PyTorch 2.15 ships with default NCCL rewrite, NVFP4 decode on Blackwell
PyTorch · x · 2026-10-29
PyTorch 2.15 is out with 3,117 commits from 583 contributors:
- The rewritten NCCL backend is now the default, with new one-sided transports via a NIXL backend
- FSDP2 and pipelining gain finer memory and communication control
- Faster 4-bit (NVFP4) LLM decoding on NVIDIA Blackwell GPUs
- FP8 support and faster decoding on Apple Silicon, plus broader platform support
A live Q&A with Meta core developers is scheduled for Nov 5, 11 AM PST.
More from Infra
- Oracle targets $30B+ free cash flow by FY29 after burning $23.7B on AI data centers — rohanpaul_ai · 2026-10-29
- EagleCorp benchmarks Nemotron 3.5 Super VL at 2.74x vanilla single-request throughput — NVIDIAAI · 2026-10-29
- Video Demo: Running DeepSeek V4.1 Flash Locally and What It Costs — RelevantEmergency707 · 2026-10-29
- Morgan Stanley projects AWS revenue to nearly triple to $374B by 2028 — Beth_Kindig · 2026-10-29
- Culper Research Goes Short on Cerebras, Claims Wafer Chips Break Down in ~2 Years — StefanoGogioso · 2026-10-29
- AWS Blueprint for an Enterprise Agentic AI Platform on Bedrock AgentCore — AWS ML Blog · 2026-10-29