PyTorch now supports FlyDSL as an optional backend for AMD MI350-series GPUs, enabling accelerated dense and grouped GEMM, MXFP8/MXFP4 scaled GEMM, fused multi-head attention, RMSNorm, and TopK through existing operator APIs without custom kernel calls1. Reported kernel-level geometric-mean speedups include 1.58x for MXFP8 and 1.68x for MXFP4 over ATen, 1.10x for BF16 dense GEMM over the faster of ATen and Triton, and 4.82x for small-K TopK over ATen. FlyDSL attention kernels ship inside AOTriton starting with the 0.14b series.
PyTorch Integrates FlyDSL Backend for AMD MI350 GPUs
PyTorch adds FlyDSL as an optional backend for AMD MI350-series GPUs, reporting kernel-level speedups up to 4.82x across supported operators.
Was this useful?
The day’s vector, in your inbox.
Vector Wire Daily: one dispatch each publishing day.