Standard sequential GraphSpec primitives

This module provides dense linear, ReLU, and axis-wise softmax primitives together with their corresponding Chain constructors. Each primitive supplies pure tensor semantics, executable TorchLean lowering, and deterministic layer initialization.

def NN.GraphSpec.Primitive.linear (inDim outDim : ℕ) : Primitive [[outDim, inDim], [outDim]] [inDim] [outDim]

The affine map $x \mapsto Wx + b$ from vectors of length inDim to vectors of length outDim.

Its parameter ABI is [[outDim, inDim], [outDim]]. The corresponding TorchLean layer initializes the weight from the layer occurrence index and initializes the bias exactly to zero.

def NN.GraphSpec.Primitive.relu (s : Spec.Shape) : Primitive [] s s

Parameter-free elementwise ReLU on tensors of shape s.

def NN.GraphSpec.Primitive.softmax (s : Spec.Shape) (axis : ℕ) [Spec.Shape.AxisInBounds axis s] : Primitive [] s s

Parameter-free softmax along axis. The axis bound is carried as an instance so ill-shaped softmax nodes cannot be constructed.

def NN.GraphSpec.Chain.linear (inDim outDim : ℕ) : Chain [[outDim, inDim], [outDim]] [inDim] [outDim]

Chain constructor for Primitive.linear.

def NN.GraphSpec.Chain.relu (s : Spec.Shape) : Chain [] s s

Chain constructor for Primitive.relu.

def NN.GraphSpec.Chain.softmax (s : Spec.Shape) (axis : ℕ) [Spec.Shape.AxisInBounds axis s] : Chain [] s s

Chain constructor for Primitive.softmax.