Up to 3.01x inference performance… optimized for AMD ($AMD) and Meta ($META)

AMD and Meta have released results from integrating FlyDSL into PyTorch to optimize AMD GPU compute, reporting up to a 3.01x throughput gain on full request workloads under MXFP4 conditions in a vLLM test on a single AMD Instinct MI355X, measured against the ATen baseline. The technical report, published October 21, details how FlyDSL connects to existing PyTorch operators on the Instinct MI350 series, letting users apply it to supported operations without hand-calling custom kernels. Supported operations include dense and grouped GEMM, attention, MXFP8 and MXFP4 scaled matrix multiplication, RMSNorm, and TopK, with coverage extending to existing PyTorch operators used on AMD GPUs. The result matters because it narrows the software-efficiency gap that has kept AMD's Instinct accelerators behind Nvidia in AI inference, where CUDA's mature kernel ecosystem remains a key lock-in. What to watch next is whether these gains hold across multi-GPU configurations and broader model workloads, and whether upstream PyTorch and vLLM adoption follows.