PyTorch 2.15, released on October 28, 2026, makes a rewritten NCCL communications backend the default for distributed training. It is designed to identify failures when multiple GPUs exchange data and to enable recovery without discarding hours of completed work. The release also introduces a backend-agnostic flight recorder that automatically captures diagnostic state during hangs and timeouts. Faster reconfiguration paths are available, but they require NCCL 2.29 or later, or RCCL 2.30.7 with ROCm. The release contains 3,117 commits from 583 contributors, while fallback options for the previous behavior are scheduled for removal in the next version.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
