Class RaggedContiguousAttention

java.lang.Object

smile.llm.attention.RaggedContiguousAttention

public final class RaggedContiguousAttention
extends Object

Ragged contiguous self-attention over packed [N, H, D] Q/K/V tensors.

Used when FlashInfer ragged is unavailable (torch_native backend or older libsmile_torch without smile_flashinfer_ragged_attention).

Method Summary

Modifier and Type	Method	Description
static Tensor	forward(Tensor query, Tensor key, Tensor value, AttentionContext ctx)	Runs non-causal (or causal) self-attention independently within each indptr segment.

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

forward

public static Tensor forward(Tensor query,
 Tensor key,
 Tensor value,
 AttentionContext ctx)

Runs non-causal (or causal) self-attention independently within each indptr segment.

Parameters:

query - [N, H, D] NHD

key - [N, H, D]

value - [N, H, D]

ctx - ragged context with AttentionContext.raggedIndptr()

Returns:

output [N, H, D]