Struct BertConfig 

pub struct BertConfig {
    pub hidden_dim: usize,
    pub num_layers: usize,
    pub num_heads: usize,
    pub intermediate_dim: usize,
    pub vocab_size: usize,
    pub max_position_embeddings: usize,
    pub type_vocab_size: usize,
    pub layer_norm_eps: f32,
    pub pad_token_id: u32,
}

BERT model configuration.

Maps to the relevant fields of HuggingFace BertConfig.

Fields

hidden_dim: usize

Hidden dimension of each token (e.g. 768 for base, 384 for MiniLM-L-6).

num_layers: usize

Number of encoder layers (e.g. 12 for base, 6 for MiniLM-L-6).

num_heads: usize

Number of attention heads. hidden_dim must divide evenly by this.

intermediate_dim: usize

FFN intermediate dimension (typically 4 × hidden_dim).

vocab_size: usize

Token vocabulary size (e.g. 30522 for bert-base-uncased).

max_position_embeddings: usize

Maximum sequence length supported by position embeddings.

type_vocab_size: usize

Token-type vocabulary size (typically 2 for [A] / [B] segments).

layer_norm_eps: f32

LayerNorm epsilon (HuggingFace default 1e-12).

pad_token_id: u32

Pad token id (typically 0). Used for attention masking.

Implementations

impl BertConfig

pub const fn head_dim(&self) -> usize

Compute the per-head dimension. hidden_dim must be divisible by num_heads.

pub fn minilm_l6() -> BertConfig

MiniLM-L-6 preset (22M params, 384 hidden, 12 heads, 6 layers).

pub fn validate(&self) -> Result<(), BertConfigError>

Check the structural invariants that model construction asserts.

MultiHeadAttention::new (reached via BertEncoder::new / CrossEncoder::new) asserts that hidden_dim is a multiple of num_heads, and head_dim() divides by num_heads. Callers that build a config from untrusted input (CLI overrides, config files) must call this first and surface the error, instead of letting the assert abort the process.

Errors

Returns BertConfigError naming the offending field when hidden_dim or num_heads is zero, or when hidden_dim is not divisible by num_heads.

pub fn validate_ids( &self, input_ids: &[u32], token_type_ids: &[u32], ) -> Result<(), BertConfigError>

Check a token-id batch against this config before a forward pass.

BertEmbeddings::forward indexes the word/position/token-type tables with the raw ids and slices them unchecked, so an id at or above the corresponding table size aborts the process with a slice-range panic. This is the fallible pre-check for any caller feeding user-supplied ids.

Errors

Returns BertConfigError for a length mismatch, a sequence longer than max_position_embeddings, an input id >= vocab_size, or a token-type id >= type_vocab_size.

Trait Implementations

impl Clone for BertConfig

fn clone(&self) -> BertConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for BertConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Default for BertConfig

fn default() -> BertConfig

bert-base-uncased defaults.

impl PartialEq for BertConfig

fn eq(&self, other: &BertConfig) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for BertConfig