Struct KvLayer 

pub struct KvLayer {
    pub k: Arc<RwLock<Storage>>,
    pub v: Arc<RwLock<Storage>>,
    pub k_layout: Layout,
    pub v_layout: Layout,
    pub k_version: u64,
    pub v_version: u64,
    pub k_authority: AuthorityState,
    pub v_authority: AuthorityState,
}

One layer’s K + V slots in a backend-erased KV cache.

Storage Arcs are dispatch-erased: each Arc points at a fuel_memory::BackendStorage enum variant (Cpu, Cuda, Vulkan, Metal). The graph references these Arcs via the persistent map in InferenceContext; readers see the bytes through whichever Layout the graph carries for the consuming NodeId.

Fields

k: Arc<RwLock<Storage>>

v: Arc<RwLock<Storage>>

k_layout: Layout

View layout into k describing the live extent. For the grow- by-replacement pattern (today) the layout matches the storage’s full shape. For pre-allocated buffers (Phase E.3.2) the layout describes the leading [..., cached_len, ...] slice.

v_layout: Layout

k_version: u64

Monotonic write version for K. Bumps on every successful Op::WriteSlice targeting this slot. Used by the future multi-device coherence protocol to detect staleness.

v_version: u64

k_authority: AuthorityState

Coherence authority per slot. K and V can in principle diverge (a future optimization that decouples their update cadence could exploit this), though standard transformer decoding updates both together every step.

v_authority: AuthorityState