PyTorch 2.15, released October 28, makes a rewritten NCCL backend the default for every process group, adding a flight recorder that dumps on hangs and rollback paths via TORCH_DIST_USE_NCCL2=0 or nccl-legacy1. The release introduces one-sided transports for torch.distributed with a NIXL backend, faster 4-bit NVFP4 LLM decoding on NVIDIA Blackwell GPUs, and FP8 support on Apple Silicon with float8_e4m3fn and _scaled_mm. Platform additions include ROCm 10.0 wheels, AMD gfx950 and gfx1250 support, CUDA 13.2 as the default wheel, and torch.compile on Python 3.15. The release comprises 3,117 commits from 583 contributors.
PyTorch 2.15 Ships Rewritten NCCL Backend, Blackwell NVFP4 Decoding, Apple Silicon FP8
PyTorch 2.15, released October 28, defaults to a rewritten NCCL backend, adds NVFP4 decoding on Blackwell GPUs, FP8 on Apple Silicon, and ROCm 10.0 support.
Was this useful?
The day’s vector, in your inbox.
Vector Wire Daily: one dispatch each publishing day.