Feeding a final prediction into the next token introduces a dependency: the current position must finish first. Two-pass training substitutes preliminary predictions so positions can be processed in parallel within a pass. This is an approximation to exact recurrence.

A one-pass memory model can still be slow if temporal matrix updates or backward-state storage dominate. Thresholding weights to zero does not reduce arithmetic when all candidate projections and matrix products still execute.

A short DC-B1 test skipped candidate work at context tails where later predictions could never read the writes. Comparisons restored the same 10M weights, Adam state, and RNG. Across four real-data windows, maximum differences in loss, gradients, parameter updates, and optimizer state were zero.

The recorded short-test speed ratio was about 1.045×. This is an example of removing provably unused work while preserving the tested updates. It does not establish the end-to-end speedup of a long training run.

Measure speed on the same real data, sequence lengths, effective batch, and precision. Separate first-compilation cost from steady updates, then include saving and evaluation in total runtime. For fusion or precision changes, compare gradients and Adam updates as well as outputs.

Record numerical-order changes separately from changes such as limiting candidate count or approximating feedback. Both may be useful, but readers need to know whether execution was accelerated or the learned computation was changed.