Module cache 

Mamba-1 Inference Caches

State carried between calls during autoregressive (token-by-token) generation. During training or prefill the whole sequence is processed at once by Mamba1::forward; during decoding one token is processed per call by Mamba1::step. Both modes thread the same two pieces of state:

1. Convolution window — the last conv_kernel pre-activation inputs to the depthwise Conv1d, kept so each step can apply the causal filter without reprocessing earlier tokens.

2. SSM hidden state — the per-channel state matrix that compresses the entire past into a fixed-size representation (independent of how many tokens have been seen).

Mirrors crate::mamba2::cache; Mamba-1 has a single SSD-free recurrence so the cache holds plain tensors rather than the head-structured state of Mamba-2.

Structs

Mamba1Cache

The mutable state carried between calls for a single Mamba-1 layer.

Mamba1CacheConfig

Configuration / factory for a single Mamba1Cache.

Mamba1Caches

A collection of per-layer caches for a complete Mamba-1 network.

Mamba1CachesConfig

Configuration / factory for Mamba1Caches.