Reproduce “FLAMIA: FLAsh MIcroscaling Attention for Fast and Causally Safe Training on Blackwell” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fedor Velikonivtsev, Mikhail Koltakov, Denis Kuznedelev
FLAMIA reduces the cost of training attention on packed, variable-length documents using Blackwell’s low-precision arithmetic. It reuses an E4M3 quantized payload across forward and transposed backward calculations, with safeguards for causal dependencies and numerical consistency. On nine evaluated causal packed-document workloads on B200 and B300, its MXFP8 path had the lowest combined attention-call latency; this is not a whole-model speedup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.