Skip to content
KernelIndex
Search⌘K
Sign in

Fused RoPE application backward

24 eligible runs
rope

Backward pass for fused rotary position embedding (RoPE) application. Computes gradients w.r.t. query, key, cosine, and sine tensors. The gradient computation mirrors the forward pass structure: grad_x = (grad_output * cos) + (rotate_half(grad_output) * sin).

Fastest reported
17.7µsmean · 82% of SOL
RIAC as wellLicense unknown · unknown

Reported evidence · last observed 2026-08-19. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencyhead_dim = 128 · half_head_dim = 6416 cases
seq_len
batch_size
num_heads
num_kv_heads
dtype
1282328fp3212816328fp3212832328fp322561328fp322931328fp325121648fp325122648fp325412328fp329971328fp32102413232fp3210241648fp3210242328fp3215714328fp3220481648fp3220482328fp3220531328fp32
scrolls · 16 cases total
Aggregatessuite of 16 cases · mean latency
HardwareBest knownImplementationRuns
NVIDIA B200 · env 117.7 µsRIAC as well14NVIDIA B200 · env 218.4 µsRecursive10

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 2026-08-19Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
RIAC as well
17.7µs82% SOL
1.00×
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 →
2
Recursive
18.5µs81% SOL
1.04×
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
Amir M. Mir | SF Tensor
19.2µs79% SOL
1.08×
Reported · license unknown · no source
2026-08-03

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

no sourcelicense unknownno install recipeRun detail →
4
doubleAI
19.3µs79% SOL
1.09×
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 →
5
Amir M. Mir | SF Tensor
19.3µs79% SOL
1.09×
Reported · license unknown · no source
2026-07-31

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

no sourcelicense unknownno install recipeRun detail →
6
ac4k
19.5µs78% SOL
1.10×
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 →
7
Databricks
19.6µs78% SOL
1.11×
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
Hyra
19.7µs78% SOL
1.11×
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 →
9
Geometric
20.0µs78% SOL
1.13×
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 →
10
Amir M. Mir | SF Tensor
20.8µs76% SOL
1.18×
Reported · license unknown · no source
2026-07-12

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

no sourcelicense unknownno install recipeRun detail →
11
Amir M. Mir | SF Tensor
21.4µs74% SOL
1.21×
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 →
12
Abhinav Gorrepati
21.5µs75% SOL
1.22×
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 →
13
Cursor
22.3µs74% SOL
1.26×
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
TH3CHARLie | TileGym
25.6µs69% SOL
1.44×
Reported · license unknown · no source
2026-07-27

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
17.7µs
1.00×
Reported
license unknown · no source
Argus
unknown
25.7µs
1.45×
Reported
license unknown · no source
doubleAI
unknown
19.3µs
1.09×
Reported
license unknown · no source
doubleAI
unknown
18.5µs
1.04×
Reported
license unknown · no source
Recursive
unknown
18.4µs
1.04×
Reported
license unknown · no source
Recursive
unknown
18.5µs
1.04×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
19.2µs
1.08×
Reported
license unknown · no source
ac4k
unknown
19.4µs
1.09×
Reported
license unknown · no source
Show all 24 implementations ›
Databricks
unknown
19.5µs
1.10×
Reported
license unknown · no source
Hyra
unknown
19.7µs
1.11×
Reported
license unknown · no source
Geometric
unknown
20.0µs
1.13×
Reported
license unknown · no source
Abhinav Gorrepati
unknown
21.5µs
1.22×
Reported
license unknown · no source
Cursor
unknown
22.3µs
1.26×
Reported
license unknown · no source
TH3CHARLie | TileGym
unknown
25.6µs
1.44×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
19.3µs
1.09×
Reported
license unknown · no source
ac4k
unknown
19.5µs
1.10×
Reported
license unknown · no source
Databricks
unknown
19.6µs
1.11×
Reported
license unknown · no source
doubleAI
unknown
20.1µs
1.13×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
20.8µs
1.18×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
21.3µs
1.21×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
21.4µs
1.21×
Reported
license unknown · no source
Cursor
unknown
22.4µs
1.26×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
46.0µs
2.60×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
243.3µs
13.74×
Reported
license unknown · no source

Semantics

Inputs and outputs
grad_q_embedfp32 [batch_size, num_heads, seq_len, head_dim]
grad_k_embedfp32 [batch_size, num_kv_heads, seq_len, head_dim]
qfp32 [batch_size, num_heads, seq_len, head_dim]
kfp32 [batch_size, num_kv_heads, seq_len, head_dim]
cosfp32 [batch_size, seq_len, head_dim]
sinfp32 [batch_size, seq_len, head_dim]
grad_qfp32 [batch_size, num_heads, seq_len, head_dim]
grad_kfp32 [batch_size, num_kv_heads, seq_len, head_dim]
grad_cosfp32 [batch_size, seq_len, head_dim]
grad_sinfp32 [batch_size, seq_len, head_dim]
Axes and behavior
seq_lenvariable
head_dimconstant = 128
num_headsvariable
batch_sizevariable
num_kv_headsvariable
half_head_dimconstant = 64
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias068_fused_rope_application_backwardmodelllama-3-1-nemotron-8b-ultralong-1m-instructsha25695f9dc75f47a…
Sources: NVIDIA SOL-ExecBench (2026-08-19)last observed 2026-08-19How records are decidedJSON