LoGo: Token-Level Dynamic Local-Global Attention
Not assessedPlan blockedMeasurementclaim-query-sparse-speedup
The paper reports that its query-sparse Triton kernel reaches a 1.99x forward-plus-backward speedup over the dense Triton baseline at 64k sequence length and a 0.5 attention budget.
Source: paper-fixed:Section 4.3 and Figure 2, p. 8
Reported and observed measurements
speedup_over_dense_triton
m-figure2-logo-speedup-64k
Reported 1.99 ratio
Observed — ratio
Assessments (0)
No immutable Assessment has been published for this Claim yet.
Runs (0)
No execution has been linked to this Claim yet.