Filip Makraduli, a machine learning engineer and author of the Transformer Tricks repository, presented work co-authored with Niels Graf on reducing the cost of the RMS norm layer in transformer architectures using two algebraic tricks—weight folding and deferred normalization—and recounted the CUDA-level debugging required to realize the speedup, including a race condition that made his model generate text with a one-step lag and repeat words from the past. The work argues that RMS norm, despite doing almost none of the arithmetic in a transformer, consumes disproportionate wall-clock time—it can be invoked roughly 33 times in a single decode step—because GPUs are inefficient at kernel launches, memory movement, and synchronization waits. The paper's contributions include Weightless normalization (Proposition 1), which folds gain and weight into a single offline-computed matrix W* and can be applied via the Transformer Tricks repo without kernel work; Deferred normalization (Proposition 2), which defers the scalar division to run in parallel with the matmul and requires custom kernel work; and Pre-normalization cancellation (Proposition 3), which applies when RMS norm appears twice in architectures such as Gemma 4. The deferred-normalization proposition requires explicit stream synchronization, as an implicit CUDA stream join silently corrupts output over long generations. The resulting "flashified" models can be deployed through Superlinked's inference engine for production testing.