Struct PaddedCausalBatch 

pub struct PaddedCausalBatch<B: Backend> {
    pub input_ids: Tensor<B, 2, Int>,
    pub labels: Tensor<B, 2, Int>,
    pub attention_mask: Tensor<B, 2, Bool>,
    pub position_ids: Tensor<B, 2, Int>,
    pub lengths: Vec<usize>,
    pub supervised_tokens: usize,
    pub ignore_index: i64,
    pub target_alignment: CausalTargetAlignment,
}

Available on crate feature std only.

Rectangular actual documents with explicitly masked batch-padding cells.

Fields

input_ids: Tensor<B, 2, Int>

Token IDs shaped [actual examples, maximum actual sequence length].

labels: Tensor<B, 2, Int>

Same shape as input_ids; all padding cells contain ignore_index.

attention_mask: Tensor<B, 2, Bool>

True exactly at real source tokens, never at padding cells.

position_ids: Tensor<B, 2, Int>

Real positions reset per document; masked padding positions are zero.

lengths: Vec<usize>

Original encoded lengths in actual batch order.

supervised_tokens: usize

Actual nonignored supervised targets under the selected alignment.

ignore_index: i64

Actual label sentinel used during collation.

target_alignment: CausalTargetAlignment

Target alignment recorded for the downstream loss contract.

Implementations

impl<B: Backend> PaddedCausalBatch<B>

pub fn examples(&self) -> usize

Actual source example count; no synthetic rows are created.

pub fn to_device(self, device: &B::Device) -> Self

Move the actual tensors while retaining the original length metadata.

impl<B: Backend> PaddedCausalBatch<B>

pub fn token_inputs(&self) -> PaddedTokenBatch<B>

Label-free input view; cloning tensor handles does not duplicate token values.

Trait Implementations

impl<B: Clone + Backend> Clone for PaddedCausalBatch<B>

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<B: Debug + Backend> Debug for PaddedCausalBatch<B>

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

Formats the value using the given formatter.