SAM hq mask decoder IoU hypernetwork fusion backward
18 eligible runs
vision
Backward pass for SAM-HQ mask decoder IoU prediction and hypernetwork weight generation. Computes gradients for 5 MLPs: 1 IoU prediction MLP (256->1024->1024->4) and 4 parallel hypernetwork MLPs (256->256->256->32).
Fastest reported
43.5µsmean · 81% of SOL
HyraLicense unknown · unknown
Reported evidence · last observed 2026-07-17. Reported by source; not independently reproduced. No public source. License unknown.
No public sourceRun detail →
Current records
Workloadsuite of 16 cases · mean latencyhidden_size = 256 · num_mask_tokens = 4 · output_channels = 32 · iou_head_hidden_dim = 102416 cases
Aggregatessuite of 16 cases · mean latency
HardwareBest knownImplementationRuns
NVIDIA B200 · env 143.5 µsHyra12NVIDIA B200 · env 257.8 µsdoubleAI6Not 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 · 12 results · last observed 2026-08-07Record history →
#
Implementation
Latency
vs #1
Trust
Observed
143.5µs81% SOL1.00×Reported · license unknown · no source2026-07-17
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
256.3µs77% SOL1.29×Reported · license unknown · no source2026-08-03
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
357.0µs76% SOL1.31×Reported · license unknown · no source2026-08-07
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
457.4µs76% SOL1.32×Reported · license unknown · no source2026-07-16
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
560.7µs75% SOL1.40×Reported · license unknown · no source2026-07-11
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
661.8µs75% SOL1.42×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
794.9µs66% SOL2.18×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
8142.0µs56% SOL3.26×Reported · license unknown · no source2026-07-14
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
9154.1µs54% SOL3.54×Reported · license unknown · no source2026-07-26
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
10239.9µs43% SOL5.51×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
11293.1µs38% SOL6.74×Reported · license unknown · no source2026-07-23
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
12419.1µs30% SOL9.64×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
no sourcelicense unknownno install recipeRun detail →
Implementations
Implementation
Runtime
Best latency
Evidence
Availability
Show all 18 implementations ›Showing all 18 implementations ⌄
Semantics
Inputs and outputs
grad_iou_scoresfp32 [batch_size, point_batch_size, num_mask_tokens]
grad_hyper_weightsfp32 [batch_size, point_batch_size, num_mask_tokens, output_channels]
iou_token_outfp32 [batch_size, point_batch_size, hidden_size]
mask_tokens_outfp32 [batch_size, point_batch_size, num_mask_tokens, hidden_size]
iou_proj_in_weightfp32 [iou_head_hidden_dim, hidden_size]
iou_proj_in_biasfp32 [iou_head_hidden_dim]
iou_hidden_weightfp32 [iou_head_hidden_dim, iou_head_hidden_dim]
iou_hidden_biasfp32 [iou_head_hidden_dim]
iou_proj_out_weightfp32 [num_mask_tokens, iou_head_hidden_dim]
iou_proj_out_biasfp32 [num_mask_tokens]
iou_hidden1fp32 [batch_size, point_batch_size, iou_head_hidden_dim]
iou_hidden1_relufp32 [batch_size, point_batch_size, iou_head_hidden_dim]
iou_hidden2fp32 [batch_size, point_batch_size, iou_head_hidden_dim]
iou_hidden2_relufp32 [batch_size, point_batch_size, iou_head_hidden_dim]
hyper_proj_in_weightsfp32 [num_mask_tokens, hidden_size, hidden_size]
hyper_proj_in_biasesfp32 [num_mask_tokens, hidden_size]
hyper_hidden_weightsfp32 [num_mask_tokens, hidden_size, hidden_size]
hyper_hidden_biasesfp32 [num_mask_tokens, hidden_size]
hyper_proj_out_weightsfp32 [num_mask_tokens, output_channels, hidden_size]
hyper_proj_out_biasesfp32 [num_mask_tokens, output_channels]
hyper_hidden1_0fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden1_1fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden1_2fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden1_3fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden1_relu_0fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden1_relu_1fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden1_relu_2fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden1_relu_3fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden2_0fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden2_1fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden2_2fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden2_3fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden2_relu_0fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden2_relu_1fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden2_relu_2fp32 [batch_size, point_batch_size, hidden_size]
hyper_hidden2_relu_3fp32 [batch_size, point_batch_size, hidden_size]
grad_iou_token_outfp32 [batch_size, point_batch_size, hidden_size]
grad_mask_tokens_outfp32 [batch_size, point_batch_size, num_mask_tokens, hidden_size]
grad_iou_proj_in_weightfp32 [iou_head_hidden_dim, hidden_size]
grad_iou_proj_in_biasfp32 [iou_head_hidden_dim]
grad_iou_hidden_weightfp32 [iou_head_hidden_dim, iou_head_hidden_dim]
grad_iou_hidden_biasfp32 [iou_head_hidden_dim]
grad_iou_proj_out_weightfp32 [num_mask_tokens, iou_head_hidden_dim]
grad_iou_proj_out_biasfp32 [num_mask_tokens]
grad_hyper_proj_in_weightsfp32 [num_mask_tokens, hidden_size, hidden_size]
grad_hyper_proj_in_biasesfp32 [num_mask_tokens, hidden_size]
grad_hyper_hidden_weightsfp32 [num_mask_tokens, hidden_size, hidden_size]
grad_hyper_hidden_biasesfp32 [num_mask_tokens, hidden_size]
grad_hyper_proj_out_weightsfp32 [num_mask_tokens, output_channels, hidden_size]
grad_hyper_proj_out_biasesfp32 [num_mask_tokens, output_channels]
Axes and behavior
batch_sizevariable
hidden_sizeconstant = 256
num_mask_tokensconstant = 4
output_channelsconstant = 32
point_batch_sizevariable
iou_head_hidden_dimconstant = 1024
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias074_sam_hq_mask_decoder_iou_hypernetwork_fusion_backwardmodelsam-hq-vit-basesha25672e25bea746d…
Sources: NVIDIA SOL-ExecBench (2026-08-07)last observed 2026-08-07How records are decidedJSON