Module bert 

BERT Model V2 - Clean implementation using solid abstractions

BERT is an encoder-only model with key differences from decoder-only models:

• Returns embeddings (ModelOutputs::Embeddings) instead of logits
• Bidirectional attention (NO causal masking)
• Uses position embeddings (learned) instead of RoPE
• Uses token type embeddings for segment distinction
• Uses LayerNorm (not RMSNorm)
• Has a pooler for [CLS] token representation

Structs

BertAttention

BERT attention: self-attention + output projection with residual

BertConfig

BertEmbeddings

BERT embeddings: word + position + token_type embeddings, then LayerNorm

BertEncoder

BERT encoder: stack of BertLayer

BertIntermediate

BERT intermediate (first FFN layer with activation)

BertLayer

BERT transformer layer

BertModelV2

Main BERT model implementation

BertOutput

BERT output (second FFN layer with residual and LayerNorm)

BertPooler

BERT pooler: takes [CLS] token and projects through dense + tanh

BertSelfAttention

BERT self-attention (bidirectional, no causal mask)

BertSelfOutput

BERT self-attention output projection