Skip to content
KernelIndex
Search⌘K
Sign in

Vision language cross attention fusion

24 eligible runs
vision

Cross-attention fusion layer for Qwen3-VL multimodal processing. Performs cross-attention between language queries (with 1D RoPE) and vision key-values (with 3D RoPE for temporal-height-width positions). Uses Grouped Query Attention (GQA) with 32 query heads and 8 key-value heads.

Fastest reported
286.2µsmean · 65% of SOL
RIAC as wellLicense unknown · unknown

Reported evidence · last observed 2026-08-15. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencythw_dim = 3 · head_dim = 128 · q_proj_dim = 4096 · hidden_size = 4096 · kv_proj_dim = 1024 · num_attention_heads = 32 · num_key_value_heads = 816 cases
lang_seq_len
vision_seq_len
batch_size
dtype
1281962bf161281968bf1612819616bf162563922bf162563928bf1625639264bf1637358832bf165127841bf165127842bf165127844bf1651278416bf165419804bf16102415681bf16102415682bf16409662721bf168192125441bf16
scrolls · 16 cases total
Aggregatessuite of 16 cases · mean latency
HardwareBest knownImplementationRuns
NVIDIA B200 · env 1286.2 µsRIAC as well16NVIDIA B200 · env 2290.3 µsAmir M. Mir | SF Tensor8

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 · 16 results · last observed 2026-08-20Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
RIAC as well
286.2µs65% SOL
1.00×
Reported · license unknown · no source
2026-08-15

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

no sourcelicense unknownno install recipeRun detail →
2
Amir M. Mir | SF Tensor
303.9µs63% SOL
1.06×
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 →
3
Recursive
305.5µs63% 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
Hyra
309.4µs62% SOL
1.08×
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 →
5
ac4k
327.2µs60% SOL
1.14×
Reported · license unknown · no source
2026-08-05

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

no sourcelicense unknownno install recipeRun detail →
6
ac4k
330.8µs59% SOL
1.16×
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 →
7
Geometric
331.4µs60% SOL
1.16×
Reported · license unknown · no source
2026-08-20

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

no sourcelicense unknownno install recipeRun detail →
8
Geometric
331.8µs60% 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 →
9
Proud Wren @ NVIDIA
333.2µs59% SOL
1.16×
Reported · license unknown · no source
2026-07-11

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

no sourcelicense unknownno install recipeRun detail →
10
Geometric
339.8µs58% SOL
1.19×
Reported · license unknown · no source
2026-08-17

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

no sourcelicense unknownno install recipeRun detail →
11
Geometric
340.6µs58% SOL
1.19×
Reported · license unknown · no source
2026-08-16

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

no sourcelicense unknownno install recipeRun detail →
12
doubleAI
405.6µs52% SOL
1.42×
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 →
13
Geometric
430.3µs51% SOL
1.50×
Reported · license unknown · no source
2026-08-20

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

no sourcelicense unknownno install recipeRun detail →
14
Geometric
431.3µs50% SOL
1.51×
Reported · license unknown · no source
2026-08-13

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

no sourcelicense unknownno install recipeRun detail →
15
Geometric
450.5µs49% SOL
1.57×
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 →
16
Cursor
467.7µs46% SOL
1.63×
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
286.2µs
1.00×
Reported
license unknown · no source
doubleAI
unknown
309.6µs
1.08×
Reported
license unknown · no source
Hyra
unknown
309.4µs
1.08×
Reported
license unknown · no source
Geometric
unknown
331.4µs
1.16×
Reported
license unknown · no source
ac4k
unknown
327.2µs
1.14×
Reported
license unknown · no source
Proud Wren @ NVIDIA
unknown
333.2µs
1.16×
Reported
license unknown · no source
Cursor
unknown
409.8µs
1.43×
Reported
license unknown · no source
Recursive
unknown
305.5µs
1.07×
Reported
license unknown · no source
Show all 24 implementations ›
doubleAI
unknown
310.6µs
1.09×
Reported
license unknown · no source
Geometric
unknown
331.8µs
1.16×
Reported
license unknown · no source
ac4k
unknown
330.8µs
1.16×
Reported
license unknown · no source
Geometric
unknown
339.8µs
1.19×
Reported
license unknown · no source
Geometric
unknown
340.6µs
1.19×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
354.6µs
1.24×
Reported
license unknown · no source
doubleAI
unknown
405.6µs
1.42×
Reported
license unknown · no source
Geometric
unknown
430.3µs
1.50×
Reported
license unknown · no source
Geometric
unknown
431.3µs
1.51×
Reported
license unknown · no source
Geometric
unknown
450.5µs
1.57×
Reported
license unknown · no source
Cursor
unknown
467.7µs
1.63×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
780.4µs
2.73×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
941.0µs
3.29×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
290.3µs
1.01×
Reported
license unknown · no source
Recursive
unknown
293.8µs
1.03×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
303.9µs
1.06×
Reported
license unknown · no source

Semantics

Inputs and outputs
language_hidden_statesbf16 [batch_size, lang_seq_len, hidden_size]
vision_hidden_statesbf16 [batch_size, vision_seq_len, hidden_size]
language_position_idsint64 [batch_size, lang_seq_len]
vision_grid_thwint64 [batch_size, vision_seq_len, thw_dim]
q_proj_weightbf16 [q_proj_dim, hidden_size]
q_proj_biasbf16 [q_proj_dim]
k_proj_weightbf16 [kv_proj_dim, hidden_size]
k_proj_biasbf16 [kv_proj_dim]
v_proj_weightbf16 [kv_proj_dim, hidden_size]
v_proj_biasbf16 [kv_proj_dim]
o_proj_weightbf16 [hidden_size, q_proj_dim]
outputbf16 [batch_size, lang_seq_len, hidden_size]
Axes and behavior
thw_dimconstant = 3
head_dimconstant = 128
batch_sizevariable
q_proj_dimq_proj_dim = num_attention_heads * head_dim
hidden_sizeconstant = 4096
kv_proj_dimkv_proj_dim = num_key_value_heads * head_dim
lang_seq_lenvariable
vision_seq_lenvariable
num_attention_headsconstant = 32
num_key_value_headsconstant = 8
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias034_vision_language_cross_attention_fusionmodelqwen3-vl-32b-instruct-nvfp4sha2563bf4a6eebf79…
Sources: NVIDIA SOL-ExecBench (2026-08-20)last observed 2026-08-20How records are decidedJSON