Skip to content
KernelIndex
Search⌘K

KV cache update with RoPE backward

19 eligible runs
attention

Backward pass for KV cache update with RoPE. Computes gradients for key_states, value_states, cos, sin, and input caches given gradients from downstream attention operations.

Fastest reported
11.2µsmean · 100% of SOL
doubleAILicense unknown · unknown

Reported evidence · last observed 2026-07-10. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencyhalf_dim = 64 · head_dim = 128 · num_kv_heads = 816 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 · 10 results · last observed 2026-08-04Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
11.2µs100% SOL
1.00×
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 →
2
11.3µs100% SOL
1.01×
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 →
3
12.7µs99% SOL
1.13×
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
18.6µs99% SOL
1.66×
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 →
5
20.5µs99% SOL
1.82×
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 →
6
25.9µs98% SOL
2.30×
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
25.9µs99% SOL
2.31×
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
31.3µs97% SOL
2.79×
Reported · license unknown · no source
2026-07-13

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

no sourcelicense unknownno install recipeRun detail →
9
33.9µs96% SOL
3.02×
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
68.0µs88% SOL
6.05×
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
12.1µs
1.09×
Reported
license unknown · no source
unknown
18.6µs
1.68×
Reported
license unknown · no source
unknown
25.9µs
2.34×
Reported
license unknown · no source
unknown
25.9µs
2.33×
Reported
license unknown · no source
unknown
31.3µs
2.82×
Reported
license unknown · no source
unknown
38.4µs
3.47×
Reported
license unknown · no source
unknown
66.7µs
6.01×
Reported
license unknown · no source
unknown
12.7µs
1.14×
Reported
license unknown · no source
Show all 19 implementations ›
unknown
20.5µs
1.84×
Reported
license unknown · no source
unknown
20.4µs
1.84×
Reported
license unknown · no source
unknown
26.1µs
2.36×
Reported
license unknown · no source
unknown
32.7µs
2.95×
Reported
license unknown · no source
unknown
33.9µs
3.05×
Reported
license unknown · no source
unknown
33.9µs
3.06×
Reported
license unknown · no source
unknown
68.0µs
6.13×
Reported
license unknown · no source
unknown
194.4µs
17.53×
Reported
license unknown · no source
unknown
11.1µs
1.00×
Reported
license unknown · no source
unknown
11.2µs
1.01×
Reported
license unknown · no source
unknown
11.3µs
1.02×
Reported
license unknown · no source

Semantics

Inputs and outputs
grad_key_cachebf16 [batch_size, num_kv_heads, max_seq_len, head_dim]
grad_value_cachebf16 [batch_size, num_kv_heads, max_seq_len, head_dim]
key_statesbf16 [batch_size, num_kv_heads, new_seq_len, head_dim]
cosbf16 [batch_size, new_seq_len, head_dim]
sinbf16 [batch_size, new_seq_len, head_dim]
cache_positionint64 [new_seq_len]
grad_key_statesbf16 [batch_size, num_kv_heads, new_seq_len, head_dim]
grad_value_statesbf16 [batch_size, num_kv_heads, new_seq_len, head_dim]
grad_cosbf16 [batch_size, new_seq_len, head_dim]
grad_sinbf16 [batch_size, new_seq_len, head_dim]
grad_key_cache_inputbf16 [batch_size, num_kv_heads, max_seq_len, head_dim]
grad_value_cache_inputbf16 [batch_size, num_kv_heads, max_seq_len, head_dim]
Axes and behavior
half_dimconstant = 64
head_dimconstant = 128
batch_sizevariable
max_seq_lenvariable
new_seq_lenvariable
num_kv_headsconstant = 8
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias062_kv_cache_update_with_rope_backwardmodeldevstral-small-2507sha256818e910899ea…
Sources: NVIDIA SOL-ExecBench (2026-08-04)last observed 2026-08-04How records are decidedJSON