Bounded numerical teaching oracle · zero network

Triton RMSNorm lab

Compare an independent scalar reference with a separately implemented power-of-two tiled model. Every accepted run exposes mean squares, reciprocal RMS values, masked lanes, affine outputs, errors, and modeled traffic.

Validated fixture

Ready with a two-row, five-column fixture.

Declared bounds and truth boundary

One to 16 rows, width two to 256, finite values and weights with magnitude at most 1e6, and epsilon from 1e-12 through 1e-2. The loop and allocation bounds are fixed by those limits.

This lab does not execute Triton, inspect generated GPU code, measure hardware, or predict speedup.

Row receipt

Run the proof to populate independent results.

RowMean squareReciprocal RMSMasked lanesMax abs / rel errorOutputResult

Mutation fixtures

    Truth boundary: numerical and modeled-work evidence for bounded synthetic inputs only. Pair its receipt with a real Triton/PyTorch test on the target device.