Module cache 

The double-SSD cache (ssm/k_state/v_state/cum_angle; no conv cache).

Mamba-3 Inference Caches

During autoregressive (token-by-token) generation, three pieces of state must be preserved between calls:

1. SSM hidden state — hₜ ∈ ℝ^{per_head_dim×state_rank} per head, compressed context.
2. Previous K state — Bₜ₋₁ per rank [batch, mimo_rank, nheads, state_rank], needed for the β term of the (double-ssd) trapezoidal recurrence.
3. Previous V state — xₜ₋₁ per head [batch, nheads, per_head_dim], paired with k_state to reconstruct β Bₜ₋₁ ⊗ xₜ₋₁.
4. Cumulative RoPE angle — the accumulated rotation angle up to position t, needed to correctly continue data-dependent rotary embeddings.

Note: Mamba-3 has no conv cache (the short 1-dimensional convolution present in Mamba-3 is removed; its role is absorbed by the trapezoidal discretization and the learnable B/C biases).

Structs

Mamba3DoubleSsdCache

The mutable state carried between decoding steps for a single Mamba-3 layer.

Mamba3DoubleSsdCacheConfig

Configuration / factory for a single Mamba3DoubleSsdCache.

Mamba3DoubleSsdCacheRecord

The record type for the module.

Mamba3DoubleSsdCacheRecordItem

The record item type for the module.

Mamba3DoubleSsdCaches

A collection of per-layer caches for a complete Mamba-3 network.

Mamba3DoubleSsdCachesConfig

Configuration / factory for Mamba3DoubleSsdCaches.

Mamba3DoubleSsdCachesRecord

The record type for the module.

Mamba3DoubleSsdCachesRecordItem

The record item type for the module.