Open-Box: Punching Above the VRAM Weight Class
A reproducible research program for running larger effective models on small GPUs.
We present a clean-room, reproducible implementation of native sparse
attention (NSA) with custom Triton kernels, gated for correctness against a
reference at every step and validated from Ampere (sm_86) to
Hopper (sm_90). We report a +53% end-to-end training throughput
optimization, a −45% model-VRAM reduction, and an honest performance
analysis showing the selection branch dominates the training step — a
finding independently corroborated by concurrent NSA-kernel work. The paper
situates NSA as the first validated component of a broader program: decouple a
model's effective capability from what must reside in GPU VRAM.
Everything needed to rebuild the results is public on GitHub.