Skip to content
KernelIndex
Search⌘K

KV shared attention with dual RoPE

23 eligible runs
attention

KV-shared attention with dual RoPE (global and local), Q/K/V normalization, soft-capping, and grouped-query attention. Implements attention where later layers can reuse KV states from earlier layers, with dual rotary position embeddings.

Fastest reported
174.6µsmean · 92% of SOL
GeometricLicense unknown · unknown

Reported evidence · last observed 2026-08-06. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencyhead_dim = 96 · q_proj_dim = 3072 · hidden_size = 3072 · kv_proj_dim = 768 · num_attention_heads = 32 · num_key_value_heads = 8 · num_key_value_groups = 416 cases

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 · 14 results · last observed 2026-08-20Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
174.6µs92% SOL
1.00×
Reported · license unknown · no source
2026-08-06

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

no sourcelicense unknownno install recipeRun detail →
2
193.1µs91% SOL
1.11×
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
222.3µs89% SOL
1.27×
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 →
4
225.2µs89% SOL
1.29×
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 →
5
268.8µs87% SOL
1.54×
Reported · license unknown · no source
2026-07-19

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

no sourcelicense unknownno install recipeRun detail →
6
270.2µs87% SOL
1.55×
Reported · license unknown · no source
2026-07-16

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

no sourcelicense unknownno install recipeRun detail →
7
299.3µs85% SOL
1.71×
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 →
8
300.4µs85% SOL
1.72×
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 →
9
301.7µs85% SOL
1.73×
Reported · license unknown · no source
2026-07-19

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

no sourcelicense unknownno install recipeRun detail →
10
314.3µs84% SOL
1.80×
Reported · license unknown · no source
2026-07-28

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

no sourcelicense unknownno install recipeRun detail →
11
322.7µs84% SOL
1.85×
Reported · license unknown · no source
2026-07-24

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

no sourcelicense unknownno install recipeRun detail →
12
337.5µs83% SOL
1.93×
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
445.7µs78% SOL
2.55×
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 →
14
1.16ms56% SOL
6.62×
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
unknown
193.1µs
1.11×
Reported
license unknown · no source
unknown
222.3µs
1.27×
Reported
license unknown · no source
unknown
268.8µs
1.54×
Reported
license unknown · no source
unknown
290.2µs
1.66×
Reported
license unknown · no source
unknown
337.5µs
1.93×
Reported
license unknown · no source
unknown
934.0µs
5.35×
Reported
license unknown · no source
unknown
1.16ms
6.62×
Reported
license unknown · no source
unknown
225.2µs
1.29×
Reported
license unknown · no source
Show all 23 implementations ›
unknown
301.7µs
1.73×
Reported
license unknown · no source
unknown
299.3µs
1.71×
Reported
license unknown · no source
unknown
300.4µs
1.72×
Reported
license unknown · no source
unknown
314.3µs
1.80×
Reported
license unknown · no source
unknown
322.7µs
1.85×
Reported
license unknown · no source
unknown
341.8µs
1.96×
Reported
license unknown · no source
unknown
437.3µs
2.50×
Reported
license unknown · no source
unknown
445.7µs
2.55×
Reported
license unknown · no source
unknown
749.5µs
4.29×
Reported
license unknown · no source
unknown
1.27ms
7.26×
Reported
license unknown · no source
unknown
1.29ms
7.37×
Reported
license unknown · no source
unknown
1.33ms
7.64×
Reported
license unknown · no source
unknown
174.6µs
1.00×
Reported
license unknown · no source
unknown
250.8µs
1.44×
Reported
license unknown · no source
unknown
270.2µs
1.55×
Reported
license unknown · no source

Semantics

Inputs and outputs
hidden_statesbf16 [batch_size, seq_len, hidden_size]
position_idsint64 [batch_size, seq_len]
attention_maskbf16 [batch_size, 1, seq_len, seq_len]
q_proj_weightbf16 [q_proj_dim, hidden_size]
k_proj_weightbf16 [kv_proj_dim, hidden_size]
v_proj_weightbf16 [kv_proj_dim, hidden_size]
o_proj_weightbf16 [hidden_size, q_proj_dim]
q_norm_weightbf16 [head_dim]
k_norm_weightbf16 [head_dim]
shared_key_statesbf16 [batch_size, num_key_value_heads, seq_len, head_dim]
shared_value_statesbf16 [batch_size, num_key_value_heads, seq_len, head_dim]
rope_thetafp32 scalar
softcapfp32 scalar
rms_norm_epsfp32 scalar
use_shared_kvbool scalar
attn_outputbf16 [batch_size, seq_len, hidden_size]
Axes and behavior
seq_lenvariable
head_dimconstant = 96
batch_sizevariable
q_proj_dimq_proj_dim = num_attention_heads * head_dim
hidden_sizeconstant = 3072
kv_proj_dimkv_proj_dim = num_key_value_heads * head_dim
num_attention_headsconstant = 32
num_key_value_headsconstant = 8
num_key_value_groupsconstant = 4
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias041_kv_shared_attention_with_dual_ropemodelgemma-3n-e4bsha256765ecc2da307…
Sources: NVIDIA SOL-ExecBench (2026-08-20)last observed 2026-08-20How records are decidedJSON