Struct BertEncoder 

pub struct BertEncoder {
    pub hp: BertHparams,
    pub tok_embd: WeightMatrix,
    pub type_embd: Option<Vec<Vec<f32>>>,
    pub pos_embd: WeightMatrix,
    pub tok_norm_w: Vec<f32>,
    pub tok_norm_b: Vec<f32>,
    pub layers: Vec<BertLayer>,
}

Fields

hp: BertHparams

tok_embd: WeightMatrix

type_embd: Option<Vec<Vec<f32>>>

token_types.weight, every row: [n_token_types, n_embd]. Row 0 is “Sentence A” and row 1 “Sentence B”. None when the checkpoint carries no table at all, which upstream allows (TENSOR_NOT_REQUIRED) and which means no segment embedding is added anywhere. Loading only row 0 — what this held before — is what made a rerank pair score both halves as Sentence A.

pos_embd: WeightMatrix

tok_norm_w: Vec<f32>

tok_norm_b: Vec<f32>

layers: Vec<BertLayer>

Implementations

impl BertEncoder

pub fn vocab_size(&self) -> usize

Trait Implementations

impl TextEncoder for BertEncoder

fn wrap_special(&self, pieces: &[u32]) -> Vec<u32>

[CLS] … [SEP], which is what llama.cpp’s WPM branch adds when add_special is set: it pushes special_bos_id before the pieces and special_sep_id after, unconditionally — the add_bos/add_eos flags are not consulted on that path (llama-vocab.cpp, case LLAMA_VOCAB_TYPE_WPM).

fn n_segments(&self) -> usize

The height of token_types.weight, or 1 when the checkpoint carries no table (nothing is added at any position, which is what a one-row table would do anyway).

fn wrap_special_pair(&self, a: &[u32], b: &[u32]) -> Option<PairSequence>

[CLS] a [SEP] b [SEP] with segments 0…0 1…1 — what HuggingFace’s tokenizer(query, document) builds for a BERT cross-encoder, which is the input these checkpoints were trained on.

The boundary is defined once, here, and both vectors are cut on it: the first [SEP] closes segment 0 (HF counts it as part of the first half) and everything after it is segment 1. Returning the ids alone and letting the graph assume a segment is how the document half came to be scored as “Sentence A”.

fn n_embd(&self) -> usize

Width of one hidden-state row, and of the pooled embedding.

fn n_ctx_train(&self) -> usize

Longest sequence this checkpoint can represent. For a learned position table this is the table’s height, and exceeding it is an error rather than a degradation.

fn pooling_type(&self) -> PoolingType

What the checkpoint’s own {arch}.pooling_type said.

fn encode( &self, tokens: &[u32], segments: Option<&[u32]>, ) -> Result<Vec<f32>, EncodeError>

n_tokens × n_embd hidden states, in row order.

fn encode_tokens(&self, tokens: &[u32]) -> Result<Vec<f32>, EncodeError>

Self::encode for a single sequence: every position is segment 0.

fn embed_tokens(&self, tokens: &[u32]) -> Result<Vec<f32>, EncodeError>

Self::encode_tokens followed by the checkpoint’s own pooling. Not L2-normalized — see crate::pooling::pool.