PyTorch 2.15 makes rewritten NCCL backend default, adds NVFP4 decode on Blackwell and FP8 on Apple Silicon
PyTorch · x · 2026-11-05
PyTorch will host a live release Q&A on November 5 at 11 AM PT with Meta's Tristan Rice, Natalia Gimelshein, Huy Do and Jane Xu, moderated by Chris Gottbrath.
Key changes in PyTorch 2.15:
- The rewritten NCCL backend is now the default, with one-sided transports via a new NIXL backend
- FSDP2 and pipelining gain finer memory and communication control
- Faster NVFP4 4-bit LLM decoding on NVIDIA Blackwell GPUs
- FP8 support and faster decode on Apple Silicon
- Broader platform support
Related event: PyTorch 2.15 Released with New NCCL Backend and NVFP4 Support(2 posts)→
More from Infra
- SkyPilot Tokens lists Kimi K3.1 Pro day-0: 2.8T params, 1M context, 40% cheaper — skypilot_org · 2026-11-05
- Goldman restricts traders' AI tools while Apple cuts internal AI budgets in same week — GaryMarcus · 2026-11-05
- Polymarket puts 29% odds a state enacts data center moratorium by year-end — Polymarket · 2026-11-05
- Running Qwen 27B Q4 at 0.7 tok/s on a 4GB-VRAM laptop as a main coding model — Felix_455-788 · 2026-11-05
- Stanford's Joules-to-GDP framework measures how much economic value AI produces per joule of energy — HazyResearch · 2026-11-05
- Beff backs space data centers: policy will only accelerate the Dyson swarm — beffjezos · 2026-11-05