Skip to content
KernelIndex
Search⌘K
Sign in

MoE expert load balancing and token capacity backward

22 eligible runs
moe

Backward pass for MoE expert load balancing and token capacity. Computes gradients through load_balance_loss, sum reduction, and scatter operations. Since topk_idx contains discrete indices and hidden_states is only used for shape, gradients w.r.t. both inputs are None.

Fastest reported
7.21µsmean · 73% of SOL
doubleAILicense unknown · unknown

Reported evidence · last observed 2026-06-25. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencyn_routed_experts = 256 · num_experts_per_tok = 816 cases
batch_seq_len
dtype
2048fp322080fp322112fp322144fp322176fp322208fp322371fp323072fp323169fp323557fp324096fp3232768fp3265536fp32131072fp32262144fp32524288fp32
scrolls · 16 cases total
Aggregatessuite of 16 cases · mean latency
HardwareBest knownImplementationRuns
NVIDIA B200 · env 17.21 µsdoubleAI12NVIDIA B200 · env 27.23 µsdoubleAI10

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 v1.0 · mean · 12 results · last observed 2026-07-23Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
doubleAI
7.21µs73% SOL
1.00×
Reported · license unknown · no source
2026-06-25

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

no sourcelicense unknownno install recipeRun detail →
2
ac4k
7.60µs71% SOL
1.05×
Reported · license unknown · no source
2026-06-26

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

no sourcelicense unknownno install recipeRun detail →
3
bismuth83
7.62µs71% SOL
1.06×
Reported · license unknown · no source
2026-04-24

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

no sourcelicense unknownno install recipeRun detail →
4
Databricks
7.64µs71% SOL
1.06×
Reported · license unknown · no source
2026-06-26

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

no sourcelicense unknownno install recipeRun detail →
5
Recursive
7.69µs71% SOL
1.07×
Reported · license unknown · no source
2026-06-07

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

no sourcelicense unknownno install recipeRun detail →
6
doubleAI
7.76µs71% SOL
1.08×
Reported · license unknown · no source
2026-05-20

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

no sourcelicense unknownno install recipeRun detail →
7
Argus
7.83µs71% SOL
1.09×
Reported · license unknown · no source
2026-07-01

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

no sourcelicense unknownno install recipeRun detail →
8
Amir M. Mir | SF Tensor
8.28µs69% SOL
1.15×
Reported · license unknown · no source
2026-04-22

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

no sourcelicense unknownno install recipeRun detail →
9
wychi888
15.1µs53% SOL
2.10×
Reported · license unknown · no source
2026-06-10

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

no sourcelicense unknownno install recipeRun detail →
10
Cursor
16.3µs51% SOL
2.27×
Reported · license unknown · no source
2026-04-14

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

no sourcelicense unknownno install recipeRun detail →
11
Amir M. Mir | SF Tensor
26.7µs37% SOL
3.70×
Reported · license unknown · no source
2026-04-21

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

no sourcelicense unknownno install recipeRun detail →
12
Igor Bragaia
103.9µs13% SOL
14.42×
Reported · license unknown · no source
2026-05-28

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

no sourcelicense unknownno install recipeRun detail →

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
Databricks
unknown
7.64µs
1.06×
Reported
license unknown · no source
bismuth83
unknown
7.62µs
1.06×
Reported
license unknown · no source
Recursive
unknown
7.69µs
1.07×
Reported
license unknown · no source
Hyra
unknown
7.81µs
1.08×
Reported
license unknown · no source
Argus
unknown
7.83µs
1.09×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
7.85µs
1.09×
Reported
license unknown · no source
TH3CHARLie | TileGym
unknown
8.14µs
1.13×
Reported
license unknown · no source
Geometric
unknown
8.82µs
1.22×
Reported
license unknown · no source
Show all 22 implementations ›
wychi888
unknown
15.1µs
2.10×
Reported
license unknown · no source
Cursor
unknown
16.1µs
2.23×
Reported
license unknown · no source
Igor Bragaia
unknown
103.9µs
14.42×
Reported
license unknown · no source
Databricks
unknown
7.71µs
1.07×
Reported
license unknown · no source
ac4k
unknown
7.70µs
1.07×
Reported
license unknown · no source
doubleAI
unknown
7.76µs
1.08×
Reported
license unknown · no source
bismuth83
unknown
7.84µs
1.09×
Reported
license unknown · no source
Recursive
unknown
7.89µs
1.09×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
8.28µs
1.15×
Reported
license unknown · no source
Cursor
unknown
16.3µs
2.27×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
26.7µs
3.70×
Reported
license unknown · no source
doubleAI
unknown
7.21µs
1.00×
Reported
license unknown · no source
doubleAI
unknown
7.23µs
1.00×
Reported
license unknown · no source
ac4k
unknown
7.60µs
1.05×
Reported
license unknown · no source

Semantics

Inputs and outputs
grad_tokens_per_expertfp32 [n_routed_experts]
grad_expert_maskfp32 [batch_seq_len, n_routed_experts]
grad_load_balance_lossfp32 [1]
topk_idxint64 [batch_seq_len, num_experts_per_tok]
expert_maskint64 [batch_seq_len, n_routed_experts]
tokens_per_expertint32 [n_routed_experts]
trainingbool [1]
grad_expert_mask_outfp32 [batch_seq_len, n_routed_experts]
Axes and behavior
batch_seq_lenvariable
n_routed_expertsconstant = 256
num_experts_per_tokconstant = 8
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias042_moe_expert_load_balancing_and_token_capacity_backwardmodeldeepseek-r1-0528sha256d28f0583e2ad…
Sources: NVIDIA SOL-ExecBench (2026-07-23)last observed 2026-07-23How records are decidedJSON