Region aware self attention with edit bias backward

32 eligible runs

diffusion

Backward pass for region-aware self-attention with edit bias. Computes gradients through QKV projection, scaled dot-product attention with position-based and edit-region biases, softmax, dropout, and output projection.

Fastest reported

295.5µsmean · 81% of SOL

RIAC as wellLicense unknown · unknown

Reported evidence · last observed 2026-08-17. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencyhead_dim = 64 · qkv_size = 3072 · hidden_size = 1024 · num_attention_heads = 16 · max_position_embeddings = 102416 cases

seq_len

batch_size

dtype

12864fp32 17616fp32 21116fp32 25664fp32 2884fp32 29332fp32 3522fp32 3731fp32 4488fp32 4492fp32 6404fp32 69116fp32 7682fp32 77332fp32 8531fp32 10241fp32

scrolls · 16 cases total

Aggregatessuite of 16 cases · mean latency

HardwareBest knownImplementationRuns

NVIDIA B200 · env 1295.5 µsRIAC as well24 NVIDIA B200 · env 2386.4 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 · 24 results · last observed 2026-08-24Record history →

#

Implementation

Latency

vs #1

Trust

Observed

1

295.5µs81% SOL

1.00×

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 →

2

Amir M. Mir | SF Tensor

335.3µs78% SOL

1.13×

Reported · license unknown · no source

2026-07-18

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

no sourcelicense unknownno install recipeRun detail →

3

393.3µs75% SOL

1.33×

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

407.2µs74% SOL

1.38×

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 →

5

413.2µs74% SOL

1.40×

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

444.4µs73% SOL

1.50×

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

469.1µs71% SOL

1.59×

Reported · license unknown · no source

2026-08-22

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

no sourcelicense unknownno install recipeRun detail →

8

483.7µs71% SOL

1.64×

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

502.3µs70% SOL

1.70×

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 →

10

506.8µs70% SOL

1.72×

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 →

11

516.9µs69% SOL

1.75×

Reported · license unknown · no source

2026-08-12

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

no sourcelicense unknownno install recipeRun detail →

12

572.3µs67% SOL

1.94×

Reported · license unknown · no source

2026-08-24

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

no sourcelicense unknownno install recipeRun detail →

13

600.8µs66% SOL

2.03×

Reported · license unknown · no source

2026-08-09

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

no sourcelicense unknownno install recipeRun detail →

14

Amir M. Mir | SF Tensor

647.5µs64% SOL

2.19×

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 →

15

Proud Wren @ NVIDIA

651.1µs64% SOL

2.20×

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 →

16

953.2µs54% SOL

3.23×

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 →

17

2.12ms34% SOL

7.19×

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 →

18

2.17ms34% SOL

7.33×

Reported · license unknown · no source

2026-07-25

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

no sourcelicense unknownno install recipeRun detail →

19

2.45ms31% SOL

8.31×

Reported · license unknown · no source

2026-07-26

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

no sourcelicense unknownno install recipeRun detail →

20

2.49ms31% SOL

8.42×

Reported · license unknown · no source

2026-07-26

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

no sourcelicense unknownno install recipeRun detail →

21

2.81ms28% SOL

9.52×

Reported · license unknown · no source

2026-07-23

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

no sourcelicense unknownno install recipeRun detail →

22

3.33ms25% SOL

11.28×

Reported · license unknown · no source

2026-08-10

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

no sourcelicense unknownno install recipeRun detail →

23

3.56ms24% SOL

12.04×

Reported · license unknown · no source

2026-08-09

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

no sourcelicense unknownno install recipeRun detail →

24

172.3ms1% SOL

582.99×

Reported · license unknown · no source

2026-08-08

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

no sourcelicense unknownno install recipeRun detail →

Implementations

Implementation

Runtime

Best latency

Evidence

Availability

Amir M. Mir | SF Tensor

unknown

335.3µs

1.13×

Reported

license unknown · no source

unknown

393.3µs

1.33×

Reported

license unknown · no source

unknown

407.2µs

1.38×

Reported

license unknown · no source

unknown

295.5µs

1.00×

Reported

license unknown · no source

unknown

386.4µs

1.31×

Reported

license unknown · no source

unknown

394.5µs

1.34×

Reported

license unknown · no source

unknown

444.4µs

1.50×

Reported

license unknown · no source

unknown

469.1µs

1.59×

Reported

license unknown · no source

Show all 32 implementations ›

Proud Wren @ NVIDIA

unknown

651.1µs

2.20×

Reported

license unknown · no source

unknown

962.8µs

3.26×

Reported

license unknown · no source

unknown

413.2µs

1.40×

Reported

license unknown · no source

unknown

483.7µs

1.64×

Reported

license unknown · no source

unknown

502.3µs

1.70×

Reported

license unknown · no source

unknown

506.8µs

1.72×

Reported

license unknown · no source

unknown

516.9µs

1.75×

Reported

license unknown · no source

unknown

563.0µs

1.91×

Reported

license unknown · no source

unknown

572.3µs

1.94×

Reported

license unknown · no source

unknown

600.8µs

2.03×

Reported

license unknown · no source

Amir M. Mir | SF Tensor

unknown

640.4µs

2.17×

Reported

license unknown · no source

Amir M. Mir | SF Tensor

unknown

647.5µs

2.19×

Reported

license unknown · no source

Amir M. Mir | SF Tensor

unknown

856.5µs

2.90×

Reported

license unknown · no source

unknown

953.2µs

3.23×

Reported

license unknown · no source

Amir M. Mir | SF Tensor

unknown

1.86ms

6.28×

Reported

license unknown · no source

unknown

2.12ms

7.19×

Reported

license unknown · no source

unknown

2.17ms

7.33×

Reported

license unknown · no source

unknown

2.45ms

8.31×

Reported

license unknown · no source

unknown

2.49ms

8.42×

Reported

license unknown · no source

Amir M. Mir | SF Tensor

unknown

2.62ms

8.87×

Reported

license unknown · no source

unknown

2.81ms

9.52×

Reported

license unknown · no source

unknown

3.33ms

11.28×

Reported

license unknown · no source

unknown

3.56ms

12.04×

Reported

license unknown · no source

unknown

172.3ms

582.99×

Reported

license unknown · no source

Semantics

Inputs and outputs

grad_outputfp32 [batch_size, seq_len, hidden_size]

hidden_statesfp32 [batch_size, seq_len, hidden_size]

edit_region_maskfp32 [batch_size, seq_len]

qkv_weightfp32 [qkv_size, hidden_size]

qkv_biasfp32 [qkv_size]

out_weightfp32 [hidden_size, hidden_size]

out_biasfp32 [hidden_size]

edit_region_biasfp32 [num_attention_heads, max_position_embeddings, max_position_embeddings]

within_edit_biasfp32 [num_attention_heads, 1, 1]

cross_edit_biasfp32 [num_attention_heads, 1, 1]

attention_maskfp32 [batch_size, seq_len]

queryfp32 [batch_size, num_attention_heads, seq_len, head_dim]

keyfp32 [batch_size, num_attention_heads, seq_len, head_dim]

valuefp32 [batch_size, num_attention_heads, seq_len, head_dim]

attention_scoresfp32 [batch_size, num_attention_heads, seq_len, seq_len]

attention_probsfp32 [batch_size, num_attention_heads, seq_len, seq_len]

attention_probs_droppedfp32 [batch_size, num_attention_heads, seq_len, seq_len]

attention_outputfp32 [batch_size, seq_len, hidden_size]

dropout_maskfp32 [batch_size, num_attention_heads, seq_len, seq_len]

scalefp32 scalar

dropout_pfp32 scalar

grad_hidden_statesfp32 [batch_size, seq_len, hidden_size]

grad_qkv_weightfp32 [qkv_size, hidden_size]

grad_qkv_biasfp32 [qkv_size]

grad_out_weightfp32 [hidden_size, hidden_size]

grad_out_biasfp32 [hidden_size]

grad_edit_region_biasfp32 [num_attention_heads, max_position_embeddings, max_position_embeddings]

grad_within_edit_biasfp32 [num_attention_heads, 1, 1]

grad_cross_edit_biasfp32 [num_attention_heads, 1, 1]

Axes and behavior

seq_lenvariable

head_dimconstant = 64

qkv_sizeqkv_size = 3 * hidden_size

batch_sizevariable

hidden_sizeconstant = 1024

num_attention_headsconstant = 16

max_position_embeddingsconstant = 1024

determinismunspecified

constraintsNo mutation or aliasing

Identity

alias072_region_aware_self_attention_with_edit_bias_backwardmodelstep1x-edit-v1p2sha2566dae081364bd…

Sources: NVIDIA SOL-ExecBench (2026-08-24)last observed 2026-08-24How records are decided JSON