Struct VocabNetworkBuilder 

pub struct VocabNetworkBuilder<C> {
    pub vocab_size: usize,
    pub pad_vocab_size_multiple: usize,
    pub layers: LayersBuilder<C>,
    pub missing_lm_head: bool,
}

Plain factory for VocabNetwork. Mirrors LatentNetworkBuilder but adds vocab padding and the tied/untied LM-head choice.

Fields

vocab_size: usize

Unpadded vocabulary size (rounded up at init).

pad_vocab_size_multiple: usize

Round vocab_size up to a multiple of this (1 disables rounding).

layers: LayersBuilder<C>

Builder for the layer stack.

missing_lm_head: bool

When true, tie the LM head to the (transposed) embedding weights.

Implementations

impl<C: MambaBlockConfig> VocabNetworkBuilder<C>

fn padded_vocab(vocab_size: usize, multiple: usize) -> usize

Round vocab_size up to the next multiple of multiple.

pub fn init(&self, device: &Device) -> VocabNetwork<C::Block>

Allocate and initialise the network on device.