AITOPAI热报
ProductsSelected

PyTorch 2.15 将 NCCL 新后端设为默认并支持 NVFP4 解码，团队开直播答疑

官方账号PyTorch@PyTorch
Post
View the post
Why we picked it

PyTorch 2.15 更新不少：NCCL 新后端转正、Blackwell 上 4-bit 解码提速、Apple Silicon 支持 FP8，周四有直播答疑，做训练的可以关注。

PyTorch 2.15 把重写后的 NCCL 通信后端设为默认，新增基于 NIXL 的单向传输，并让 FSDP2 和流水线并行获得更细粒度的显存与通信控制。该版本在 NVIDIA Blackwell GPU 上支持更快的 NVFP4 4-bit LLM 解码，同时带来 Apple Silicon 上的 FP8 与解码加速。Meta 的 Tristan Rice、Natalia Gimelshein、Huy Do、Jane Xu 将在 11 月 5 日 11 AM PT 的直播 Q&A 中介绍版本内容并回答社区问题。

Original · PyTorch

Bring your questions about PyTorch 2.15 to our live Q&A this Thursday, November 5, at 11 AM PT.

Tristan Rice (@rice_fry), Natalia Gimelshein, Huy Do, and Jane Xu (@janeyx99) (all @Meta), with moderator Chris Gottbrath (Gottbrath Tech), will share an overview of the release and answer community questions about PyTorch and the new capabilities in the release.

PyTorch 2.15 makes the rewritten NCCL backend the default, adds one-sided transports with a NIXL backend, and gives FSDP2 and pipelining finer memory and communication control. The release also brings faster 4-bit (NVFP4) LLM decoding on NVIDIA Blackwell GPUs, FP8 and faster decode on Apple Silicon, and broader platform support.

Register: https://t.co/opFmSkAUNb

#PyTorch#NCCL#FSDP2#NVFP4#开源模型

Keep reading

Recent stories with the same tag or section
ProductsBuildersMulti-sourceSelected02:06
Ray Launches Special Interest Group for Accelerators

Ray becoming a PyTorch Foundation project shifts accelerator support from a single-vendor effort to a community-governed roadmap, which should reduce fragmentation across GPU and TPU fleets. Teams running distributed training or inference on mixed hardware can follow the SIG on GitHub or Slack to influence scheduling and transport priorities early.

#Ray#PyTorch#NVIDIA#TPU
Event
官方账号PyTorch@PyTorch3 sourcesSource
ProductsBuildersMulti-sourceSelected71°01:52
PyTorch 2.15 ships with 3,117 commits and NCCL rewrite

The NCCL rewrite becoming the default signals a maturing distributed-training stack, and NVFP4 decode targets Blackwell-class GPUs that Singapore and ASEAN data centres are actively procuring. Teams running FSDP2 at scale should review the new memory and communication controls before their next training run. Register for the 5 November Q&A to put migration questions directly to the release engineers.

#PyTorch#NCCL#FSDP2#NVFP4
Event
官方账号PyTorch@PyTorch8 sourcesSource
How-toBuildersSelected00:56
DeepSpeed team to host virtual office hours on October 27

The office hours are a direct channel to influence DeepSpeed's Q4 priorities, which is rare for a framework of this scale. Practitioners benchmarking training frameworks should attend to question the parallelism results against their own workloads, since the claims of matching state-of-the-art frameworks were made only on selected workloads. The benchmark comparisons against rival frameworks set the baseline for distributed training choices going into 2027.

#DeepSpeed#PyTorch#distributed training#open-source models
官方账号PyTorch@PyTorchSource