Expand description
§Serial SSD with a custom, memory-efficient backward (Mamba-3 double-SSD)
The SerialRecalculated path for the double-SSD pathway. The forward is the
same serial scan as super::super::serial, routed through the
Mamba3DoubleSsdBackendExt trait so that Autodiff backends substitute a
custom backward that recomputes per-chunk intermediates instead of storing
them (see super::backward / super::combined_backward). Plain
backends use the trait’s default body, which replays the serial kernels.
The default body runs under a generic backend B, where the high-level
Tensor (pinned to Dispatch) is unavailable, so the K1–K5 math goes
through the rank-tagged F primitive wrapper. K1/K2/K4 are reused by the
recompute backward in super::combined_backward; K5 is forward-only.
Traits§
- Mamba3
Double SsdAutodiff Backend Ext - Marker for an autodiff-capable backend that also implements the SSD extension trait (so the custom memory-efficient backward is available).
- Mamba3
Double SsdBackend Ext - Extends the backend for the memory-efficient serial recalculated SSD.
Functions§
- k1_
ssd_ 🔒chunk_ cumsum - Primitive port of
super::super::serial::k1_ssd_chunk_cumsum. - k2_
ssd_ 🔒bmm - Primitive port of
super::super::serial::k2_ssd_bmm(fusedL·M). - k3_
ssd_ 🔒chunk_ state - Primitive port of
super::super::serial::k3_ssd_chunk_state(lean: returns only the chunk-end state). - k4_
ssd_ 🔒state_ passing - Primitive port of
super::super::serial::k4_ssd_state_passing. - k5_
ssd_ 🔒chunk_ scan - Primitive port of
super::super::serial::k5_ssd_chunk_scan.