🔥PyTorch 2.15 Rewrites NCCL Backend With 3,117 Commits From 583 Contributors

Your multi-GPU training jobs are about to get easier to debug

PyTorch 2.15 swaps out its NCCL backend for a rewrite, and if you run distributed training, you'll feel it. The old ProcessGroupNCCL was messy. It could spawn dozens of communicators per process group, and those communicators could interfere with each other. The new backend stops that crosstalk by default. Honestly, that's the kind of cleanup that should have happened years ago. If you're training across multiple GPUs, this changes your debugging life. The new release adds a flight recorder that captures every operation as a series of snapshots you can inspect after something blows up. Stragglers are easier to handle too, since reconfiguration is simpler. The catch? If your program needs parallelism, it now dispatches more process groups, so check your setup before you upgrade blindly. The update landed with 3,117 commits from 583 contributors. The legacy ProcessGroupNCCL ships in the 2.15 package as a rollback option. It gets removed in 2.16, which is planned for early next year. Bottom line: test on 2.15 now, while you still have a safety net.