Docs / nablatensor-core / com.nablatensor.engine
final class
AadTape
A recorded scalar computation graph: the "kernel" in kernel-AAD terms.
Nodes are appended in evaluation order, so the array order is already a valid topological order for the forward sweep and its reverse is a valid order for the adjoint sweep. Recording happens once; the tape is then translated into a device kernel and replayed for every scenario.
Methods
Input values as seen during recording; the starting point for a replay.
Normal draws per scenario on the default stream — the common single-stream total.
Number of independent random streams; 1 unless rec.stream(name) was used.
Standard-normal draws per scenario on stream.
Uniform [0,1) draws per scenario on stream.
For a RANDN/RANDU node: which stream it draws from.
For a RANDN/RANDU node: its 0-based index within its (stream, kind).
Total random draws per scenario across every stream and both kinds.
Flat-buffer offset of stream's draws in the layout [s0 normals | s0 uniforms | s1 normals | ...].
Flat-buffer index of a RANDN/RANDU node's draw.
Whether the tape uses anything beyond a single stream of standard-normal draws.
Node index of the primary (first-recorded) output.
Number of recorded outputs; 1 for a single rec.output(...).
Node index of output i, in recording order.
Name of output i; "value" for a single unnamed output.