PyTorch has released version 2.15, making its rewritten NCCL backend the default for jobs using NCCL, the communication system used to coordinate NVIDIA GPUs. If a group of GPUs hangs during training, a new flight recorder can save what each rank was doing, helping teams investigate the failure.
The release also adds one-sided data transfers through a NIXL backend and gives developers more control over memory and communication in distributed training. Other updates include faster Apple Silicon inference kernels, CUDA 13.2 as the default wheel, and support for Python 3.15 in torch.compile.
Why it matters
Teams training models across NVIDIA GPUs get diagnostic records when communication hangs, though the release notes do not quantify how much this will reduce recovery time.
