eredu-nn

eredu-nn is Eredu's backend-neutral neural-compute boundary. Architecture crates are generic over the NeuralBackend trait. Each backend supplies its opaque tensor handle as the trait's associated Tensor type; that handle implements Tensor and identifies its execution context.

The boundary does not exchange tensor data. Architecture calls are statically dispatched after monomorphization, and operations such as linear projection, layer normalization, rotary encoding, and scaled dot-product attention remain backend fusion points. A backend therefore retains control of storage, device placement, graph construction, laziness, synchronization, and kernel fusion.

Shared tensor-independent geometry validates normalization, grouped gates and causal-mask positions before backend work. Native and scalar implementations use these contracts without sharing layout, dtype or kernel policy.

eredu-nn has no concrete-backend features or accelerator dependencies. Each backend implements NeuralBackend for a local backend type and Tensor for a local tensor newtype, which keeps the neutral contracts independent while satisfying Rust's orphan rules. A new backend implements these contracts once; models in backend-neutral architecture crates can then be reused without being ported.