Multimodal RoPE position computation with grid based indexing
Computes 3D rotary position embeddings for multimodal inputs by generating position indices from grid_thw (temporal, height, width) parameters, performing bilinear interpolation for variable resolution support, and creating interleaved MRoPE embeddings. Combines grid-based position index generation, learned position embedding lookup with bilinear interpolation weights, spatial permutation to merge_size blocks, and rotary embedding frequency computation with interleaving.
Reported evidence · last observed 2026-06-25. Reported by source; not independently reproduced. No public source. License unknown.
Current records
Workloadsuite of 16 cases · mean latencyhead_dim = 128 · num_heads = 12 · hidden_size = 1536 · quarter_head_dim = 32 · num_grid_per_side = 35 · spatial_merge_size = 2 · num_position_embeddings = 122516 cases
Not measured on H100 for this workload. Challenges →
14.00µs99% SOL1.00×Reported · license unknown · no source2026-06-25
Measured exactly what you asked. Reported by source; not independently reproduced.
24.13µs99% SOL1.03×Reported · license unknown · no source2026-07-06
Measured exactly what you asked. Reported by source; not independently reproduced.
35.32µs98% SOL1.33×Reported · license unknown · no source2026-07-01
Measured exactly what you asked. Reported by source; not independently reproduced.
45.86µs98% SOL1.46×Reported · license unknown · no source2026-06-18
Measured exactly what you asked. Reported by source; not independently reproduced.
55.90µs98% SOL1.47×Reported · license unknown · no source2026-06-20
Measured exactly what you asked. Reported by source; not independently reproduced.
66.20µs98% SOL1.55×Reported · license unknown · no source2026-04-15
Measured exactly what you asked. Reported by source; not independently reproduced.
76.49µs98% SOL1.62×Reported · license unknown · no source2026-07-05
Measured exactly what you asked. Reported by source; not independently reproduced.
86.66µs98% SOL1.66×Reported · license unknown · no source2026-06-06
Measured exactly what you asked. Reported by source; not independently reproduced.
97.32µs97% SOL1.83×Reported · license unknown · no source2026-06-14
Measured exactly what you asked. Reported by source; not independently reproduced.
107.56µs97% SOL1.89×Reported · license unknown · no source2026-05-20
Measured exactly what you asked. Reported by source; not independently reproduced.
118.60µs96% SOL2.15×Reported · license unknown · no source2026-07-08
Measured exactly what you asked. Reported by source; not independently reproduced.
1210.6µs96% SOL2.66×Reported · license unknown · no source2026-04-14
Measured exactly what you asked. Reported by source; not independently reproduced.
1316.4µs93% SOL4.10×Reported · license unknown · no source2026-07-09
Measured exactly what you asked. Reported by source; not independently reproduced.
1418.4µs93% SOL4.60×Reported · license unknown · no source2026-07-08
Measured exactly what you asked. Reported by source; not independently reproduced.
1518.9µs92% SOL4.73×Reported · license unknown · no source2026-07-08
Measured exactly what you asked. Reported by source; not independently reproduced.
1621.1µs91% SOL5.28×Reported · license unknown · no source2026-07-07
Measured exactly what you asked. Reported by source; not independently reproduced.
1730.5µs88% SOL7.62×Reported · license unknown · no source2026-05-18
Measured exactly what you asked. Reported by source; not independently reproduced.
1836.5µs86% SOL9.13×Reported · license unknown · no source2026-04-27
Measured exactly what you asked. Reported by source; not independently reproduced.
1985.0µs72% SOL21.23×Reported · license unknown · no source2026-04-25
Measured exactly what you asked. Reported by source; not independently reproduced.
20724.1µs25% SOL180.92×Reported · license unknown · no source2026-04-22
Measured exactly what you asked. Reported by source; not independently reproduced.
211.56ms14% SOL389.95×Reported · license unknown · no source2026-05-28
Measured exactly what you asked. Reported by source; not independently reproduced.