Vision cu seqlens variable length attention
Vision encoder attention with variable-length sequence support using cumulative sequence lengths (cu_seqlens). Processes multiple images/videos with different resolutions in a single batch by using cu_seqlens to define sequence boundaries. Includes QKV projection, rotary position embeddings, scaled dot-product attention per sequence, and output projection.
Reported evidence · last observed 2026-07-15. Reported by source; not independently reproduced. No public source. License unknown.
Current records
Workloadsuite of 16 cases · mean latencyhead_dim = 80 · num_heads = 16 · hidden_size = 1280 · qkv_out_size = 384016 cases
Not measured on H100 for this workload. Challenges →
155.6µs91% SOL1.00×Reported · license unknown · no source2026-07-15
Measured exactly what you asked. Reported by source; not independently reproduced.
258.6µs90% SOL1.05×Reported · license unknown · no source2026-08-24
Measured exactly what you asked. Reported by source; not independently reproduced.
364.4µs89% SOL1.16×Reported · license unknown · no source2026-07-26
Measured exactly what you asked. Reported by source; not independently reproduced.
466.9µs89% SOL1.20×Reported · license unknown · no source2026-08-19
Measured exactly what you asked. Reported by source; not independently reproduced.
574.5µs87% SOL1.34×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
681.2µs86% SOL1.46×Reported · license unknown · no source2026-08-24
Measured exactly what you asked. Reported by source; not independently reproduced.
785.9µs86% SOL1.54×Reported · license unknown · no source2026-08-17
Measured exactly what you asked. Reported by source; not independently reproduced.
888.6µs85% SOL1.59×Reported · license unknown · no source2026-08-11
Measured exactly what you asked. Reported by source; not independently reproduced.
988.9µs85% SOL1.60×Reported · license unknown · no source2026-08-06
Measured exactly what you asked. Reported by source; not independently reproduced.
1089.5µs85% SOL1.61×Reported · license unknown · no source2026-08-22
Measured exactly what you asked. Reported by source; not independently reproduced.
1190.2µs85% SOL1.62×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
1293.3µs84% SOL1.68×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
1394.7µs84% SOL1.70×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
14100.5µs83% SOL1.81×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
15115.7µs81% SOL2.08×Reported · license unknown · no source2026-07-20
Measured exactly what you asked. Reported by source; not independently reproduced.
16167.1µs74% SOL3.00×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
17253.0µs65% SOL4.55×Reported · license unknown · no source2026-07-28
Measured exactly what you asked. Reported by source; not independently reproduced.
18360.7µs57% SOL6.49×Reported · license unknown · no source2026-07-25
Measured exactly what you asked. Reported by source; not independently reproduced.
19395.5µs54% SOL7.11×Reported · license unknown · no source2026-07-27
Measured exactly what you asked. Reported by source; not independently reproduced.
20399.4µs54% SOL7.18×Reported · license unknown · no source2026-07-25
Measured exactly what you asked. Reported by source; not independently reproduced.
21508.0µs48% SOL9.14×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.