v1.18.0 (2026-01)
The 1.18.0 release series supports NCCL v2.29.2-1 while maintaining backward compatibility with older NCCL versions (NCCL v2.17.1 and later).
With this release, building with platform-aws requires Libfabric v1.22.0amzn4.0 or greater.
Bug Fixes and Improvements:
- P6-B300 support: added custom tuner decisions for P6-B300
- Improved performance of PAT on P6-B200 by reducing channel count for smaller message sizes
- Changed protocol defaults:
- Default to RDMA protocol on Trn1
- Default to SENDRECV protocol on g7e.8xlarge
- Dynamic platform selection: added feature to enable AWS optimizations at runtime based on presence of AWS NICs. This allows a single plugin binary to be used for both AWS and non-AWS platforms. AWS optimizations can still be disabled at compile-time.
- Fixed support for non-FI_MR_VIRT_ADDR providers in RDMA protocol
- Improved NIC PCIe link speed and width reporting to NCCL
- Redesigned threading model to support multi-threaded applications without requiring a separate Libfabric domain for each thread.
- Fixed support for FI_MR_ENDPOINT providers (supports SENDRECV protocol only) by cleaning up resources in correct order
Checksum (sha512) for the release tarball aws-ofi-nccl-1.18.0.tar.gz:
768767b93d07b4200a2b080ee687881328d8aef5b7428d9e9cdb4c8c6e3b79808c4f31574266612506b144dc1ce508dea46659302d6211f49ca1082ae3e14771 aws-ofi-nccl-1.18.0.tar.gz