Reproduce “ZYFA: Speed and Accuracy of FP8 FlashAttention on AMD CDNA4” 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
Quentin Anthony, Ganesh Bikshandi, Praneeth Medepalli, Stephen Ebert
ZYFA provides causal attention kernels for long-prompt prefill on AMD’s MI355X. It makes FP8 probabilities cheaper and overlaps softmax with matrix operations. On captured Mistral-7B activations at 16K, its FP8 kernel reaches 1.16× AITER’s throughput; at 4K, its single-call arithmetic error averages 0.22× AITER’s on identical quantized inputs. These are kernel-level findings: full-model FP8 attention remains outside the BF16 noise band.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.