Reference and Portable Backend Capsules

Portable capsules for paths that do not require CUDA or LibTorch.

These are not meant to win benchmarks. They keep TorchLean runnable on CPU-only machines and give the planner a clear fallback vocabulary for reference/spec-aligned execution.

def NN.Backend.Reference.capsule (name : String) (op : BackendOp) (valueSummary vjpSummary : String) (vjpMode : VJPMode := VJPMode.torchLeanTape) : KernelCapsule

Build a checked portable CPU capsule with explicit value, VJP, shape, and layout contracts.

def NN.Backend.Reference.pointwiseCapsule (op : BackendOp) : KernelCapsule

Build the standard portable capsule for a pointwise operation.

def NN.Backend.Reference.reductionCapsule (op : BackendOp) : KernelCapsule

Build a portable reduction capsule with deterministic left-to-right accumulation.

def NN.Backend.Reference.accumulationCapsule (name : String) (op : BackendOp) (valueSummary vjpSummary : String) : KernelCapsule

Reference kernel whose scalar result contains an explicit left-to-right accumulation.

Matrix products, affine layers, convolutions, and averaging operations all reduce several products or samples into one output entry. Keeping this constructor separate from pointwise kernels prevents the numerical audit from incorrectly reporting that reduction order is irrelevant.

def NN.Backend.Reference.viewCapsule (op : BackendOp) : KernelCapsule

Build the standard portable capsule for a shape or layout transformation.

def NN.Backend.Reference.forwardOnlyCapsule (op : BackendOp) (valueSummary : String) : KernelCapsule

Build a portable forward-only capsule with no registered reverse derivative.

def NN.Backend.Reference.convPoolCapsule (op : BackendOp) : KernelCapsule

Build a portable capsule for channel-first convolution or pooling.

def NN.Backend.Reference.selectionCapsule (op : BackendOp) : KernelCapsule

Reference window selection with deterministic traversal and tie handling.

def NN.Backend.Reference.relu : KernelCapsule

Reference ReLU activation.

def NN.Backend.Reference.gelu : KernelCapsule

Reference GELU activation.

def NN.Backend.Reference.add : KernelCapsule

Reference pointwise addition.

def NN.Backend.Reference.sub : KernelCapsule

Reference pointwise subtraction.

def NN.Backend.Reference.mul : KernelCapsule

Reference pointwise multiplication.

def NN.Backend.Reference.scale : KernelCapsule

Reference scalar multiplication.

def NN.Backend.Reference.abs : KernelCapsule

Reference pointwise absolute value.

def NN.Backend.Reference.sqrt : KernelCapsule

Reference pointwise square root.

def NN.Backend.Reference.clamp : KernelCapsule

Reference pointwise interval clamp.

def NN.Backend.Reference.max : KernelCapsule

Reference pointwise maximum.

def NN.Backend.Reference.min : KernelCapsule

Reference pointwise minimum.

def NN.Backend.Reference.sigmoid : KernelCapsule

Reference pointwise sigmoid.

def NN.Backend.Reference.tanh : KernelCapsule

Reference pointwise hyperbolic tangent.

def NN.Backend.Reference.softplus : KernelCapsule

Reference pointwise softplus.

def NN.Backend.Reference.exp : KernelCapsule

Reference pointwise exponential.

def NN.Backend.Reference.log : KernelCapsule

Reference pointwise natural logarithm.

def NN.Backend.Reference.sin : KernelCapsule

Reference pointwise sine.

def NN.Backend.Reference.cos : KernelCapsule

Reference pointwise cosine.

def NN.Backend.Reference.inv : KernelCapsule

Reference pointwise reciprocal.

def NN.Backend.Reference.safeLog : KernelCapsule

Reference guarded logarithm used by numerically defensive programs.

def NN.Backend.Reference.logSoftmax : KernelCapsule

Reference log-softmax reduction and normalization.

def NN.Backend.Reference.softmax : KernelCapsule

Reference softmax path.

def NN.Backend.Reference.hardMaskedSoftmax : KernelCapsule

Reference hard-masked softmax with exact zero weight at blocked coordinates.

def NN.Backend.Reference.reduceSum : KernelCapsule

Reference sum reduction.

def NN.Backend.Reference.reduceMean : KernelCapsule

Reference arithmetic-mean reduction.

def NN.Backend.Reference.reshape : KernelCapsule

Reference shape-preserving reshape view.

def NN.Backend.Reference.permute : KernelCapsule

Reference axis permutation.

def NN.Backend.Reference.broadcast : KernelCapsule

Reference tensor broadcasting.

def NN.Backend.Reference.concat : KernelCapsule

Reference tensor concatenation.

def NN.Backend.Reference.slice : KernelCapsule

Reference contiguous tensor slice.

def NN.Backend.Reference.gather : KernelCapsule

Reference indexed gather.

def NN.Backend.Reference.scatterAdd : KernelCapsule

Reference indexed scatter-add.

def NN.Backend.Reference.randUniform : KernelCapsule

Reference seeded uniform-random tensor generation.

def NN.Backend.Reference.bernoulliMask : KernelCapsule

Reference seeded Bernoulli-mask generation.

def NN.Backend.Reference.matmul : KernelCapsule

Reference matmul path.

def NN.Backend.Reference.linear : KernelCapsule

Reference linear layer.

def NN.Backend.Reference.mseLoss : KernelCapsule

Reference mean-squared-error loss.

def NN.Backend.Reference.layerNorm : KernelCapsule

Reference layer normalization.

def NN.Backend.Reference.batchNorm : KernelCapsule

Reference batch normalization.

def NN.Backend.Reference.conv : KernelCapsule

Reference generic channel-first convolution.

def NN.Backend.Reference.convTranspose : KernelCapsule

Reference generic channel-first transpose convolution.

def NN.Backend.Reference.maxPool : KernelCapsule

Reference max pooling.

def NN.Backend.Reference.smoothMaxPool : KernelCapsule

Reference smooth max pooling.

def NN.Backend.Reference.avgPool : KernelCapsule

Reference average pooling.

def NN.Backend.Reference.attention : KernelCapsule

Reference attention path using the composed TorchLean expression.

def NN.Backend.Reference.capsules : Array KernelCapsule

Cross-platform reference capsules.