KV cache update with RoPE backward
Backward pass for KV cache update with RoPE. Computes gradients for key_states, value_states, cos, sin, and input caches given gradients from downstream attention operations.
Reported evidence · last observed 2026-07-10. Reported by source; not independently reproduced. No public source. License unknown.
Current records
Workloadsuite of 16 cases · mean latencyhalf_dim = 64 · head_dim = 128 · num_kv_heads = 816 cases
Not measured on H100 for this workload. Challenges →
111.2µs100% SOL1.00×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
211.3µs100% SOL1.01×Reported · license unknown · no source2026-08-04
Measured exactly what you asked. Reported by source; not independently reproduced.
312.7µs99% SOL1.13×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
418.6µs99% SOL1.66×Reported · license unknown · no source2026-07-24
Measured exactly what you asked. Reported by source; not independently reproduced.
520.5µs99% SOL1.82×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
625.9µs98% SOL2.30×Reported · license unknown · no source2026-07-16
Measured exactly what you asked. Reported by source; not independently reproduced.
725.9µs99% SOL2.31×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
831.3µs97% SOL2.79×Reported · license unknown · no source2026-07-13
Measured exactly what you asked. Reported by source; not independently reproduced.
933.9µs96% SOL3.02×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
1068.0µs88% SOL6.05×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.