Module class 

Learnable [CLS]-style class tokens/latents spliced into the sequence.

Macros

impl_class_marker 🔒

Enums

ClassLatent

Position marker for a learnable class latent inserted into a layer’s working sequence (embedding width = d_model).

ClassLatentSerde 🔒

ClassToken

Position marker for a learnable class token inserted into a network’s input sequence (embedding width = the network input width / “d_input”).

ClassTokenSerde 🔒

Traits

ClassMarker

Shared behaviour of the ClassToken / ClassLatent position markers, letting one generic helper place either kind.

Functions

assert_step_compatible 🔒

Panic if any marker is incompatible with single-token step().

class_marker_output_indices 🔒

The output-sequence position of each marker (in Vec order) for an input of length orig_len, without materialising any tensor. Mirrors the placement in insert_class_markers — useful for reading a class token back out.

class_step_injections 🔒

The output-sequence position of each step-injectable marker (in Vec order), for use by step’s cursor. Asserts no Middle/End (those need the full length — forward only). Start/Custom positions are length-independent, so an unbounded orig_len resolves them exactly.

init_class_emb 🔒

Build the embedding param for n class markers of the given width (None when there are none — Burn has no zero-width tensors).

insert_class_markers 🔒

Insert the learnable class tokens emb ([k, width], row i ↔ markers[i]) into x ([batch, orig_len, width]) per the markers, returning the lengthened sequence ([batch, orig_len + k, width]) and, for each marker in Vec order, its position in the output sequence.