gm.nn.Gemma3_12B

class gemma.gm.nn.Gemma3_12B(

config: _config.TransformerConfig = TransformerConfig(num_embed=262144,

embed_dim=3840,

hidden_dim=15360,

num_heads=16,

head_dim=256,

num_kv_heads=8,

final_logit_softcap=None,

use_post_attn_norm=True,

use_post_ffw_norm=True,

attention_types=(<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>),

query_pre_attn_norm=<QueryPreAttentionNormalisation.BY_ONE_OVER_SQRT_HEAD_DIM: 1>,

attn_logits_soft_cap=None,

sliding_window_size=1024,

transpose_gating_einsum=True,

use_qk_norm=True,

local_base_frequency=10000,

global_base_frequency=1000000,

local_scale_factor=1.0,

global_scale_factor=8.0,

vision_encoder=SigLiPFromPatches(     # attributes     siglip_encoder = ViTModel(         # attributes         patch_size = (14,

14)         width=1152         depth = 27         mlp_dim = 4304         num_heads = 16         posemb = 'learn'         dropout = 0.0         scan = False         remat_policy = 'nothing_saveable'         dtype_mm = 'float32'     )     siglip_exit = VisionExit(         # attributes         output_length = 256     )     num_mm_tokens_per_image_prepool = 4096     num_mm_tokens_per_image = 256     image_height = 896     image_width = 896     image_channels = 3     apply_stop_gradient = True )),

text_only: bool = False,

parent: flax.linen.module.Module | flax.core.scope.Scope | flax.linen.module._Sentinel | None = <flax.linen.module._Sentinel object>,

name: str | None = None,

*,

return_last_only: bool | None = None,

dtype: jnp.dtype = <class 'jax.numpy.bfloat16'>,

tokens: kontext.Key = '__KEY_REQUIRED__',

images: kontext.Key | None = None,

positions: kontext.Key | None = None,

attention_mask: kontext.Key | None = None,

)

Bases: gemma.gm.nn._gemma._Gemma3Base

Gemma3 transformer architecture.

config: gemma.gm.nn._config.TransformerConfig = TransformerConfig(num_embed=262144, embed_dim=3840, hidden_dim=15360, num_heads=16, head_dim=256, num_kv_heads=8, final_logit_softcap=None, use_post_attn_norm=True, use_post_ffw_norm=True, attention_types=(<AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>), query_pre_attn_norm=<QueryPreAttentionNormalisation.BY_ONE_OVER_SQRT_HEAD_DIM: 1>, attn_logits_soft_cap=None, sliding_window_size=1024, transpose_gating_einsum=True, use_qk_norm=True, local_base_frequency=10000, global_base_frequency=1000000, local_scale_factor=1.0, global_scale_factor=8.0, vision_encoder=SigLiPFromPatches(     # attributes     siglip_encoder = ViTModel(         # attributes         patch_size = (14, 14)         width = 1152         depth = 27         mlp_dim = 4304         num_heads = 16         posemb = 'learn'         dropout = 0.0         scan = False         remat_policy = 'nothing_saveable'         dtype_mm = 'float32'     )     siglip_exit = VisionExit(         # attributes         output_length = 256     )     num_mm_tokens_per_image_prepool = 4096     num_mm_tokens_per_image = 256     image_height = 896     image_width = 896     image_channels = 3     apply_stop_gradient = True ))

INFO: ClassVar[gemma.gm.nn._transformer.ModelInfo] = ModelInfo(tokenizer_version=3, default_ckpt=None)

name: str | None = None

parent: flax.linen.module.Module | flax.core.scope.Scope | flax.linen.module._Sentinel | None = None

scope: flax.core.scope.Scope | None = None