AITOPAI热报
ProductsOfficial

Zyphra 发布 ZYFA：AMD MI355X 上的 FP8 FlashAttention 内核

Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x Faster

官方一手marktechpost@Michal Sutter
Blog/Media· 3 sources
Read the original
Why we picked it

Zyphra 开源了 MI355X 上的 FP8 attention 内核，比 AMD 官方 AITER 最快快 1.76 倍，误差还更低，搞 AMD 推理的可以看看。

Zyphra 开源了 ZYFA，一套面向 AMD Instinct MI355X 的 FP8 与 BF16 prefill attention 内核。在真实模型激活数据上，其 FP8 内核比 AMD 的 AITER 快 1.16-1.76 倍。在 head dimension 128 下，误差降低 62-78%。代码开源，面向在 AMD GPU 上做大模型推理的开发者。

Image · marktechpost
Original · marktechpost

Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x Faster

Zyphra released ZYFA, open-source FP8 and BF16 prefill attention kernels for AMD Instinct MI355X. On real model activations, its FP8 kernels run 1.16-1.76x faster than AMD’s AITER, with 62-78% lower error at head dimension 128. The post Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x Faster appeared first on MarkTechPost .

  • arXiv: DeepSeek11-05 05:58原文
  • Tri Dao (FlashAttention)11-05 19:01原文
#ZYFA#Zyphra#AMD MI355X#FlashAttention#开源模型

Keep reading

Recent stories with the same tag or section
PapersBuildersSelected11:52
ZYFA Kernels Double FP8 Attention Speed on AMD MI355X

AMD's FP8 attention path has lagged NVIDIA's, partly because the MI355X lacks hardware tracking of matrix instruction completion, making software scheduling critical. ZYFA shows that careful kernel design can close much of that gap, and its open-source release gives engineers a drop-in option for MI355X prefill workloads. Teams benchmarking AMD accelerators against H100 or B200 deployments, including those running DeepSeek-style MLA, should test it against their current AITER setup.

#ZYFA#AMD#MI355X#FlashAttention
Event
arXiv: DeepSeek@Quentin Anthony et al. (4)2 sourcesSource