Skip to content
KernelIndex
Search⌘K
Sign in

SAM hq mask decoder two way transformer

28 eligible runs
vision

Two-way transformer block for SAM-HQ mask decoder performing bidirectional cross-attention between sparse prompt tokens and dense image embeddings. Includes: (1) self-attention on sparse tokens, (2) cross-attention from tokens to image, (3) MLP on tokens, (4) cross-attention from image to tokens. Uses downsampled internal dimensions (256->128) for cross-attention efficiency.

Fastest reported
142.4µsmean · 80% 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 latencymlp_dim = 2048 · hidden_size = 256 · internal_dim = 128 · num_attention_heads = 816 cases
batch_size
n_query_tokens
n_key_tokens
point_batch_size
dtype
1440961fp321440964fp321510242fp321640961fp321840961fp322440961fp322440962fp322640961fp322840962fp324510242fp324520481fp324640961fp328440961fp328440962fp328520481fp3216440961fp32
scrolls · 16 cases total
Aggregatessuite of 16 cases · mean latency
HardwareBest knownImplementationRuns
NVIDIA B200 · env 1142.4 µsRIAC as well20NVIDIA B200 · env 2184.3 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 · 20 results · last observed 2026-08-21Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
RIAC as well
142.4µs80% 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
Amir M. Mir | SF Tensor
173.8µs77% SOL
1.22×
Reported · license unknown · no source
2026-07-15

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

no sourcelicense unknownno install recipeRun detail →
3
Hyra
174.3µs77% 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 →
4
doubleAI
193.0µs75% SOL
1.36×
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
ac4k
198.5µs75% SOL
1.39×
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 →
6
Geometric
199.4µs74% SOL
1.40×
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 →
7
Geometric
236.1µs71% SOL
1.66×
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 →
8
Geometric
253.7µs69% SOL
1.78×
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 →
9
Geometric
289.9µs66% SOL
2.04×
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
Proud Wren @ NVIDIA
295.5µs66% SOL
2.08×
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
306.9µs65% SOL
2.16×
Reported · license unknown · no source
2026-08-07

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

no sourcelicense unknownno install recipeRun detail →
12
Geometric
309.8µs65% SOL
2.18×
Reported · license unknown · no source
2026-08-15

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

no sourcelicense unknownno install recipeRun detail →
13
Geometric
349.2µs62% SOL
2.45×
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 →
14
Geometric
361.4µs61% SOL
2.54×
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 →
15
Geometric
369.9µs61% SOL
2.60×
Reported · license unknown · no source
2026-08-16

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

no sourcelicense unknownno install recipeRun detail →
16
Geometric
394.6µs59% SOL
2.77×
Reported · license unknown · no source
2026-08-13

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

no sourcelicense unknownno install recipeRun detail →
17
Amir M. Mir | SF Tensor
443.8µs56% SOL
3.12×
Reported · license unknown · no source
2026-07-14

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

no sourcelicense unknownno install recipeRun detail →
18
Amir M. Mir | SF Tensor
474.7µs55% SOL
3.33×
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 →
19
Geometric
680.4µs45% SOL
4.78×
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 →
20
Cursor
1.20ms32% SOL
8.46×
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
142.4µs
1.00×
Reported
license unknown · no source
doubleAI
unknown
188.4µs
1.32×
Reported
license unknown · no source
ac4k
unknown
198.5µs
1.39×
Reported
license unknown · no source
Geometric
unknown
199.4µs
1.40×
Reported
license unknown · no source
Proud Wren @ NVIDIA
unknown
295.5µs
2.08×
Reported
license unknown · no source
Cursor
unknown
1.02ms
7.17×
Reported
license unknown · no source
doubleAI
unknown
193.0µs
1.36×
Reported
license unknown · no source
Geometric
unknown
236.1µs
1.66×
Reported
license unknown · no source
Show all 28 implementations ›
doubleAI
unknown
254.1µs
1.79×
Reported
license unknown · no source
Geometric
unknown
253.7µs
1.78×
Reported
license unknown · no source
Geometric
unknown
289.9µs
2.04×
Reported
license unknown · no source
Geometric
unknown
306.9µs
2.16×
Reported
license unknown · no source
Geometric
unknown
309.8µs
2.18×
Reported
license unknown · no source
Geometric
unknown
349.2µs
2.45×
Reported
license unknown · no source
Geometric
unknown
361.4µs
2.54×
Reported
license unknown · no source
Geometric
unknown
369.9µs
2.60×
Reported
license unknown · no source
Geometric
unknown
394.6µs
2.77×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
443.8µs
3.12×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
463.9µs
3.26×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
474.7µs
3.33×
Reported
license unknown · no source
doubleAI
unknown
620.5µs
4.36×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
673.7µs
4.73×
Reported
license unknown · no source
Geometric
unknown
680.4µs
4.78×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
1.06ms
7.47×
Reported
license unknown · no source
Cursor
unknown
1.20ms
8.46×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
173.8µs
1.22×
Reported
license unknown · no source
Hyra
unknown
174.3µs
1.22×
Reported
license unknown · no source
Recursive
unknown
184.3µs
1.29×
Reported
license unknown · no source

Semantics

Inputs and outputs
queriesfp32 [batch_size, point_batch_size, n_query_tokens, hidden_size]
keysfp32 [batch_size, point_batch_size, n_key_tokens, hidden_size]
query_point_embeddingfp32 [batch_size, point_batch_size, n_query_tokens, hidden_size]
key_point_embeddingfp32 [batch_size, point_batch_size, n_key_tokens, hidden_size]
self_attn_q_weightfp32 [hidden_size, hidden_size]
self_attn_q_biasfp32 [hidden_size]
self_attn_k_weightfp32 [hidden_size, hidden_size]
self_attn_k_biasfp32 [hidden_size]
self_attn_v_weightfp32 [hidden_size, hidden_size]
self_attn_v_biasfp32 [hidden_size]
self_attn_out_weightfp32 [hidden_size, hidden_size]
self_attn_out_biasfp32 [hidden_size]
layer_norm1_weightfp32 [hidden_size]
layer_norm1_biasfp32 [hidden_size]
cross_t2i_q_weightfp32 [internal_dim, hidden_size]
cross_t2i_q_biasfp32 [internal_dim]
cross_t2i_k_weightfp32 [internal_dim, hidden_size]
cross_t2i_k_biasfp32 [internal_dim]
cross_t2i_v_weightfp32 [internal_dim, hidden_size]
cross_t2i_v_biasfp32 [internal_dim]
cross_t2i_out_weightfp32 [hidden_size, internal_dim]
cross_t2i_out_biasfp32 [hidden_size]
layer_norm2_weightfp32 [hidden_size]
layer_norm2_biasfp32 [hidden_size]
mlp_lin1_weightfp32 [mlp_dim, hidden_size]
mlp_lin1_biasfp32 [mlp_dim]
mlp_lin2_weightfp32 [hidden_size, mlp_dim]
mlp_lin2_biasfp32 [hidden_size]
layer_norm3_weightfp32 [hidden_size]
layer_norm3_biasfp32 [hidden_size]
cross_i2t_q_weightfp32 [internal_dim, hidden_size]
cross_i2t_q_biasfp32 [internal_dim]
cross_i2t_k_weightfp32 [internal_dim, hidden_size]
cross_i2t_k_biasfp32 [internal_dim]
cross_i2t_v_weightfp32 [internal_dim, hidden_size]
cross_i2t_v_biasfp32 [internal_dim]
cross_i2t_out_weightfp32 [hidden_size, internal_dim]
cross_i2t_out_biasfp32 [hidden_size]
layer_norm4_weightfp32 [hidden_size]
layer_norm4_biasfp32 [hidden_size]
skip_first_layer_pebool scalar
output_queriesfp32 [batch_size, point_batch_size, n_query_tokens, hidden_size]
output_keysfp32 [batch_size, point_batch_size, n_key_tokens, hidden_size]
Axes and behavior
mlp_dimconstant = 2048
batch_sizevariable
hidden_sizeconstant = 256
internal_dimconstant = 128
n_key_tokensvariable
n_query_tokensvariable
point_batch_sizevariable
num_attention_headsconstant = 8
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias075_sam_hq_mask_decoder_two_way_transformermodelsam-hq-vit-basesha256449479eb7709…
Sources: NVIDIA SOL-ExecBench (2026-08-21)last observed 2026-08-21How records are decidedJSON