Expand description
Mamba-2: Structured State Space Duality (SSD).
§Mamba-2
Structured State Space Duality (SSD). Mamba-2 recasts the selective SSM
recurrence as a chunkwise algorithm built from batched GEMMs, making it
tensor-core friendly for training while remaining exactly equivalent to the
recurrent form used for decoding. See mamba2 for the full SSD math.
mamba2— the SSD block. The residual layer stack, the full language model, and the bidirectional wrappers are the family-generic types in [crate::generic] (MambaLatentNet/MambaVocabNet/MambaBidiLayers).- [
cache] — the conv-window + SSM-state carried between calls. - [
ssd] — the pluggable chunkwise SSD algorithms (Minimal / Serial / SerialRecalculated) and the backend extension trait.