PyTorch 2.15 Brings Rewritten NCCL Backend and Broader Hardware Support

Published 29 October 2026

PyTorch 2.15 makes its rewritten NCCL backend the default, adds one-sided distributed transports, improves FSDP2 and pipeline parallelism, and accelerates inference across NVIDIA, AMD, Apple, and Intel hardware. The release also adds CUDA 13.2 and ROCm 10.0 wheels, Python 3.15 support, and expanded profiling tools.