PyTorch 2.15 ships a rewritten backend for Nvidia’s NCCL (“Nickel”) collective communications library, aimed at better fault diagnosis and recovery when lots of GPUs work together, The Register reports.
What changed
The old ProcessGroupNCCL could spawn dozens of communicators per process group, leading to intermittent deadlocks and unpredictable performance. Core developer Tristan Rice said it was drowning in technical debt. The new backend enforces one communicator per process group at initialization, which removes the crosstalk and makes failures easier to isolate.
My take
This is the boring plumbing that makes big training jobs survivable. One flaky GPU shouldn’t mean starting over from scratch.
God Bless, and Happy hacking!
