PyTorch 2.15 Release Blog

PyTorch 2.15 makes the rewritten NCCL backend the default for every nccl process group, with TORCH_DIST_USE_NCCL2=0 or nccl-legacy as rollbacks, and adds a flight recorder that dumps on hangs. It also drops Python 3.10 and macOS 14 support, and changes FSDP2 gradient dtype defaults. The release adds NIXL one-sided transports, FP8 and faster decode on Apple Silicon, and CUDA 13.2 as the default wheel. The MPS speedups, such as 1.32–1.34x MoE decode on an M3 Ultra, are the project's own figures and could not be independently checked.