Module driver 

Hardware-agnostic compute driver trait.

The Driver trait exposes low-level compute primitives (GEMM, layer-norm, activations, etc.) that each hardware backend implements. Model architectures are generic over D: Driver and compose these primitives into a forward pass.

Design

• Associated type Tensor: each driver defines its own opaque tensor handle (Metal: buffer+offset, CUDA: device pointer, CPU: ndarray).
• Not object-safe: architectures use D: Driver generics so the compiler can monomorphize and inline driver calls.
• Send + Sync: drivers are shared across the pipeline.

Structs

BatchInputs

Batch input tensors on device, produced by Driver::prepare_batch.

Traits

Driver

Hardware-agnostic compute primitives for BERT inference.