gm.nn.Gemma2_2B

class gemma.gm.nn.Gemma2_2B(

config: transformer.TransformerConfig = TransformerConfig(num_layers=26,

num_embed=256128,

embed_dim=2304,

hidden_dim=9216,

num_heads=8,

head_dim=256,

num_kv_heads=4,

final_logit_softcap=30.0,

use_post_attn_norm=True,

use_post_ffw_norm=True,

attention_types=(<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>,

<AttentionType.LOCAL_SLIDING: 2>,

<AttentionType.GLOBAL: 1>),

max_cache_length=None,

query_pre_attn_norm=<QueryPreAttentionNormalisation.BY_ONE_OVER_SQRT_HEAD_DIM: 1>,

attn_logits_soft_cap=50.0,

sliding_window_size=4096,

transpose_gating_einsum=False,

use_qk_norm=False,

local_base_frequency=10000,

global_base_frequency=10000,

local_scale_factor=1.0,

global_scale_factor=1.0,

mm_extra_vocab_size=0,

vision_encoder=None),

return_last_only: bool | None = None,

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

tokens: kontext.Key = '__KEY_REQUIRED__',

images: kontext.Key | None = None,

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

name: str | None = None,

)

基类： gemma.gm.nn._transformer.Transformer

Gemma2 Transformer 架构。

config: transformer.TransformerConfig = TransformerConfig(num_layers=26, num_embed=256128, embed_dim=2304, hidden_dim=9216, num_heads=8, head_dim=256, num_kv_heads=4, final_logit_softcap=30.0, use_post_attn_norm=True, use_post_ffw_norm=True, attention_types=(<AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>, <AttentionType.LOCAL_SLIDING: 2>, <AttentionType.GLOBAL: 1>), max_cache_length=None, query_pre_attn_norm=<QueryPreAttentionNormalisation.BY_ONE_OVER_SQRT_HEAD_DIM: 1>, attn_logits_soft_cap=50.0, sliding_window_size=4096, transpose_gating_einsum=False, use_qk_norm=False, local_base_frequency=10000, global_base_frequency=10000, local_scale_factor=1.0, global_scale_factor=1.0, mm_extra_vocab_size=0, vision_encoder=None)

INFO: ClassVar[ModelInfo] = ModelInfo(tokenizer_version=2, default_ckpt=<CheckpointPath.GEMMA2_2B_IT: 'gs://gemma-data/checkpoints/gemma2-2b-it'>)

name: str | None = None

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

scope: Scope | None = None