Audio encoder varlen attention with chunking backward
Backward pass for variable-length audio encoder attention with chunking. Computes gradients for hidden_states and all projection weights/biases through the attention mechanism, respecting chunk boundaries defined by cu_seqlens.
Reported evidence · last observed 2026-08-22. Reported by source; not independently reproduced. No public source. License unknown.
Current records
Workloadsuite of 16 cases · mean latencyd_model = 1280 · qkv_dim = 1280 · head_dim = 64 · num_heads = 2016 cases
Not measured on H100 for this workload. Challenges →
164.9µs91% SOL1.00×Reported · license unknown · no source2026-08-22
Measured exactly what you asked. Reported by source; not independently reproduced.
271.7µs90% SOL1.10×Reported · license unknown · no source2026-08-20
Measured exactly what you asked. Reported by source; not independently reproduced.
391.2µs87% SOL1.41×Reported · license unknown · no source2026-07-17
Measured exactly what you asked. Reported by source; not independently reproduced.
493.3µs87% SOL1.44×Reported · license unknown · no source2026-08-19
Measured exactly what you asked. Reported by source; not independently reproduced.
5107.7µs85% SOL1.66×Reported · license unknown · no source2026-08-04
Measured exactly what you asked. Reported by source; not independently reproduced.
6109.1µs85% SOL1.68×Reported · license unknown · no source2026-08-08
Measured exactly what you asked. Reported by source; not independently reproduced.
7109.3µs85% SOL1.68×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
8113.5µs84% SOL1.75×Reported · license unknown · no source2026-08-03
Measured exactly what you asked. Reported by source; not independently reproduced.
9114.3µs84% SOL1.76×Reported · license unknown · no source2026-07-11
Measured exactly what you asked. Reported by source; not independently reproduced.
10117.5µs84% SOL1.81×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
11119.6µs84% SOL1.84×Reported · license unknown · no source2026-08-15
Measured exactly what you asked. Reported by source; not independently reproduced.
12119.6µs84% SOL1.84×Reported · license unknown · no source2026-08-15
Measured exactly what you asked. Reported by source; not independently reproduced.
13119.6µs84% SOL1.84×Reported · license unknown · no source2026-08-15
Measured exactly what you asked. Reported by source; not independently reproduced.
14119.8µs84% SOL1.85×Reported · license unknown · no source2026-07-20
Measured exactly what you asked. Reported by source; not independently reproduced.
15148.5µs80% SOL2.29×Reported · license unknown · no source2026-08-23
Measured exactly what you asked. Reported by source; not independently reproduced.
16204.1µs74% SOL3.14×Reported · license unknown · no source2026-07-14
Measured exactly what you asked. Reported by source; not independently reproduced.
17259.0µs69% SOL3.99×Reported · license unknown · no source2026-08-05
Measured exactly what you asked. Reported by source; not independently reproduced.
18308.7µs65% SOL4.76×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
19640.9µs47% SOL9.88×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
201.30ms31% SOL20.09×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.