Zyphra releases ZYFA attention kernels for AMD MI355XMachine translation

Zyphra has released open-source ZYFA for long-context prefill on AMD Instinct MI355X. In tests using real model activations, its FP8 kernels delivered 1.16–1.76 times AITER's throughput. Support is limited to forward prefill on MI355X, with no decode or paged KV.