# coding=utf-8 # 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 import warnings from collections.abc import Callable from typing import Any, Optional, Union import torch import torch.nn as nn import torch.utils.checkpoint from ...cache_utils import Cache, DynamicCache from ...configuration_utils import PretrainedConfig, layer_type_validation from ...masking_utils import create_causal_mask, create_masks_for_generate, create_sliding_window_causal_mask from ...modeling_flash_attention_utils import FlashAttentionKwargs from ...modeling_layers import GradientCheckpointingLayer from ...modeling_outputs import BaseModelOutputWithPast, SequenceClassifierOutputWithPast from ...modeling_rope_utils import rope_config_validation from ...modeling_utils import ALL_ATTENTION_FUNCTIONS from ...processing_utils import Unpack from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging from ..gemma2.configuration_gemma2 import Gemma2Config from ..gemma2.modeling_gemma2 import ( Gemma2Attention, Gemma2ForCausalLM, Gemma2MLP, Gemma2Model, Gemma2PreTrainedModel, Gemma2RMSNorm, Gemma2RotaryEmbedding, apply_rotary_pos_emb, eager_attention_forward, ) from ..paligemma.modeling_paligemma import ( PaligemmaCausalLMOutputWithPast, PaliGemmaForConditionalGeneration, PaliGemmaModel, PaligemmaModelOutputWithPast, ) from ..siglip import SiglipVisionConfig logger = logging.get_logger(__name__) class Gemma3TextConfig(Gemma2Config, PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Gemma3TextModel`]. It is used to instantiate an Gemma3Text 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 Gemma3Text-7B. e.g. [google/gemma3_text-7b](https://huggingface.co/google/gemma3_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 Gemma3Text model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`Gemma3TextModel`] 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. tie_word_embeddings (`bool`, *optional*, defaults to `True`): Whether to tie weight embeddings rope_theta (`float`, *optional*, defaults to 1000000.0): The base period of the RoPE embeddings. 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 Gemma3Text, 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_scaling (`Dict`, *optional*): Dictionary containing the scaling configuration for the RoPE embeddings used in global attention. NOTE: if you apply new rope type and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value accordingly. Expected contents: `rope_type` (`str`): The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope', 'llama3'], with 'default' being the original RoPE implementation. `factor` (`float`, *optional*): Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In most scaling types, a `factor` of x will enable the model to handle sequences of length x * original maximum pre-trained length. `original_max_position_embeddings` (`int`, *optional*): Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during pretraining. `attention_factor` (`float`, *optional*): Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention computation. If unspecified, it defaults to value recommended by the implementation, using the `factor` field to infer the suggested value. `beta_fast` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear ramp function. If unspecified, it defaults to 32. `beta_slow` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear ramp function. If unspecified, it defaults to 1. `short_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to short contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `long_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to long contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `low_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE `high_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE rope_local_base_freq (float, *optional*, defaults to 10000.0): The base period of the RoPE embeddings for local attention. ```python >>> from transformers import Gemma3TextModel, Gemma3TextConfig >>> # Initializing a Gemma3Text gemma3_text-7b style configuration >>> configuration = Gemma3TextConfig() >>> # Initializing a model from the gemma3_text-7b style configuration >>> model = Gemma3TextModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` """ model_type = "gemma3_text" def __init__( self, vocab_size=262_208, hidden_size=2304, intermediate_size=9216, num_hidden_layers=26, num_attention_heads=8, num_key_value_heads=4, head_dim=256, hidden_activation="gelu_pytorch_tanh", max_position_embeddings=131_072, initializer_range=0.02, rms_norm_eps=1e-6, use_cache=True, pad_token_id=0, eos_token_id=1, bos_token_id=2, tie_word_embeddings=True, rope_theta=1_000_000.0, attention_bias=False, attention_dropout=0.0, query_pre_attn_scalar=256, sliding_window=4096, layer_types=None, final_logit_softcapping=None, attn_logit_softcapping=None, rope_scaling=None, rope_local_base_freq=10_000.0, **kwargs, ): PretrainedConfig.__init__( pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs, ) 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.rope_theta = rope_theta 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 self.rope_local_base_freq = rope_local_base_freq self.rope_scaling = rope_scaling rope_config_validation(self) # 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) @property def sliding_window_pattern(self): warnings.warn( "The `sliding_window_pattern` attribute is deprecated and will be removed in v4.55.0.", FutureWarning, ) return self._sliding_window_pattern @sliding_window_pattern.setter def sliding_window_pattern(self, value): self._sliding_window_pattern = value class Gemma3Config(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Gemma3ForConditionalGeneration`]. It is used to instantiate an Gemma3ForConditionalGeneration according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the PaliGemma-2B. e.g. [google/gemma-3-4b](https://huggingface.co/google/gemma-3-4b) Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: text_config (`Union[Gemma3TextConfig, dict]`, *optional*): The config object of the text backbone. vision_config (`Union[AutoConfig, dict]`, *optional*): Custom vision config or dict. mm_tokens_per_image (`int`, *optional*, defaults to 256): The number of tokens per image embedding. boi_token_index (`int`, *optional*, defaults to 255999): The begin-of-image token index to wrap the image prompt. eoi_token_index (`int`, *optional*, defaults to 256000): The end-of-image token index to wrap the image prompt. image_token_index (`int`, *optional*, defaults to 262144): The image token index to encode the image prompt. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. Example: ```python >>> from transformers import Gemma3ForConditionalGeneration, Gemma3Config, SiglipVisionConfig, Gemma3TextConfig >>> # Initializing a Siglip-like vision config >>> vision_config = SiglipVisionConfig() >>> # Initializing a Gemma3 Text config >>> text_config = Gemma3TextConfig() >>> # Initializing a Gemma3 gemma-3-4b style configuration >>> configuration = Gemma3Config(vision_config, text_config) >>> # Initializing a model from the gemma-3-4b style configuration >>> model = Gemma3TextConfig(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "gemma3" attribute_map = { "image_token_id": "image_token_index", "boi_token_id": "boi_token_index", "eoi_token_id": "eoi_token_index", } sub_configs = { "text_config": Gemma3TextConfig, "vision_config": SiglipVisionConfig, } def __init__( self, text_config: Optional[Union[Gemma3TextConfig, dict[str, Any]]] = None, vision_config: Optional[Union[SiglipVisionConfig, dict[str, Any]]] = None, mm_tokens_per_image: int = 256, boi_token_index: int = 255_999, eoi_token_index: int = 256_000, image_token_index: int = 262_144, initializer_range: float = 0.02, **kwargs, ): if text_config is None: text_config = Gemma3TextConfig() logger.info("text_config is None, using default Gemma3TextConfig text config.") elif isinstance(text_config, dict): text_config = Gemma3TextConfig(**text_config) if isinstance(vision_config, dict): vision_config = SiglipVisionConfig(**vision_config) elif vision_config is None: vision_config = SiglipVisionConfig() logger.info("vision_config is None, using default SiglipVisionConfig vision config.") self.text_config = text_config self.vision_config = vision_config self.mm_tokens_per_image = mm_tokens_per_image self.boi_token_index = boi_token_index self.eoi_token_index = eoi_token_index self.image_token_index = image_token_index self.initializer_range = initializer_range super().__init__(**kwargs) class Gemma3ModelOutputWithPast(PaligemmaModelOutputWithPast): pass class Gemma3CausalLMOutputWithPast(PaligemmaCausalLMOutputWithPast): pass class Gemma3TextScaledWordEmbedding(nn.Embedding): """ This module overrides nn.Embeddings' forward by multiplying with embeddings scale. """ def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: float = 1.0): super().__init__(num_embeddings, embedding_dim, padding_idx) self.register_buffer("embed_scale", torch.tensor(embed_scale), persistent=False) def forward(self, input_ids: torch.Tensor): return super().forward(input_ids) * self.embed_scale.to(self.weight.dtype) class Gemma3MLP(Gemma2MLP): def __init__(self, config: Gemma3TextConfig): super().__init__(config) class Gemma3RMSNorm(Gemma2RMSNorm): def __init__(self, dim: int, eps: float = 1e-6): super().__init__() class Gemma3RotaryEmbedding(Gemma2RotaryEmbedding): def __init__(self, config: Gemma3TextConfig, device=None): super().__init__(config) # Weird way to inherit but otherwise the sliding window gets defined first and can't access `is_sliding` class Gemma3Attention(Gemma2Attention): def __init__(self, config: Gemma3TextConfig, layer_idx: int): self.is_sliding = config.layer_types[layer_idx] == "sliding_attention" super().__init__() self.sliding_window = config.sliding_window if self.is_sliding else None self.q_norm = Gemma3RMSNorm(dim=config.head_dim, eps=config.rms_norm_eps) self.k_norm = Gemma3RMSNorm(dim=config.head_dim, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, position_embeddings: torch.Tensor, attention_mask: Optional[torch.Tensor], past_key_value: Optional[Cache] = None, cache_position: Optional[torch.LongTensor] = None, **kwargs: Unpack[FlashAttentionKwargs], ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]: input_shape = hidden_states.shape[:-1] hidden_shape = (*input_shape, -1, self.head_dim) 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_value is not None: # 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} key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) attention_interface: Callable = eager_attention_forward if self.config._attn_implementation != "eager": attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=self.attention_dropout if self.training else 0.0, scaling=self.scaling, sliding_window=self.sliding_window, **kwargs, ) attn_output = attn_output.reshape(*input_shape, -1).contiguous() attn_output = self.o_proj(attn_output) return attn_output, attn_weights class Gemma3DecoderLayer(GradientCheckpointingLayer): def __init__(self, config: Gemma3TextConfig, layer_idx: int): super().__init__() self.config = config self.hidden_size = config.hidden_size self.layer_idx = layer_idx self.attention_type = config.layer_types[layer_idx] self.self_attn = Gemma3Attention(config=config, layer_idx=layer_idx) self.mlp = Gemma3MLP(config) self.input_layernorm = Gemma3RMSNorm(self.hidden_size, eps=config.rms_norm_eps) self.post_attention_layernorm = Gemma3RMSNorm(self.hidden_size, eps=config.rms_norm_eps) self.pre_feedforward_layernorm = Gemma3RMSNorm(self.hidden_size, eps=config.rms_norm_eps) self.post_feedforward_layernorm = Gemma3RMSNorm(self.hidden_size, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, position_embeddings_global: torch.Tensor, position_embeddings_local: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]: residual = hidden_states hidden_states = self.input_layernorm(hidden_states) # apply global RoPE to non-sliding layer only if self.self_attn.is_sliding: position_embeddings = position_embeddings_local else: position_embeddings = position_embeddings_global hidden_states, self_attn_weights = self.self_attn( hidden_states=hidden_states, position_embeddings=position_embeddings, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = residual + 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 + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) return outputs GEMMA3_START_DOCSTRING = None class Gemma3PreTrainedModel(Gemma2PreTrainedModel): base_model_prefix = "" _no_split_modules = [ "Gemma3DecoderLayer", "SiglipVisionEmbeddings", "SiglipEncoderLayer", "SiglipMultiheadAttentionPoolingHead", ] def _init_weights(self, module): Gemma2PreTrainedModel._init_weights(module) if isinstance(module, Gemma3MultiModalProjector): module.mm_input_projection_weight.data.zero_() class Gemma3TextModel(Gemma2Model): config: Gemma3TextConfig def __init__(self, config: Gemma3TextConfig): super().__init__(config) # Gemma3 downcasts the below to bfloat16, causing sqrt(3072)=55.4256 to become 55.5. See https://github.com/huggingface/transformers/pull/29402 self.embed_tokens = Gemma3TextScaledWordEmbedding( config.vocab_size, config.hidden_size, self.padding_idx, embed_scale=self.config.hidden_size**0.5 ) # TODO: raushan fix this after RoPE refactor. For now we hack it by reassigning thetas # when we want to create a local RoPE layer. Config defaults should hold values for global RoPE config = copy.deepcopy(config) config.rope_theta = config.rope_local_base_freq config.rope_scaling = {"rope_type": "default"} self.rotary_emb_local = Gemma3RotaryEmbedding(config=config) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutputWithPast: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if self.gradient_checkpointing and self.training and use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." ) use_cache = False if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) if use_cache and past_key_values is None and not self.training: past_key_values = 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) # It may already have been prepared by e.g. `generate` if not isinstance(causal_mask_mapping := attention_mask, dict): # Prepare mask arguments mask_kwargs = { "config": self.config, "input_embeds": inputs_embeds, "attention_mask": attention_mask, "cache_position": cache_position, "past_key_values": past_key_values, "position_ids": position_ids, } # Create the masks causal_mask_mapping = { "full_attention": create_causal_mask(**mask_kwargs), "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs), } # embed positions hidden_states = inputs_embeds # create position embeddings to be shared across the decoder layers position_embeddings_global = self.rotary_emb(hidden_states, position_ids) position_embeddings_local = self.rotary_emb_local(hidden_states, position_ids) # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None for decoder_layer in self.layers[: self.config.num_hidden_layers]: if output_hidden_states: all_hidden_states += (hidden_states,) layer_outputs = decoder_layer( hidden_states, position_embeddings_global=position_embeddings_global, position_embeddings_local=position_embeddings_local, attention_mask=causal_mask_mapping[decoder_layer.attention_type], position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = layer_outputs[0] if output_attentions: all_self_attns += (layer_outputs[1],) hidden_states = self.norm(hidden_states) if output_hidden_states: all_hidden_states += (hidden_states,) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values, hidden_states=all_hidden_states, attentions=all_self_attns, ) class Gemma3ForCausalLM(Gemma2ForCausalLM): config: Gemma3TextConfig base_model_prefix = "language_model" def __init__(self, config: Gemma3TextConfig): super().__init__(config) self.model = Gemma3TextModel(config) class Gemma3MultiModalProjector(nn.Module): def __init__(self, config: Gemma3Config): super().__init__() self.mm_input_projection_weight = nn.Parameter( torch.zeros(config.vision_config.hidden_size, config.text_config.hidden_size) ) self.mm_soft_emb_norm = Gemma3RMSNorm( config.vision_config.hidden_size, eps=config.vision_config.layer_norm_eps ) self.patches_per_image = int(config.vision_config.image_size // config.vision_config.patch_size) self.tokens_per_side = int(config.mm_tokens_per_image**0.5) self.kernel_size = self.patches_per_image // self.tokens_per_side self.avg_pool = nn.AvgPool2d(kernel_size=self.kernel_size, stride=self.kernel_size) def forward(self, vision_outputs: torch.Tensor): batch_size, _, seq_length = vision_outputs.shape reshaped_vision_outputs = vision_outputs.transpose(1, 2) reshaped_vision_outputs = reshaped_vision_outputs.reshape( batch_size, seq_length, self.patches_per_image, self.patches_per_image ) reshaped_vision_outputs = reshaped_vision_outputs.contiguous() pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs) pooled_vision_outputs = pooled_vision_outputs.flatten(2) pooled_vision_outputs = pooled_vision_outputs.transpose(1, 2) normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs) projected_vision_outputs = torch.matmul(normed_vision_outputs, self.mm_input_projection_weight) return projected_vision_outputs.type_as(vision_outputs) def token_type_ids_mask_function( token_type_ids: Optional[torch.Tensor], image_group_ids: Optional[torch.Tensor], tokens_per_image: int, ) -> Optional[Callable]: """ This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths, not start and end indices. """ # Do not return an additional mask in this case if token_type_ids is None: return None def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool: # If it's 1 for both query and key/value, we are in an image block # NOTE: static cache shape goes beyond input seq length, while token_type_ids.shape[1] == input seq length # Since vmap doesn't support `if statement` we workaround it with `torch.where` safe_idx = torch.where(kv_idx < token_type_ids.shape[1], kv_idx, 0) token_type_ids_at_kv_idx = token_type_ids[batch_idx, safe_idx] token_type_ids_at_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], token_type_ids_at_kv_idx, 0) image_group_ids_at_kv_idx = image_group_ids[batch_idx, safe_idx] image_group_ids_at_kv_idx = torch.where(kv_idx < image_group_ids.shape[1], image_group_ids_at_kv_idx, -1) is_image_block = (token_type_ids[batch_idx, q_idx] == 1) & (token_type_ids_at_kv_idx == 1) same_image_block = image_group_ids[batch_idx, q_idx] == image_group_ids_at_kv_idx # This is bidirectional attention whenever we are dealing with image tokens return is_image_block & same_image_block return inner_mask class Gemma3Model(PaliGemmaModel): # we are filtering the logits/labels so we shouldn't divide the loss based on num_items_in_batch accepts_loss_kwargs = False def get_image_features(self, pixel_values: torch.Tensor) -> torch.Tensor: """ Projects the last hidden state from the vision model into language model space. Args: pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`) The tensors corresponding to the input images. Returns: image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`). """ vision_outputs = self.vision_tower(pixel_values=pixel_values).last_hidden_state image_features = self.multi_modal_projector(vision_outputs) return image_features def _update_causal_mask(self, **super_kwargs): raise AttributeError("We don't want to inherit it") @can_return_tuple @auto_docstring def forward( self, input_ids: torch.LongTensor = None, pixel_values: torch.FloatTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Union[list[torch.FloatTensor], Cache]] = None, token_type_ids: Optional[torch.LongTensor] = None, cache_position: Optional[torch.LongTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, **lm_kwargs, ) -> Union[tuple, Gemma3ModelOutputWithPast]: if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict # Replace image id woth PAD if the image token if OOV, to avoid index-errors if input_ids is not None and self.config.image_token_id >= self.vocab_size: special_image_mask = input_ids == self.config.image_token_id llm_input_ids = input_ids.clone() llm_input_ids[special_image_mask] = 0 else: llm_input_ids = input_ids if inputs_embeds is None: inputs_embeds = self.get_input_embeddings()(llm_input_ids) 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 ) # Merge text and images if pixel_values is not None: image_features = self.get_image_features(pixel_values) image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype) special_image_mask = self.get_placeholder_mask( input_ids, inputs_embeds=inputs_embeds, image_features=image_features ) inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features) # It may already have been prepared by e.g. `generate` if not isinstance(causal_mask_mapping := attention_mask, dict): # Prepare mask arguments mask_kwargs = { "config": self.config.get_text_config(), "input_embeds": inputs_embeds, "attention_mask": attention_mask, "cache_position": cache_position, "past_key_values": past_key_values, "position_ids": position_ids, } if token_type_ids is not None and inputs_embeds.shape[1] != 1: # We need to pass an additional mask function to account for token type ids, and it needs to be an `or` # First find where a new image block starts: 1 if image and previous not image # The images cannot attend to future images, but can attend to all prev images and to itself bidirectionally is_image = (token_type_ids == 1).to(cache_position.device) new_image_start = is_image & ~nn.functional.pad(is_image, (1, 0), value=0)[:, :-1] image_group_ids = torch.cumsum(new_image_start.int(), dim=1) - 1 image_group_ids = torch.where(is_image, image_group_ids, torch.full_like(token_type_ids, -1)) mask_kwargs["or_mask_function"] = token_type_ids_mask_function( token_type_ids.to(cache_position.device), image_group_ids, self.config.mm_tokens_per_image ) # Create the masks causal_mask_mapping = { "full_attention": create_causal_mask(**mask_kwargs), "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs), } outputs = self.language_model( attention_mask=causal_mask_mapping, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=True, cache_position=cache_position, **lm_kwargs, ) return Gemma3ModelOutputWithPast( last_hidden_state=outputs.last_hidden_state, past_key_values=outputs.past_key_values if use_cache else None, hidden_states=outputs.hidden_states, attentions=outputs.attentions, image_hidden_states=image_features if pixel_values is not None else None, ) class Gemma3ForConditionalGeneration(PaliGemmaForConditionalGeneration): @auto_docstring def forward( self, input_ids: torch.LongTensor = None, pixel_values: torch.FloatTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Union[list[torch.FloatTensor], Cache]] = None, token_type_ids: Optional[torch.LongTensor] = None, cache_position: Optional[torch.LongTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, logits_to_keep: Union[int, torch.Tensor] = 0, **lm_kwargs, ) -> Union[tuple, Gemma3CausalLMOutputWithPast]: r""" 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.text_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.text_config.vocab_size]`. Example: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, Gemma3ForConditionalGeneration >>> model = Gemma3ForConditionalGeneration.from_pretrained("google/gemma-3-4b-it") >>> processor = AutoProcessor.from_pretrained("google/gemma-3-4b-it") >>> messages = [ ... { ... "role": "system", ... "content": [ ... {"type": "text", "text": "You are a helpful assistant."} ... ] ... }, ... { ... "role": "user", "content": [ ... {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"}, ... {"type": "text", "text": "Where is the cat standing?"}, ... ] ... }, ... ] >>> inputs = processor.apply_chat_template( ... messages, ... tokenize=True, ... return_dict=True, ... return_tensors="pt", ... add_generation_prompt=True ... ) >>> # Generate >>> generate_ids = model.generate(**inputs) >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] "user\nYou are a helpful assistant.\n\n\n\n\n\nWhere is the cat standing?\nmodel\nBased on the image, the cat is standing in a snowy area, likely outdoors. It appears to" ``` """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict outputs = self.model( input_ids=input_ids, pixel_values=pixel_values, token_type_ids=token_type_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, labels=labels, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, cache_position=cache_position, **lm_kwargs, ) hidden_states = outputs[0] # 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, :]) loss = None if labels is not None: # Upcast to float if we need to compute the loss to avoid potential precision issues logits = logits.float() shift_logits = logits[..., :-1, :] shift_labels = labels[..., 1:] if attention_mask is not None: # we use the input attention mask to shift the logits and labels, because it is 2D. # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft shift_attention_mask = attention_mask[:, -shift_logits.shape[1] :].to(logits.device) shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous() shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous() else: shift_logits = shift_logits.contiguous() shift_labels = shift_labels.contiguous() # Flatten the tokens loss_fct = nn.CrossEntropyLoss() flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size) flat_labels = shift_labels.view(-1).to(shift_logits.device) loss = loss_fct(flat_logits, flat_labels) if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return Gemma3CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, image_hidden_states=outputs.image_hidden_states, ) def prepare_inputs_for_generation( self, input_ids, past_key_values=None, inputs_embeds=None, cache_position=None, position_ids=None, pixel_values=None, attention_mask=None, token_type_ids=None, use_cache=True, logits_to_keep=None, labels=None, **kwargs, ): # Overwritten -- custom `position_ids` and `pixel_values` handling model_inputs = super().prepare_inputs_for_generation( input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, attention_mask=attention_mask, position_ids=position_ids, cache_position=cache_position, use_cache=use_cache, logits_to_keep=logits_to_keep, token_type_ids=token_type_ids, **kwargs, ) # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore # Otherwise we need pixel values to be passed to model. NOTE: use_cache=False needs pixel_values always if cache_position[0] == 0: model_inputs["pixel_values"] = pixel_values return model_inputs def _prepare_4d_causal_attention_mask_with_cache_position(self, **super_kwargs): raise AttributeError("We don't want to inherit it") @staticmethod def create_masks_for_generate( config: PretrainedConfig, input_embeds: torch.Tensor, attention_mask: Optional[torch.Tensor], cache_position: torch.Tensor, past_key_values: Optional[Cache], position_ids: Optional[torch.Tensor], token_type_ids: Optional[torch.Tensor] = None, **kwargs, ) -> dict: # Prepare mask arguments mask_kwargs = { "config": config.get_text_config(), "input_embeds": input_embeds, "attention_mask": attention_mask, "cache_position": cache_position, "past_key_values": past_key_values, "position_ids": position_ids, } # Add the token type ids mask for generate as well if token_type_ids is not None and input_embeds.shape[1] != 1: # We need to pass an additional mask function to account for token type ids, and it needs to be an `or` # First find where a new image block starts: 1 if image and previous not image # The images cannot attend to future images, but can attend to all prev images and to itself bidirectionally is_image = (token_type_ids == 1).to(cache_position.device) new_image_start = is_image & ~nn.functional.pad(is_image, (1, 0), value=0)[:, :-1] image_group_ids = torch.cumsum(new_image_start.int(), dim=1) - 1 image_group_ids = torch.where(is_image, image_group_ids, torch.full_like(token_type_ids, -1)) mask_kwargs["or_mask_function"] = token_type_ids_mask_function( token_type_ids.to(cache_position.device), image_group_ids, config.mm_tokens_per_image ) return create_masks_for_generate(**mask_kwargs) class Gemma3ForSequenceClassification(Gemma3PreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.model = Gemma3Model(config) self.score = nn.Linear(config.text_config.hidden_size, self.num_labels, bias=False) # Initialize weights and apply final processing 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, pixel_values: Optional[torch.FloatTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, inputs_embeds: Optional[torch.FloatTensor] = None, token_type_ids: Optional[torch.LongTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, **kwargs: Unpack[TransformersKwargs], ) -> SequenceClassifierOutputWithPast: r""" 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). """ transformer_outputs = self.model( input_ids, attention_mask=attention_mask, pixel_values=pixel_values, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, token_type_ids=token_type_ids, use_cache=use_cache, **kwargs, ) hidden_states = transformer_outputs.last_hidden_state logits = self.score(hidden_states) if input_ids is not None: batch_size = input_ids.shape[0] else: batch_size = inputs_embeds.shape[0] if self.config.text_config.pad_token_id is None and batch_size != 1: raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") if self.config.text_config.pad_token_id is None: last_non_pad_token = -1 elif input_ids is not None: # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id non_pad_mask = (input_ids != self.config.text_config.pad_token_id).to(logits.device, torch.int32) token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32) last_non_pad_token = (token_indices * non_pad_mask).argmax(-1) else: last_non_pad_token = -1 logger.warning_once( f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be " "unexpected if using padding tokens in conjunction with `inputs_embeds.`" ) 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 SequenceClassifierOutputWithPast( loss=loss, logits=pooled_logits, past_key_values=transformer_outputs.past_key_values, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions, ) __all__ = [ "Gemma3Config", "Gemma3TextConfig", "Gemma3PreTrainedModel", # noqa: F822 "Gemma3TextModel", "Gemma3ForCausalLM", "Gemma3ForConditionalGeneration", "Gemma3Model", "Gemma3ForSequenceClassification", ]