PyTorch updates NCCL Extensions pin from 0.2
A change was made in the PyTorch trunk to update the NCCL Extensions version from 0.2. The new NCCL-EP v0.3 introduces new features such as Vera Rubin support and Attention–FFN Disaggregation (AFD).
A change was made in the PyTorch trunk to update the NCCL Extensions pin from 0.2. Additionally, nccl-ep will not be built when the architecture list is not specified as sm_90-or-newer.
The newly released NCCL-EP v0.3 adds Vera Rubin support, optimized low-latency NVLink dispatch and combine kernels, broader hidden-dimension support, and zero-copy DeepSeek FP8 dispatch. It also introduces Attention–FFN Disaggregation (AFD), which separates attention and expert processing across rank groups, including placement on separate RDMA-connected nodes.
Furthermore, it adds high-throughput pull-based dispatch and push-based combine for eligible NVLink configurations, MXFP8 combine, and opt-in Compute Fabric Transport (CFT) for supported Blackwell-or-later NVLink configurations. These features retain compatibility with applications compiled against EP v0.1 and v0.2.
This item is a summary. The full, checkable version is at the original source:
Read the original at PyTorch ↗