Skip to content
KernelIndex
Search⌘K
Sign in

Multimodal rotary embedding attention

26 eligible runs
rope

Grouped Query Attention with Multi-modal 3D Rotary Position Embeddings. Supports 28 query heads and 4 KV heads (7:1 GQA ratio) with specialized 3D RoPE for temporal/height/width dimensions in vision tokens.

Fastest reported
166.2µsmean · 67% of SOL
RIAC as wellLicense unknown · unknown

Reported evidence · last observed 2026-08-23. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencyone = 1 · head_dim = 128 · num_heads = 28 · q_proj_dim = 3584 · rope_dim_0 = 3 · hidden_size = 3584 · kv_proj_dim = 512 · num_kv_heads = 4 · num_kv_groups = 716 cases
seq_len
batch_size
dtype
1281bf161284bf161288bf1612832bf162561bf162564bf1625632bf1625664bf165121bf165128bf1654132bf166131bf1610241bf1610242bf1610244bf1640961bf16
scrolls · 16 cases total
Aggregatessuite of 16 cases · mean latency
HardwareBest knownImplementationRuns
NVIDIA B200 · env 1166.2 µsRIAC as well18NVIDIA B200 · env 2170.1 µsRecursive8

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload suite of 16 cases · mean latency · Protocol SOL-ExecBench evaluation stack v1.1 · mean · 18 results · last observed 2026-08-23Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
RIAC as well
166.2µs67% SOL
1.00×
Reported · license unknown · no source
2026-08-23

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
2
Amir M. Mir | SF Tensor
174.0µs65% SOL
1.05×
Reported · license unknown · no source
2026-08-19

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
3
Recursive
177.9µs65% SOL
1.07×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
4
Geometric
181.3µs63% SOL
1.09×
Reported · license unknown · no source
2026-08-19

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
5
ac4k
191.1µs63% SOL
1.15×
Reported · license unknown · no source
2026-08-04

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
6
Geometric
193.4µs61% SOL
1.16×
Reported · license unknown · no source
2026-08-18

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
7
Geometric
197.1µs60% SOL
1.19×
Reported · license unknown · no source
2026-08-19

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
8
Hyra
200.7µs60% SOL
1.21×
Reported · license unknown · no source
2026-07-17

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
9
Amir M. Mir | SF Tensor
210.5µs59% SOL
1.27×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
10
Geometric
236.7µs54% SOL
1.42×
Reported · license unknown · no source
2026-07-20

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
11
Geometric
237.2µs54% SOL
1.43×
Reported · license unknown · no source
2026-07-21

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
12
Geometric
239.6µs53% SOL
1.44×
Reported · license unknown · no source
2026-07-21

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
13
Geometric
240.9µs53% SOL
1.45×
Reported · license unknown · no source
2026-07-21

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
14
Geometric
242.4µs53% SOL
1.46×
Reported · license unknown · no source
2026-07-20

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
15
Proud Wren @ NVIDIA
253.7µs53% SOL
1.53×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
16
Geometric
257.1µs51% SOL
1.55×
Reported · license unknown · no source
2026-07-20

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
17
doubleAI
355.4µs42% SOL
2.14×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
18
Cursor
568.3µs31% SOL
3.42×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
RIAC as well
unknown
166.2µs
1.00×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
174.0µs
1.05×
Reported
license unknown · no source
ac4k
unknown
191.1µs
1.15×
Reported
license unknown · no source
Geometric
unknown
181.3µs
1.09×
Reported
license unknown · no source
Hyra
unknown
200.7µs
1.21×
Reported
license unknown · no source
Proud Wren @ NVIDIA
unknown
207.7µs
1.25×
Reported
license unknown · no source
Cursor
unknown
507.3µs
3.05×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
197.7µs
1.19×
Reported
license unknown · no source
Show all 26 implementations ›
doubleAI
unknown
201.8µs
1.21×
Reported
license unknown · no source
Geometric
unknown
193.4µs
1.16×
Reported
license unknown · no source
Geometric
unknown
197.1µs
1.19×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
210.5µs
1.27×
Reported
license unknown · no source
Geometric
unknown
236.7µs
1.42×
Reported
license unknown · no source
Geometric
unknown
237.2µs
1.43×
Reported
license unknown · no source
Geometric
unknown
239.6µs
1.44×
Reported
license unknown · no source
Geometric
unknown
240.9µs
1.45×
Reported
license unknown · no source
Geometric
unknown
242.4µs
1.46×
Reported
license unknown · no source
Proud Wren @ NVIDIA
unknown
253.7µs
1.53×
Reported
license unknown · no source
Geometric
unknown
257.1µs
1.55×
Reported
license unknown · no source
doubleAI
unknown
355.4µs
2.14×
Reported
license unknown · no source
Cursor
unknown
568.3µs
3.42×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
585.2µs
3.52×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
825.4µs
4.97×
Reported
license unknown · no source
Recursive
unknown
170.1µs
1.02×
Reported
license unknown · no source
doubleAI
unknown
173.7µs
1.05×
Reported
license unknown · no source
Recursive
unknown
177.9µs
1.07×
Reported
license unknown · no source

Semantics

Inputs and outputs
hidden_statesbf16 [batch_size, seq_len, hidden_size]
q_weightbf16 [q_proj_dim, hidden_size]
q_biasbf16 [q_proj_dim]
k_weightbf16 [kv_proj_dim, hidden_size]
k_biasbf16 [kv_proj_dim]
v_weightbf16 [kv_proj_dim, hidden_size]
v_biasbf16 [kv_proj_dim]
o_weightbf16 [hidden_size, q_proj_dim]
cosbf16 [rope_dim_0, batch_size, seq_len, head_dim]
sinbf16 [rope_dim_0, batch_size, seq_len, head_dim]
attention_maskbf16 [batch_size, one, seq_len, seq_len]
outputbf16 [batch_size, seq_len, hidden_size]
Axes and behavior
oneconstant = 1
seq_lenvariable
head_dimconstant = 128
num_headsconstant = 28
batch_sizevariable
q_proj_dimq_proj_dim = num_heads * head_dim
rope_dim_0constant = 3
hidden_sizeconstant = 3584
kv_proj_dimkv_proj_dim = num_kv_heads * head_dim
num_kv_headsconstant = 4
num_kv_groupsconstant = 7
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias007_multimodal_rotary_embedding_attentionmodelqwen2-5-vl-7b-instructsha256484a3d62631c…
Sources: NVIDIA SOL-ExecBench (2026-08-23)last observed 2026-08-23How records are decidedJSON