Tape Nodes

Public entrypoint for the analytic tape-node proof library.

The modules below provide NodeFDerivCorrect and NodeFDerivCorrectAt facts for the primitive nodes used by graph-level reverse-mode autograd proofs:

• Context: vectorized contexts, projections, injections, and generic node construction;
• Elementwise: pointwise scalar functions and activations;
• Arithmetic: affine maps, pointwise arithmetic, and fixed-mask stochastic nodes;
• Matrix: matmul, transpose, and row/column matrix adapters;
• Batched: head-wise matmul and last-axis softmax on rank-3 tensors;
• Softmax: last-axis softmax and log-softmax;
• Shape: reshape, flatten, and reindexing adapters;
• Reductions: sums, broadcasts, axis reductions, and concatenation;
• Losses: training losses;
• Piecewise: branch-selected min/max nodes;
• GraphComposition: the DGraph layer for composing node-local proofs into model-level VJP theorems.