Latent KV expansion with split
DeepSeek-V3 latent KV expansion: RMSNorm on compressed KV (kv_lora_rank=512), linear projection to expanded space (num_heads * (qk_nope_head_dim + v_head_dim) = 128*256 = 32768), reshape, transpose, and split into K_nope and V tensors. Fuses normalization, projection, and memory layout transformations.
Reported evidence · last observed 2026-08-21. Reported by source; not independently reproduced. No public source. License unknown.
Current records
Workloadsuite of 16 cases · mean latencynum_heads = 128 · v_head_dim = 128 · kv_lora_rank = 512 · kv_expanded_dim = 32768 · qk_nope_head_dim = 12816 cases
Not measured on H100 for this workload. Challenges →
169.9µs79% SOL1.00×Reported · license unknown · no source2026-08-21
Measured exactly what you asked. Reported by source; not independently reproduced.
272.3µs77% SOL1.03×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
375.5µs76% SOL1.08×Reported · license unknown · no source2026-08-24
Measured exactly what you asked. Reported by source; not independently reproduced.
476.8µs75% SOL1.10×Reported · license unknown · no source2026-08-06
Measured exactly what you asked. Reported by source; not independently reproduced.
576.8µs75% SOL1.10×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
677.3µs74% SOL1.11×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
777.7µs75% SOL1.11×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
878.3µs74% SOL1.12×Reported · license unknown · no source2026-07-16
Measured exactly what you asked. Reported by source; not independently reproduced.
978.6µs74% SOL1.12×Reported · license unknown · no source2026-08-13
Measured exactly what you asked. Reported by source; not independently reproduced.
1083.6µs71% SOL1.20×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.
11101.6µs63% SOL1.45×Reported · license unknown · no source2026-07-24
Measured exactly what you asked. Reported by source; not independently reproduced.
12212.8µs37% SOL3.04×Reported · license unknown · no source2026-07-10
Measured exactly what you asked. Reported by source; not independently reproduced.