder.dropout_rate = dropout_rate decoder.attention_dropout = attention_dropout self.decoder = decoder for special_token_key in ["bos_token_id", "pad_token_id", "eos_token_id", "vocab_size"]: if special_token_key not in kwargs: kwargs[special_token_key] = getattr(decoder, special_token_key) self.classifier_dropout_rate = classifier_dropout_rate self.initializer_range = initializer_range self.eoi_token_index = encoder.eoi_token_index self.image_token_index = image_token_index self.tie_word_embeddings = tie_word_embeddings super().__init__(is_encoder_decoder=is_encoder_decoder, **kwargs) class T5Gemma2RMSNorm(Gemma3RMSNorm): pass class T5Gemma2MLP(Gemma3MLP): def __init__(self, config: T5Gemma2TextConfig): super().__init__(config) self.dropout = nn.Dropout(config.dropout_rate) def forward(self, x): hidden_states = self.act_fn(self.gate_proj(x)) * self.up_proj(x) hidden_states = self.dropout(hidden_states) down_proj = self.down_proj(hidden_states) return down_proj class T5Gemma2RotaryEmbedding(Gemma3RotaryEmbedding): def __init__(self, config: T5Gemma2TextConfig, device=None): super().__init__(config, device) @staticmethod def compute_default_rope_parameters( config: T5Gemma2TextConfig | None = None, device: Optional["torch.device"] = None, seq_len: int | None = None, layer_type: str | None = None, ) -> tuple["torch.Tensor", float]: return super().compute_default_rope_parameters(config, device, seq_len, layer_type) class T5Gemma2SelfAttention(Gemma3Attention): def __init__(self, config: T5Gemma2TextConfig, layer_idx: int): super().__init__(config, layer_idx) self.is_causal = False # Only used by the encoder class T5Gemma2MergedAttention(Gemma3Attention): """Merged self-attention and cross-attention for decoder.""" def __init__(self, config: T5Gemma2TextConfig, layer_idx: int): super().__init__(config, layer_idx) self.is_causal = False # Fused causal and encoder mask def forward( self, # decoder self-attention inputs hidden_states: torch.Tensor, position_embeddings: tuple[torch.Tensor, torch.Tensor], merged_attention_mask: torch.Tensor | None, # cross-attention inputs encoder_hidden_states: torch.Tensor, # cache inputs past_key_values: EncoderDecoderCache | None = None, cache_position: torch.LongTensor | None = None, # others **kwargs: Unpack[FlashAttentionKwargs], ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]: # attention shapes. input_shape = hidden_states.shape[:-1] hidden_shape = (*input_shape, -1, self.head_dim) cross_input_shape = encoder_hidden_states.shape[:-1] cross_hidden_shape = (*cross_input_shape, -1, self.head_dim) # self-attention. query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2) key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2) value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2) query_states = self.q_norm(query_states) key_states = self.k_norm(key_states) cos, sin = position_embeddings query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_values is not None: # self-attention. # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} self_attention_cache = past_key_values.self_attention_cache key_states, value_states = self_attention_cache.update( key_states, value_states, self.layer_idx, cache_kwargs ) # cross-attention. is_updated = past_key_values.is_updated.get(self.layer_idx) cross_attention_cache = past_key_values.cross_attention_cache if past_key_values is None or not is_updated: cross_key_states = self.k_proj(encoder_hidden_states).view(cross_hidden_shape).transpose(1, 2) cross_value_states = self.v_proj(encoder_hidden_states).view(cross_hidden_shape).transpose(1, 2) cross_key_states = self.k_norm(cross_key_states) if past_key_values is not None: cross_key_states, cross_value_states = cross_attention_cache.update( cross_key_states, cross_value_states, self.layer_idx ) past_key_values.is_updated[self.layer_idx] = True else: cross_key_states = cross_attention_cache.layers[self.layer_idx].keys cross_value_states = cross_attention_cache.layers[self.layer_idx].values # merged attention. query_states = query_states cross_key_size = cross_input_shape[1] key_states = torch.cat([key_states, cross_key_states], dim=2) value_states = torch.cat([value_states, cross_value_states], dim=2) attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface( self.config._attn_implementation, eager_attention_forward ) attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, merged_attention_mask, dropout=self.attention_dropout if self.training else 0.0, scaling=self.scaling, **kwargs, ) attn_output = attn_output.reshape(*input_shape, -1).contiguous() attn_output = self.o_proj(attn_output) # decompose merged attention weights into self & cross attention weights if attn_weights is not None: self_attn_weights = attn_weights[..., :-cross_key_size] cross_attn_weights = attn_weights[..., -cross_key_size:] else: self_attn_weights, cross_attn_weights = None, None return attn_output, self_attn_weights, cross_attn_weights def sliding_window_mask_function(sliding_window: int, is_causal=True) -> Callable: """ This creates uni/bidirectional attention mask with sliding window. """ def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: if is_causal: left_window_size, right_window_size = sliding_window, 0 else: left_window_size, right_window_size = ((sliding_window + 1) // 2, (sliding_window) // 2 + 1) dist = q_idx - kv_idx left_mask = (dist >= 0) & (dist < left_window_size) right_mask = (dist < 0) & (-dist < right_window_size) return left_mask | right_mask return inner_mask class T5Gemma2EncoderLayer(T5GemmaEncoderLayer): pass class T5Gemma2DecoderLayer(T5GemmaEncoderLayer): """Decoder sub-layer: merged attention instead of vanilla self-attention.""" def __init__(self, config, layer_idx: int): super().__init__(config, layer_idx) # replace vanilla self-attention with merged attention to support joint cross-attention. self.self_attn = T5Gemma2MergedAttention( config=config, layer_idx=layer_idx, ) def forward( self, hidden_states: torch.Tensor, position_embeddings: tuple[torch.Tensor, torch.Tensor], merged_attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: EncoderDecoderCache | None = None, use_cache: bool | None = False, cache_position: torch.LongTensor | None = None, encoder_hidden_states: torch.Tensor | None = None, **kwargs, ) -> torch.FloatTensor: residual = hidden_states hidden_states = self.pre_self_attn_layernorm(hidden_states) hidden_states, _, _ = self.self_attn( hidden_states=hidden_states, position_embeddings=position_embeddings, merged_attention_mask=merged_attention_mask, position_ids=position_ids, past_key_values=past_key_values, use_cache=use_cache, cache_position=cache_position, encoder_hidden_states=encoder_hidden_states, **kwargs, ) hidden_states = self.post_self_attn_layernorm(hidden_states) hidden_states = residual + self.dropout(hidden_states) residual = hidden_states hidden_states = self.pre_feedforward_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = self.post_feedforward_layernorm(hidden_states) hidden_states = residual + self.dropout(hidden_states) return hidden_states class T5Gemma2LMHead(T5GemmaLMHead): pass class T5Gemma2ClassificationHead(T5GemmaClassificationHead): pass class T5Gemma2MultiModalProjector(Gemma3MultiModalProjector): def __init__(self, config: T5Gemma2EncoderConfig): super().__init__(config) class T5Gemma2TextScaledWordEmbedding(Gemma3TextScaledWordEmbedding): """T5Gemma2 Embedding: override to add eoi token embedding separately.""" def __init__( self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: float = 1.0, eoi_token_index: int = 256_000, ): super().__init__(num_embeddings, embedding_dim, padding_idx, embed_scale) self.eoi_token_index = eoi_token_index self.eoi_embedding = nn.Parameter(torch.zeros(self.embedding_dim)) def forward(self, input_ids: torch.Tensor): input_embeddings = super().forward(input_ids) * self.embed_scale.to(self.weight.dtype) input_embeddings[input_ids == self.eoi_token_index] = self.eoi_embedding.to(input_embeddings.dtype) return input_embeddings @auto_docstring class T5Gemma2PreTrainedModel(Gemma3PreTrainedModel): config: T5Gemma2Config base_model_prefix = "model" supports_gradient_checkpointing = True # Mask creation is incompatible # FA due to non-default creation / SWA _supports_flash_attn = False # Flex due to custom masks not compatible to be merged after creation _supports_flex_attn = False _no_split_modules = [ "T5Gemma2EncoderLayer", "T5Gemma2DecoderLayer", "SiglipVisionEmbeddings", "SiglipEncoderLayer", "SiglipMultiheadAttentionPoolingHead", ] _can_record_outputs = { "hidden_states": [T5Gemma2EncoderLayer, T5Gemma2DecoderLayer], "attentions": [ OutputRecorder(T5Gemma2SelfAttention, index=1, layer_name="self_attn"), OutputRecorder(T5Gemma2MergedAttention, index=1, layer_name="self_attn"), OutputRecorder(T5Gemma2MergedAttention, index=2, layer_name="cross_attn"), ], } def _init_weights(self, module): PreTrainedModel._init_weights(self, module) if isinstance(module, T5Gemma2MultiModalProjector): init.zeros_(module.mm_input_projection_weight) elif isinstance(module, T5Gemma2TextScaledWordEmbedding): init.zeros_(module.eoi_embedding) init.constant_(module.embed_scale, module.scalar_embed_scale) elif isinstance(module, T5Gemma2ClassificationHead): scale = module.out_proj.weight.shape[0] ** -0.5 init.normal_(module.out_proj.weight, mean=0.0, std=self.config.initializer_range * scale) if hasattr(module.out_proj, "bias") and module.out_proj.bias is not None: init.zeros_(module.out_proj.bias) # We initialize with 0s to be 1 centered as the RMSNorm here does (1 + weight) elif "RMSNorm" in module.__class__.__name__: init.zeros_(module.weight) elif isinstance(module, T5Gemma2RotaryEmbedding): for layer_type in module.layer_types: rope_init_fn = module.compute_default_rope_parameters if module.rope_type[layer_type] != "default": rope_init_fn = ROPE_INIT_FUNCTIONS[module.rope_type[layer_type]] curr_inv_freq, _ = rope_init_fn(module.config, layer_type=layer_type) init.copy_(getattr(module, f"{layer_type}_inv_freq"), curr_inv_freq) init.copy_(getattr(module, f"{layer_type}_original_inv_freq"), curr_inv_freq) def prepare_decoder_input_ids_from_labels(self, input_ids): """ Shifts input_ids to the right, prepends the decoder_start_token_id, and handles pad_token_id replacement for labels that were -100. This is a common preparation step for decoder inputs in sequence-to-sequence models. """ decoder_config = self.config.decoder decoder_start_token_id = decoder_config.bos_token_id pad_token_id = decoder_config.pad_token_id if decoder_start_token_id is None: raise ValueError("self.model.config.decoder.bos_token_id has to be defined. ") # shift inputs to the right shifted_input_ids = input_ids.new_zeros(input_ids.shape) shifted_input_ids[..., 1:] = input_ids[..., :-1].clone() shifted_input_ids[..., 0] = decoder_start_token_id if pad_token_id is None: raise ValueError("self.model.config.decoder.pad_token_id has to be defined.") # Is this T5 specific? # replace possible -100 values in labels by `pad_token_id` shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id) return shifted_input_ids class T5Gemma2TextEncoder(T5Gemma2PreTrainedModel): config: T5Gemma2TextConfig _can_record_outputs = { "attentions": T5Gemma2SelfAttention, "hidden_states": T5Gemma2EncoderLayer, } def __init__( self, config: T5Gemma2TextConfig, eoi_token_index: int = 256_000, ): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = T5Gemma2TextScaledWordEmbedding( config.vocab_size, config.hidden_size, self.padding_idx, embed_scale=config.hidden_size**0.5, eoi_token_index=eoi_token_index, ) self.norm = T5Gemma2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.gradient_checkpointing = False self.layers = nn.ModuleList( [T5Gemma2EncoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self.dropout = nn.Dropout(config.dropout_rate) self.rotary_emb = T5Gemma2RotaryEmbedding(config) # Initialize weights and apply final processing self.post_init() @merge_with_config_defaults @capture_outputs @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, inputs_embeds: torch.FloatTensor | None = None, # Unused for processor compatibility kept in signature. token_type_ids: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutput: if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") # As we want to pass `past_key_values=None` explicitly everywhere, we need to pop them from kwargs if present kwargs.pop("past_key_values", None) if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) if position_ids is None: position_ids = torch.arange(0, inputs_embeds.shape[1], device=inputs_embeds.device).unsqueeze(0) if not isinstance(self_attn_mask_mapping := attention_mask, dict): mask_kwargs = { "config": self.config, "inputs_embeds": inputs_embeds, "attention_mask": attention_mask, } self_attn_mask_mapping = { "full_attention": create_bidirectional_mask(**mask_kwargs), "sliding_attention": create_bidirectional_mask( **mask_kwargs, and_mask_function=sliding_window_mask_function(self.config.sliding_window, is_causal=False), ), } # input layer hidden_states = inputs_embeds # global and local position embeddings position_embeddings = {} for layer_type in self.config.layer_types: position_embeddings[layer_type] = self.rotary_emb(hidden_states, position_ids, layer_type) # dropout hidden_states = self.dropout(hidden_states) for layer_module in self.layers[: self.config.num_hidden_layers]: hidden_states = layer_module( hidden_states, position_embeddings[layer_module.attention_type], self_attn_mask_mapping[layer_module.attention_type], position_ids, **kwargs, ) hidden_states = self.norm(hidden_states) hidden_states = self.dropout(hidden_states) return BaseModelOutput( last_hidden_state=hidden_states, ) class T5Gemma2Encoder(T5Gemma2PreTrainedModel): config: T5Gemma2EncoderConfig def __init__( self, config: T5Gemma2EncoderConfig, eoi_token_index: int = 256_000, ): super().__init__(config) self.text_model = T5Gemma2TextEncoder._from_config(config.text_config, eoi_token_index=eoi_token_index) self.vision_tower = AutoModel.from_config(config=config.vision_config) self.multi_modal_projector = T5Gemma2MultiModalProjector(config) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.text_model.get_input_embeddings() def set_input_embeddings(self, new_embeddings): return self.text_model.set_input_embeddings(new_embeddings) @can_return_tuple @auto_docstring def get_image_features( self, pixel_values: torch.Tensor, **kwargs: Unpack[TransformersKwargs] ) -> tuple | BaseModelOutputWithPooling: # pixel_values: (batch_size, channels, height, width) # image_features: Image feature tensor of shape (num_images, image_length, embed_dim). vision_outputs = self.vision_tower(pixel_values=pixel_values, return_dict=True, **kwargs) last_hidden_state = vision_outputs.last_hidden_state image_features = self.multi_modal_projector(last_hidden_state) vision_outputs.pooler_output = image_features return vision_outputs def get_image_placeholder_mask( self, input_ids: torch.LongTensor | None, inputs_embeds: torch.FloatTensor | None, image_features: torch.FloatTensor, ): """ Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is equal to the length of multimodal features. If the lengths are different, an error is raised. """ image_token_id = self.config.image_token_id if input_ids is None: if inputs_embeds is None: raise ValueError("Either `input_ids` or `inputs_embeds` has to be provided.") special_image_mask = inputs_embeds == self.get_input_embeddings()( torch.tensor(image_token_id, dtype=torch.long, device=inputs_embeds.device) ) special_image_mask = special_image_mask.all(-1) else: special_image_mask = input_ids == image_token_id n_image_tokens = special_image_mask.sum() special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device) n_image_features = image_features.shape[0] * image_features.shape[1] torch_compilable_check( inputs_embeds[special_image_mask].numel() == image_features.numel(), f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}", ) return special_image_mask @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, inputs_embeds: torch.FloatTensor | None = None, pixel_values: torch.FloatTensor | None = None, # Unused for processor compatibility kept in signature. token_type_ids: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutput: if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if inputs_embeds is None: inputs_embeds = self.text_model.embed_tokens(input_ids) if pixel_values is not None: image_features = self.get_image_features(pixel_values, return_dict=True).pooler_output image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype) image_mask = self.get_image_placeholder_mask( input_ids, inputs_embeds=inputs_embeds, image_features=image_features ) inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_features) outputs = self.text_model( inputs_embeds=inputs_embeds, attention_mask=attention_mask, position_ids=position_ids, **kwargs, ) return outputs class T5Gemma2Decoder(T5Gemma2PreTrainedModel): config: T5Gemma2DecoderConfig _can_record_outputs = { "attentions": OutputRecorder(T5Gemma2MergedAttention, index=1), "cross_attentions": OutputRecorder(T5Gemma2MergedAttention, index=2), "hidden_states": T5Gemma2DecoderLayer, } def __init__(self, config: T5Gemma2DecoderConfig, eoi_token_index: int = 256_000): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = T5Gemma2TextScaledWordEmbedding( config.vocab_size, config.hidden_size, config.pad_token_id, embed_scale=config.hidden_size**0.5, eoi_token_index=eoi_token_index, ) self.norm = T5Gemma2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.gradient_checkpointing = False self.layers = nn.ModuleList( [T5Gemma2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self.dropout = nn.Dropout(config.dropout_rate) self.rotary_emb = T5Gemma2RotaryEmbedding(config) self.post_init() @merge_with_config_defaults @capture_outputs @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: EncoderDecoderCache | None = None, inputs_embeds: torch.FloatTensor | None = None, use_cache: bool | None = None, cache_position: torch.LongTensor | None = None, encoder_hidden_states: torch.Tensor | None = None, encoder_attention_mask: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutputWithPastAndCrossAttentions: if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if encoder_hidden_states is None: raise ValueError("`encoder_hidden_states` must be given in decoder") if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) if not self.training and use_cache and past_key_values is None: past_key_values = EncoderDecoderCache(DynamicCache(config=self.config), DynamicCache()) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 cache_position = torch.arange( past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) if position_ids is None: position_ids = cache_position.unsqueeze(0) if not isinstance(self_attn_mask_mapping := attention_mask, dict): mask_kwargs = { "config": self.config, "inputs_embeds": inputs_embeds, "attention_mask": attention_mask, "cache_position": cache_position, "past_key_values": past_key_values.self_attention_cache if past_key_values is not None else None, "position_ids": position_ids, } # this masking function did nothing to masking but forces `allow_is_causal_skip` to be False # as we always need a mask during decoding for merged attention. mask_kwargs["and_mask_function"] = lambda *args: torch.tensor(True, dtype=torch.bool) self_attn_mask_mapping = { "full_attention": create_causal_mask(**mask_kwargs), "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs), } if not isinstance(cross_attn_mask_mapping := encoder_attention_mask, dict): mask_kwargs = { "config": self.config, "inputs_embeds": encoder_hidden_states, "attention_mask": encoder_attention_mask, "cache_position": cache_position, "past_key_values": None, "position_ids": None, } cross_attn_mask_mapping = { "full_attention": create_causal_mask( **mask_kwargs, or_mask_function=bidirectional_mask_function(encoder_attention_mask), ), } merged_attn_mask_mapping = { "full_attention": torch.cat( [self_attn_mask_mapping["full_attention"], cross_attn_mask_mapping["full_attention"]], dim=-1 ), "sliding_attention": torch.cat( [self_attn_mask_mapping["sliding_attention"], cross_attn_mask_mapping["full_attention"]], dim=-1 ), } # input layer hidden_states = inputs_embeds # global and local position embeddings position_embeddings = {} for layer_type in self.config.layer_types: position_embeddings[layer_type] = self.rotary_emb(hidden_states, position_ids, layer_type) # dropout hidden_states = self.dropout(hidden_states) for layer_module in self.layers[: self.config.num_hidden_layers]: hidden_states = layer_module( hidden_states, position_embeddings[layer_module.attention_type], merged_attn_mask_mapping[layer_module.attention_type], position_ids, past_key_values, use_cache, cache_position, encoder_hidden_states, **kwargs, ) hidden_states = self.norm(hidden_states) hidden_states = self.dropout(hidden_states) return BaseModelOutputWithPastAndCrossAttentions( last_hidden_state=hidden_states, past_key_values=past_key_values, ) @auto_docstring class T5Gemma2Model(T5Gemma2PreTrainedModel): _tied_weights_keys = { "decoder.embed_tokens.weight": "encoder.text_model.embed_tokens.weight", "decoder.embed_tokens.eoi_embedding": "encoder.text_model.embed_tokens.eoi_embedding", } def __init__(self, config: T5Gemma2Config): super().__init__(config) # setup encoder and decoder self.encoder = T5Gemma2Encoder(config.encoder, config.eoi_token_index) self.decoder = T5Gemma2Decoder(config.decoder, config.eoi_token_index) self.post_init() def get_encoder(self): return self.encoder def get_decoder(self): return self.decoder def get_input_embeddings(self): return self.encoder.get_input_embeddings() def set_input_embeddings(self, new_embeddings): return self.encoder.set_input_embeddings(new_embeddings) @can_return_tuple @auto_docstring def forward( self, # encoder inputs input_ids: torch.LongTensor | None = None, pixel_values: torch.FloatTensor | None = None, attention_mask: torch.FloatTensor | None = None, position_ids: torch.LongTensor | None = None, # decoder inputs decoder_input_ids: torch.LongTensor | None = None, decoder_attention_mask: torch.BoolTensor | None = None, decoder_position_ids: torch.LongTensor | None = None, # others (mainly inference or cache related) encoder_outputs: BaseModelOutput | None = None, past_key_values: EncoderDecoderCache | None = None, inputs_embeds: torch.Tensor | None = None, decoder_inputs_embeds: torch.Tensor | None = None, use_cache: bool | None = None, cache_position: torch.LongTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> Seq2SeqModelOutput: r""" decoder_position_ids (`torch.LongTensor` of shape `(batch_size, decoder_sequence_length)`, *optional*): Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the range `[0, config.decoder.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) """ # encoder if encoder_outputs is None: encoder_outputs = self.encoder( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, inputs_embeds=inputs_embeds, pixel_values=pixel_values, return_dict=True, **kwargs, ) encoder_hidden_states = encoder_outputs.last_hidden_state # decoder decoder_outputs = self.decoder( input_ids=decoder_input_ids, attention_mask=decoder_attention_mask, position_ids=decoder_position_ids, inputs_embeds=decoder_inputs_embeds, past_key_values=past_key_values, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=attention_mask, use_cache=use_cache, cache_position=cache_position, return_dict=True, **kwargs, ) return Seq2SeqModelOutput( last_hidden_state=decoder_outputs.last_hidden_state, past_key_values=decoder_outputs.past_key_values, decoder_hidden_states=decoder_outputs.hidden_states, decoder_attentions=decoder_outputs.attentions, cross_attentions=decoder_outputs.cross_attentions, encoder_last_hidden_state=encoder_outputs.last_hidden_state, encoder_hidden_states=encoder_outputs.hidden_states, encoder_attentions=encoder_outputs.attentions, ) class T5Gemma2ForConditionalGeneration(T5Gemma2PreTrainedModel, GenerationMixin): _tied_weights_keys = { "lm_head.out_proj.weight": "model.encoder.text_model.embed_tokens.weight", } _tp_plan = {"lm_head.out_proj": "colwise_gather_output"} _pp_plan = {"lm_head.out_proj": (["hidden_states"], ["logits"])} def __init__(self, config: T5Gemma2Config): super().__init__(config) self.model = T5Gemma2Model(config) self.vocab_size = config.decoder.vocab_size self.lm_head = T5Gemma2LMHead(config.decoder.hidden_size, self.vocab_size) self.loss_type = "ForMaskedLM" self.post_init() def set_output_embeddings(self, new_embeddings): self.lm_head.out_proj = new_embeddings def get_output_embeddings(self): return self.lm_head.out_proj def get_input_embeddings(self): return self.model.get_input_embeddings() def set_input_embeddings(self, value): self.model.set_input_embeddings(value) def get_encoder(self): return self.model.get_encoder() def get_decoder(self): return self.model.get_decoder() @can_return_tuple @auto_docstring def get_image_features( self, pixel_values: torch.Tensor, **kwargs: Unpack[TransformersKwargs] ) -> tuple | BaseModelOutputWithPooling: return self.get_encoder().get_image_features(pixel_values, **kwargs) @property def vision_tower(self): return self.get_encoder().vision_tower @can_return_tuple @auto_docstring def forward( self, # encoder inputs input_ids: torch.LongTensor | None = None, pixel_values: torch.FloatTensor | None = None, attention_mask: torch.FloatTensor | None = None, position_ids: torch.LongTensor | None = None, # decoder inputs decoder_input_ids: torch.LongTensor | None = None, decoder_attention_mask: torch.BoolTensor | None = None, decoder_position_ids: torch.LongTensor | None = None, # others (mainly inference or cache related) encoder_outputs: BaseModelOutput | None = None, past_key_values: EncoderDecoderCache | None = None, inputs_embeds: torch.FloatTensor | None = None, decoder_inputs_embeds: torch.FloatTensor | None = None, labels: torch.LongTensor | None = None, use_cache: bool | None = None, cache_position: torch.LongTensor | None = None, logits_to_keep: int | torch.Tensor = 0, **kwargs: Unpack[TransformersKwargs], ) -> tuple[torch.FloatTensor] | Seq2SeqLMOutput: r""" decoder_position_ids (`torch.LongTensor` of shape `(batch_size, decoder_sequence_length)`, *optional*): Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the range `[0, config.decoder.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. """ if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None: # get decoder inputs from shifting lm labels to the right decoder_input_ids = self.prepare_decoder_input_ids_from_labels(labels) decoder_outputs: Seq2SeqModelOutput = self.model( input_ids=input_ids, pixel_values=pixel_values, attention_mask=attention_mask, position_ids=position_ids, decoder_input_ids=decoder_input_ids, decoder_attention_mask=decoder_attention_mask, decoder_position_ids=decoder_position_ids, encoder_outputs=encoder_outputs, past_key_values=past_key_values, inputs_embeds=inputs_embeds, decoder_inputs_embeds=decoder_inputs_embeds, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = decoder_outputs.last_hidden_state # Only compute necessary logits, and do not upcast them to float if we are not computing the loss slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep logits = self.lm_head(hidden_states[:, slice_indices, :]) decoder_config = self.config.decoder if decoder_config.final_logit_softcapping is not None: logits = logits / decoder_config.final_logit_softcapping logits = torch.tanh(logits) logits = logits * decoder_config.final_logit_softcapping loss = None if labels is not None: # Input has right-shifted so we directly perform masked lm loss loss = self.loss_function(logits, labels, self.vocab_size, **kwargs) return Seq2SeqLMOutput( loss=loss, logits=logits, past_key_values=decoder_outputs.past_key_values, decoder_hidden_states=decoder_outputs.decoder_hidden_states, decoder_attentions=decoder_outputs.decoder_attentions, cross_attentions=decoder_outputs.cross_attentions, encoder_last_hidden_state=decoder_outputs.encoder_last_hidden_state, encoder_hidden_states=decoder_outputs.encoder_hidden_states, encoder_attentions=decoder_outputs.encoder_attentions, ) def _prepare_cache_for_generation( self, generation_config: GenerationConfig, model_kwargs: dict, generation_mode: GenerationMode, batch_size: int, max_cache_length: int, ) -> bool: """Override cache preparation to support T5Gemma2-specific EncoderDecoder Cache.""" # Build cache and past_key_values structure first and then override as needed. super()._prepare_cache_for_generation( generation_config, model_kwargs, generation_mode, batch_size, max_cache_length, ) # If use_cache is False, do not prepare the cache. if generation_config.use_cache is False: return cache_implementation = generation_config.cache_implementation if cache_implementation is None: offload_cache = False else: offload_cache = "offloaded" in generation_config.cache_implementation # Main change: use full cache for cross-attention. cross_attn_config = copy.deepcopy(self.config.get_text_config(decoder=True)) # cross-attention does not use sliding window del cross_attn_config.sliding_window del cross_attn_config.layer_types cross_attn_cache_kwargs = { "config": cross_attn_config, "offloading": offload_cache, } past_key_values = model_kwargs.get("past_key_values") if past_key_values is not None: if not isinstance(past_key_values, EncoderDecoderCache): raise ValueError( "The `past_key_values` in `model_kwargs` must be of type `EncoderDecoderCache` for T5Gemma2 model." ) # Cache already established, no need to re-initialize. if len(past_key_values.is_updated) > 0 and past_key_values.is_updated.get(0): return cross_attn_cls = type(past_key_values.cross_attention_cache) if cross_attn_cls == StaticCache: cross_attn_cache_kwargs["max_cache_len"] = model_kwargs["encoder_outputs"][0].shape[1] # Update cross-attention cache only (switch from sliding_window to full). past_key_values.cross_attention_cache = cross_attn_cls(**cross_attn_cache_kwargs) else: # Initialize new cache. model_kwargs["past_key_values"] = EncoderDecoderCache( DynamicCache( **{ "config": self.config.get_text_config(decoder=True), "offloading": offload_cache, } ), # self-attention cache DynamicCache(), # cross-attention cache ) if hasattr(self, "_cache") and self._cache is not None: if not isinstance(self._cache, EncoderDecoderCache): raise ValueError("The internal cache must be of type `EncoderDecoderCache` for T5Gemma2 model.") self._cache = model_kwargs["past_key_values"] @auto_docstring class T5Gemma2ForSequenceClassification(T5Gemma2PreTrainedModel): def __init__(self, config: T5Gemma2Config): super().__init__(config) self.num_labels = config.num_labels self.hidden_size = config.decoder.hidden_size self.model = T5Gemma2Model(config) classifier_dropout = getattr(config, "classifier_dropout_rate", 0.1) self.score = T5Gemma2ClassificationHead(self.hidden_size, self.num_labels, classifier_dropout) self.post_init() def get_input_embeddings(self): return self.model.get_input_embeddings() def set_input_embeddings(self, value): self.model.set_input_embeddings(value) @can_return_tuple @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, pixel_values: torch.FloatTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, decoder_input_ids: torch.LongTensor | None = None, decoder_attention_mask: torch.Tensor | None = None, decoder_position_ids: torch.LongTensor | None = None, encoder_outputs: BaseModelOutput | None = None, inputs_embeds: torch.FloatTensor | None = None, decoder_inputs_embeds: torch.FloatTensor | None = None, labels: torch.LongTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> SequenceClassifierOutput: r""" decoder_position_ids (`torch.LongTensor` of shape `(batch_size, decoder_sequence_length)`, *optional*): Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the range `[0, config.decoder.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). """ if inputs_embeds is not None or decoder_inputs_embeds is not None: raise NotImplementedError( f"Passing input embeddings is currently not supported for {self.__class__.__name__}." ) if input_ids is None: raise ValueError("You have to specify input_ids") if decoder_input_ids is None: decoder_input_ids = self.prepare_decoder_input_ids_from_labels(input_ids) outputs: Seq2SeqModelOutput = self.model( input_ids, pixel_values=pixel_values, attention_mask=attention_mask, position_ids=position_ids, decoder_input_ids=decoder_input_ids, decoder_attention_mask=decoder_attention_mask, decoder_position_ids=decoder_position_ids, encoder_outputs=encoder_outputs, inputs_embeds=inputs_embeds, decoder_inputs_embeds=decoder_inputs_embeds, use_cache=False, **kwargs, ) last_hidden_state = outputs.last_hidden_state hidden_states = outputs.decoder_hidden_states attentions = outputs.decoder_attentions logits = self.score(last_hidden_state) batch_size = input_ids.shape[0] # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id non_pad_mask = (decoder_input_ids != self.config.pad_token_id).to(logits.device, torch.int32) token_indices = torch.arange(decoder_input_ids.shape[-1], device=logits.device, dtype=torch.int32) last_non_pad_token = (token_indices * non_pad_mask).argmax(-1) last_non_pad_token = torch.clamp(last_non_pad_token, max=decoder_input_ids.shape[-1] - 1) pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token] loss = None if labels is not None: loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config) return SequenceClassifierOutput( loss=loss, logits=pooled_logits, hidden_states=hidden_states, attentions=attentions, ) @auto_docstring class T5Gemma2ForTokenClassification(T5Gemma2PreTrainedModel): def __init__(self, config: T5Gemma2Config): super().__init__(config) self.num_labels = config.num_labels self.hidden_size = config.decoder.hidden_size self.model = T5Gemma2Model(config) classifier_dropout = getattr(config, "classifier_dropout_rate", 0.1) self.score = T5Gemma2ClassificationHead(self.hidden_size, self.num_labels, classifier_dropout) self.post_init() def get_input_embeddings(self): return self.model.get_input_embeddings() def set_input_embeddings(self, value): self.model.set_input_embeddings(value) @can_return_tuple @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, pixel_values: torch.FloatTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, decoder_input_ids: torch.LongTensor | None = None, decoder_attention_mask: torch.Tensor | None = None, decoder_position_ids: torch.LongTensor | None = None, encoder_outputs: BaseModelOutput | None = None, inputs_embeds: torch.FloatTensor | None = None, decoder_inputs_embeds: torch.FloatTensor | None = None, labels: torch.LongTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> TokenClassifierOutput: r""" decoder_position_ids (`torch.LongTensor` of shape `(batch_size, decoder_sequence_length)`, *optional*): Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the range `[0, config.decoder.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). """ if inputs_embeds is not None or decoder_inputs_embeds is not None: raise NotImplementedError( f"Passing input embeddings is currently not supported for {self.__class__.__name__}." ) if input_ids is None: raise ValueError("You have to specify input_ids") if decoder_input_ids is None: decoder_input_ids = self.prepare_decoder_input_ids_from_labels(input_ids) outputs: Seq2SeqModelOutput = self.model( input_ids, pixel_values=pixel_values, attention_mask=attention_mask, position_ids=position_ids, decoder_input_ids=decoder_input_ids, decoder_attention_mask=decoder_attention_mask, decoder_position_ids=decoder_position_ids, encoder_outputs=encoder_outputs, inputs_embeds=inputs_embeds, decoder_inputs_embeds=decoder_inputs_embeds, use_cache=False, **kwargs, ) last_hidden_state = outputs.last_hidden_state hidden_states = outputs.decoder_hidden_states attentions = outputs.decoder_attentions logits = self.score(last_hidden_state) loss = None if labels is not None: loss = self.loss_function(logits, labels, self.config) return TokenClassifierOutput( loss=loss, logits=logits, hidden_states=hidden_states, attentions=attentions, ) __all__ = [ "T5Gemma2Config", "T5Gemma2TextConfig", "T5Gemma2EncoderConfig", "T5Gemma2DecoderConfig", "T5Gemma2ForConditionalGeneration", "T5Gemma2Model", "T5Gemma2Encoder", "T5Gemma2PreTrainedModel", "T5Gemma2ForSequenceClassification", "T5Gemma2ForTokenClassification", ]