Expand description
§Double-SSD pathway (VikramLex-style)
Realises the Mamba-3 trapezoidal recurrence as two standard SSD calls that reuse the Mamba-2-like kernels:
- γ-SSM:
hᵞₜ = αₜ hᵞₜ₋₁ + γₜ Bₜ xₜ(current token) - β-SSM:
hᵝₜ = αₜ hᵝₜ₋₁ + βₜ Bₜ₋₁ xₜ₋₁(previous token) hₜ = hᵞₜ + hᵝₜ.
Simple and easy to verify, at the cost of ~2× the intra-chunk and
chunk-state memory of the single_ssd pathway.
Modules§
- cache
- The double-SSD cache (
ssm/k_state/v_state/cum_angle; no conv cache). - double_
ssd forward_double_ssd/step_double_ssdand the RoPE helpers.- prelude
- Public re-exports for the double-SSD pathway.
- ssd
- The standard SSD kernels reused by both the γ and β passes. Standard MIMO-first SSD kernels reused by the double-SSD γ and β passes.