# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/embedding_gemma2/modular_embedding_gemma2.py. # Do NOT edit this file manually as any edits will be overwritten by the generation of # the file from the modular. If any change should be done, please apply the change to the # modular_embedding_gemma2.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # Copyright 2026 the HuggingFace 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. from typing import Any, Literal from huggingface_hub.dataclasses import strict from ...configuration_utils import PreTrainedConfig, SubConfigSpec from ...utils import ( auto_docstring, logging, ) from ..auto import AutoConfig logger = logging.get_logger(__name__) @auto_docstring(checkpoint="google/embeddinggemma-2") @strict class EmbeddingGemma2TextConfig(PreTrainedConfig): r""" sliding_window (`int`, *optional*, defaults to 512): Inclusive radius of the bidirectional sliding window: a `sliding_attention` layer attends to every position with `abs(q_idx - kv_idx) <= sliding_window`. It is a radius and not the one-sided width the causal Gemma models configure, because the mask here is symmetric, so the default is half of the 1024-wide window the reference implementation states. hidden_size_per_layer_input (`int`, *optional*, defaults to 512): Dimensionality of the per-layer (PLE) residual signal. EmbeddingGemma 2 uses *projection-only* PLE: the signal is derived from `inputs_embeds` alone, with no auxiliary token lookup table. embedding_dim (`int`, *optional*, defaults to 768): Dimensionality of the sentence embedding produced by `embedding_projection`. """ model_type = "embedding_gemma2_text" base_model_tp_plan = { "layers.*.self_attn.q_proj": "colwise", "layers.*.self_attn.k_proj": "colwise", "layers.*.self_attn.v_proj": "colwise", "layers.*.self_attn.q_norm": "replicated_with_grad_allreduce", "layers.*.self_attn.k_norm": "replicated_with_grad_allreduce", "layers.*.self_attn.v_norm": "replicated_with_grad_allreduce", "layers.*.self_attn.o_proj": "rowwise", "layers.*.mlp.gate_proj": "colwise", "layers.*.mlp.up_proj": "colwise", "layers.*.mlp.down_proj": "rowwise", } base_model_pp_plan = { "embed_tokens": (["input_ids"], ["inputs_embeds"]), "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), "norm": (["hidden_states"], ["hidden_states"]), } vocab_size: int = 262_144 hidden_size: int = 512 intermediate_size: int = 2048 num_hidden_layers: int = 24 num_attention_heads: int = 4 num_key_value_heads: int = 2 head_dim: int = 256 hidden_activation: str = "gelu_pytorch_tanh" max_position_embeddings: int = 262_144 initializer_range: float = 0.02 rms_norm_eps: float = 1e-6 pad_token_id: int | None = 0 eos_token_id: int | list[int] | None = 1 bos_token_id: int | None = 2 rope_parameters: dict | None = None attention_bias: bool = False attention_dropout: int | float | None = 0.0 sliding_window: int = 512 layer_types: list[str] | None = None hidden_size_per_layer_input: int = 512 embedding_dim: int = 768 def __post_init__(self, **kwargs): # `sliding_window_pattern`, `global_head_dim` and `num_global_key_value_heads` are builder-only # kwargs, as in Gemma 4: they shape `layer_types` and `per_layer_config` and are not kept on # the config, which exposes the derived values instead. sliding_window_pattern = kwargs.pop("sliding_window_pattern", 6) if sliding_window_pattern < 1: raise ValueError(f"`sliding_window_pattern` must be a positive integer, got {sliding_window_pattern}.") if self.layer_types is None: self.layer_types = [ "sliding_attention" if bool((i + 1) % sliding_window_pattern) else "full_attention" for i in range(self.num_hidden_layers) ] if self.layer_types and (last_layer_type := self.layer_types[-1]) != "full_attention": logger.warning( f"Last layer must use `full_attention`, but got `{last_layer_type}`. Forcing last layer to `full_attention`." ) self.layer_types[-1] = "full_attention" default_rope_params: dict[Literal["full_attention", "sliding_attention"] : dict[str, Any]] = { "sliding_attention": {"rope_type": "default", "rope_theta": 10_000.0}, "full_attention": {"rope_type": "default", "rope_theta": 1_000_000.0}, } if self.rope_parameters is None: self.rope_parameters = default_rope_params # Full-attention layers are wider, hence the separate head dimension and kv head count. global_head_dim = kwargs.pop("global_head_dim", 512) num_global_key_value_heads = kwargs.pop("num_global_key_value_heads", 1) if "per_layer_config" not in kwargs: kwargs["per_layer_config"] = { layer_idx: {"head_dim": global_head_dim, "num_key_value_heads": num_global_key_value_heads} for layer_idx, layer_type in enumerate(self.layer_types) if layer_type == "full_attention" } super().__post_init__(**kwargs) def convert_rope_params_to_dict(self, **kwargs): # No need to handle BC for new models, because they have no old-format `rope_scaling` return kwargs @auto_docstring(checkpoint="google/embeddinggemma-2") @strict class EmbeddingGemma2Config(PreTrainedConfig): r""" text_config (`EmbeddingGemma2TextConfig`, *optional*): Configuration of the text backbone. vision_config (`PreTrainedConfig` or `dict`, *optional*): Configuration of the vision tower. Reused verbatim from Gemma 4; the tower itself is resolved at runtime through `AutoModel`. audio_config (`PreTrainedConfig` or `dict`, *optional*): Configuration of the audio tower. Reused verbatim from Gemma 4; the tower itself is resolved at runtime through `AutoModel`. boi_token_id (`int`, *optional*, defaults to 255999): The begin-of-image token index to wrap the image prompt. eoi_token_id (`int`, *optional*, defaults to 258882): The end-of-image token index to wrap the image prompt. boa_token_id (`int`, *optional*, defaults to 256000): The begin-of-audio token index to wrap the audio prompt. eoa_token_index (`int`, *optional*, defaults to 258883): The end-of-audio token index to wrap the audio prompt. """ model_type = "embedding_gemma2" sub_configs_defaults = { "vision_config": SubConfigSpec(config_class=AutoConfig, model_type="gemma4_vision", optional=True), "text_config": SubConfigSpec(config_class=EmbeddingGemma2TextConfig), "audio_config": SubConfigSpec(config_class=AutoConfig, model_type="gemma4_audio", optional=True), } text_config: EmbeddingGemma2TextConfig | dict[str, Any] | None = None vision_config: PreTrainedConfig | dict[str, Any] | None = None audio_config: PreTrainedConfig | dict[str, Any] | None = None boi_token_id: int | None = 255_999 eoi_token_id: int | None = 258_882 image_token_id: int | None = 258_880 video_token_id: int | None = 258_884 boa_token_id: int | None = 256_000 eoa_token_index: int | None = 258_883 audio_token_id: int | None = 258_881 initializer_range: float | None = 0.02 __all__ = ["EmbeddingGemma2Config", "EmbeddingGemma2TextConfig"]