Expand description
Trained-state storage — authoritative data produced by fit().
States are distinct from CacheStore entries:
- Cache entries are discardable — the system can recompute them.
- States are authoritative — they are the product of training and belong to the Graph that produced them. They must not be evicted arbitrarily.
StateStore is the trait; implementations may keep states in memory,
on local disk, or in object storage. States are returned as
Arc<Value> so the hot forward path can borrow them (&*arc) without
cloning potentially-large tensors.
Structs§
- Memory
State Store - In-memory
StateStore— the default backend.
Traits§
- State
Store - Storage for trained filter states, keyed by node id.