PyTorch Monarch Ported to AMD Instinct GPUs via ROCm
July 25, 2026
PyTorch Monarch now supports AMD GPUs, enabling elastic distributed training and dynamic recovery from node failures on ROCm. The system previously demonstrated 96.16% scaling efficiency on a 1024-GPU MI325 cluster training DeepSeekV3-671B.
HOW THIS AFFECTS YOU
●
builderYou can now implement fault-tolerant, large-scale training runs on AMD hardware.
●
researcherYou can scale training for billion-parameter models across ROCm clusters with higher reliability.