PyTorch 2.15 makes rewritten NCCL backend default, speeds LLM decode on Apple and Blackwell

PyTorch 2.15 promotes a rewritten NCCL backend to default and accelerates LLM decoding across hardware stacks—from NVIDIA Blackwell to Apple Silicon. The release, spanning 3,117 commits from 583 contributors, also adds FP8 support on Apple, 4-bit decode optimizations on Blackwell, and expanded platform coverage for AMD and Intel GPUs.

• The rewritten NCCL backend is now default, with a flight recorder that dumps on hangs and rollback via TORCH_DIST_USE_NCCL2=0 or nccl-legacy
• Apple Silicon gains FP8 (float8_e4m3fn, _scaled_mm) and a grouped_mm decode fast path delivering 1.32–1.34x tokens/s on MoE decode
• NVIDIA Blackwell GPUs get faster 4-bit (NVFP4) LLM decoding through fewer, better-overlapped kernel launches in NVGEMM
• FSDP2 and pipelining add fp32 for precision-sensitive parameters, CUDA-graph-captured gradient accumulation, and gradient reduction overlapped with compute
• Platform support expands to FlyDSL kernels on AMD gfx950, Tensor Data Mover for gfx1250 (MI450), ROCm 10.0 wheels, simpler SYCL kernel authoring on Intel GPUs, and CUDA 13.2 as the default wheel

Bottom line Upgrade if you train or serve LLMs on Apple Silicon or Blackwell, or need the new NCCL flight recorder; otherwise, plan your next regular update.