Attention Backend Capsules

Attention is the first place where the backend-contract distinction matters in practice.

The proof-facing FlashAttention spec and every registered runtime provider use hard-mask semantics: blocked entries have exactly zero softmax numerator. Additive attention biases are a separate operation and are never used to encode a boolean mask.

def NN.Backend.Attention.torchLeanComposed : KernelCapsule

Composed TorchLean attention path.

TorchLean owns the tape and local VJP. CUDA batched matrix multiplication evaluates the two dense contractions, while TorchLean's hard-masked softmax supplies the attention weights. Unlike the direct reference kernel, this implementation has the expected quadratic dependence on sequence length.

def NN.Backend.Attention.nativeDirectAttention : KernelCapsule

Direct native CUDA reference path.

The implementation computes attention without materializing the score matrix, but it is not the IO-tiled FlashAttention algorithm. Its backward kernel recomputes row statistics and is intended for parity checks and small inputs rather than large-model training.

def NN.Backend.Attention.libTorchSDPAForward : KernelCapsule

LibTorch SDPA forward provider while TorchLean keeps the graph/tape boundary.

def NN.Backend.Attention.capsules : Array KernelCapsule

Built-in attention capsules in default planner order. Optional external providers register their own capsules in their provider modules.