Understand this story
On October 29, 2026, the PyTorch Foundation released PyTorch 2.15. According to information published on the PyTorch Blog the same day, this version switches all process groups that use nccl to the rewritten NCCL backend by default, and includes a flight recorder that dumps on-site information when collective communication hangs, with fallback available via TORCH_DIST_USE_NCCL2=0 or nccl-legacy. This version also adds float8_e4m3fn and _scaled_mm support for Apple Silicon, and the new grouped_mm improves MoE decoding throughput by 1.32–1.34x on M3 Ultra. The default wheel is upgraded to CUDA 13.2, while a CUDA 13.4 wheel is also provided with experimental support for NVIDIA Rubin GPUs; a ROCm 10.0 wheel is newly added, and Python 3.10 and macOS 14 are no longer supported. There are currently no subsequent developments.
AI-generated from public reports
Latest development · Report
PyTorch 2.15 Released: Rewritten NCCL Backend Becomes Default, Apple Silicon Gains FP8 SupportDevelopments
1 confirmed developments
- Source:PyTorch Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 29
PyTorch BlogFirst-party
PyTorch 2.15 Released: Rewritten NCCL Backend Becomes Default, Apple Silicon Gains FP8 Support
The PyTorch Foundation has released PyTorch 2.15. All process groups using nccl now default to the rewritten NCCL backend, accompanied by a flight recorder that dumps on-site information when collective communication hangs, and can be reverted via TORCH_DIST_USE_NCCL2=0 or nccl-legacy. This version also adds float8_e4m3fn and _scaled_mm support for Apple Silicon, and the new grouped_mm improves MoE decoding throughput by 1.32–1.34x on M3 Ultra. The default wheel is upgraded to CUDA 13.2, while a CUDA 13.4 wheel is also provided with experimental support for NVIDIA Rubin GPUs; a ROCm 10.0 wheel is newly added, and Python 3.10 and macOS 14 are no longer supported.