PyTorch 2.15 updates ProcessGroupNCCL for NCCL faults

PyTorch 2.15 introduces a rewritten backend for Nvidia's Collective Communications Library designed to resolve intermittent deadlocks and improve fault diagnosis in multi-GPU training clusters. The updated ProcessGroupNCCL backend enforces a strict one-communicator-per-process-group rule, incorporates a flight recorder for capturing operational failures, and allows third-party orchestrators to bypass or replace lagging workers. Additional updates in the release include asynchronous one-way data movement and improved support for Apple hardware and Nvidia Blackwell GPUs. Open article →