Skip to content
KernelIndex
Search⌘K
Sign in

Cross attention QKV projection with GQA repeat backward

26 eligible runs
other

Backward pass for cross-attention QKV projection with GQA repeat. Computes gradients for decoder_hidden_states, encoder_hidden_states, q_weight, k_weight, and v_weight given upstream gradients from query_states, key_states, and value_states.

Fastest reported
151.1µsmean · 70% of SOL
RIAC as wellLicense unknown · unknown

Reported evidence · last observed 2026-08-21. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencyhead_dim = 256 · num_heads = 16 · q_out_dim = 4096 · kv_out_dim = 1024 · hidden_size = 1536 · cross_hidden_size = 1024 · num_key_value_heads = 4 · num_key_value_groups = 416 cases
seq_len_dec
seq_len_enc
batch_size
dtype
1281282fp321281288fp3212812832fp3212825616fp3221137316fp322562562fp3225625664fp322565121fp3225651216fp3244954132fp325125122fp326136911fp329199978fp32102420481fp32204840964fp32409681921fp32
scrolls · 16 cases total
Aggregatessuite of 16 cases · mean latency
HardwareBest knownImplementationRuns
NVIDIA B200 · env 1151.1 µsRIAC as well16NVIDIA B200 · env 2149.5 µsac4k10

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 2026-08-22Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
RIAC as well
151.1µs70% SOL
1.00×
Reported · license unknown · no source
2026-08-21

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

no sourcelicense unknownno install recipeRun detail →
2
Recursive
152.9µs69% SOL
1.01×
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
ac4k
159.1µs68% SOL
1.05×
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 →
4
Databricks
168.0µs67% 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 →
5
Geometric
168.7µs67% SOL
1.12×
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 →
6
Hyra
168.9µs66% SOL
1.12×
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
doubleAI
174.2µs65% SOL
1.15×
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
Geometric
177.2µs65% SOL
1.17×
Reported · license unknown · no source
2026-08-21

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

no sourcelicense unknownno install recipeRun detail →
9
TH3CHARLie | TileGym
181.2µs64% SOL
1.20×
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 →
10
Amir M. Mir | SF Tensor
183.1µs64% SOL
1.21×
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 →
11
Geometric
184.3µs64% SOL
1.22×
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 →
12
ac4k
189.0µs64% SOL
1.25×
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
Amir M. Mir | SF Tensor
195.9µs62% SOL
1.30×
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
Geometric
200.2µs62% SOL
1.33×
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 →
15
Geometric
362.6µs43% SOL
2.40×
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 →
16
Cursor
386.0µs41% 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 →

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
RIAC as well
unknown
151.1µs
1.01×
Reported
license unknown · no source
doubleAI
unknown
158.3µs
1.06×
Reported
license unknown · no source
Databricks
unknown
161.3µs
1.08×
Reported
license unknown · no source
Geometric
unknown
168.7µs
1.13×
Reported
license unknown · no source
Hyra
unknown
168.9µs
1.13×
Reported
license unknown · no source
TH3CHARLie | TileGym
unknown
181.2µs
1.21×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
183.1µs
1.22×
Reported
license unknown · no source
Argus
unknown
288.0µs
1.93×
Reported
license unknown · no source
Show all 26 implementations ›
Cursor
unknown
377.8µs
2.53×
Reported
license unknown · no source
ac4k
unknown
159.1µs
1.06×
Reported
license unknown · no source
doubleAI
unknown
163.5µs
1.09×
Reported
license unknown · no source
Databricks
unknown
168.0µs
1.12×
Reported
license unknown · no source
doubleAI
unknown
174.2µs
1.17×
Reported
license unknown · no source
Geometric
unknown
177.2µs
1.19×
Reported
license unknown · no source
Geometric
unknown
184.3µs
1.23×
Reported
license unknown · no source
ac4k
unknown
189.0µs
1.26×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
189.4µs
1.27×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
195.9µs
1.31×
Reported
license unknown · no source
Geometric
unknown
200.2µs
1.34×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
273.5µs
1.83×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
321.2µs
2.15×
Reported
license unknown · no source
Geometric
unknown
362.6µs
2.43×
Reported
license unknown · no source
Cursor
unknown
386.0µs
2.58×
Reported
license unknown · no source
ac4k
unknown
149.5µs
1.00×
Reported
license unknown · no source
Recursive
unknown
152.0µs
1.02×
Reported
license unknown · no source
Recursive
unknown
152.9µs
1.02×
Reported
license unknown · no source

Semantics

Inputs and outputs
grad_query_statesfp32 [batch_size, num_heads, seq_len_dec, head_dim]
grad_key_statesfp32 [batch_size, num_heads, seq_len_enc, head_dim]
grad_value_statesfp32 [batch_size, num_heads, seq_len_enc, head_dim]
decoder_hidden_statesfp32 [batch_size, seq_len_dec, hidden_size]
encoder_hidden_statesfp32 [batch_size, seq_len_enc, cross_hidden_size]
q_weightfp32 [q_out_dim, hidden_size]
k_weightfp32 [kv_out_dim, cross_hidden_size]
v_weightfp32 [kv_out_dim, cross_hidden_size]
grad_decoder_hidden_statesfp32 [batch_size, seq_len_dec, hidden_size]
grad_encoder_hidden_statesfp32 [batch_size, seq_len_enc, cross_hidden_size]
grad_q_weightfp32 [q_out_dim, hidden_size]
grad_k_weightfp32 [kv_out_dim, cross_hidden_size]
grad_v_weightfp32 [kv_out_dim, cross_hidden_size]
Axes and behavior
head_dimconstant = 256
num_headsconstant = 16
q_out_dimq_out_dim = num_heads * head_dim
batch_sizevariable
kv_out_dimkv_out_dim = num_key_value_heads * head_dim
hidden_sizeconstant = 1536
seq_len_decvariable
seq_len_encvariable
cross_hidden_sizeconstant = 1024
num_key_value_headsconstant = 4
num_key_value_groupsconstant = 4
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias072_cross_attention_qkv_projection_with_gqa_repeat_backwardmodelparakeet-tdt-0-6b-v3sha256c80125700b20…
Sources: NVIDIA SOL-ExecBench (2026-08-22)last observed 2026-08-22How records are decidedJSON