Zyphra has announced the release of ZYFA (Zyphra FlashAttention on AMD), an open-source library of FP8 and BF16 prefill attention kernels optimized for AMD Instinct MI355X GPUs. Released under an Apache 2.0 license, ZYFA delivers a 1.16x to 1.76x execution speedup on real model activations compared to AMD’s default AITER library, while dramatically reducing numerical error.

To maximize hardware utilization, ZYFA introduces a fast probability approximation technique dubbed “lean P,” which cuts softmax vector computation cycles per block from 896 to 384. This optimization raises the steady-state matrix pipeline active time from 60% to 86%. On real-activation benchmarks, ZYFA achieved sustained speedups of 1.16x on Mistral-7B, 1.17x on Qwen3-8B, and 1.76x on DeepSeek-V2-Lite.

ZYFA serves as the core attention engine for the Zyphra Optimization Engine (ZOE) behind Zyphra Cloud. By lowering FP8 error by 62% to 78% compared to AITER at head dimension 128, the library demonstrates enhanced precision stability for low-precision inference on AMD hardware.

Why it matters

  • Infrastructure engineers can achieve higher throughput on AMD MI355X hardware using ZYFA’s open-source kernels.

  • Founders running long-context models can lower compute costs via improved FP8 efficiency without losing numerical accuracy.

  • AMD ecosystem developers gain production-grade performance enhancements that narrow the software stack gap with NVIDIA.

Source: marktechpost.com