- Amazon Linux
- Amazon Linux 2
- Redhat Enterprise Linux 7.0
- Ubuntu 16.04 LTS
- CentOS 7
This release introduces changes required to support NCCLv2.4 and fixes race condition during connection establishment by removing FI_SOURCE requirement.
New Features:
- Support NCCL provided MR register/deregister APIs.
Bug Fixes:
- Remove FI_SOURCE requirement for providers.
- Fix travis CI to build with NCCLv2.4.
Testing: The plugin has been tested with following libfabric providers:
- tcp;ofi_rxm
- sockets
- verbs;ofi_rxm
This release makes improvements to the building and CI infrastructure. It also includes several bug fixes. Details below:
New Features:
- Change build system to use autoconf, automake and libtool
- Add support for continuous integration using Travis CI
- Add official support for libfabric v1.7.x
Bug Fixes:
- Remove hard-coded CUDA path when linking test binaries.
- Provide request contexts to all libfabric send/recv calls
- Readme updates and other minor fixes
Testing: The plugin has been tested with following libfabric providers:
- tcp;ofi_rxm
- sockets
- verbs;ofi_rxm
- psm2
- efa;ofi_rxr
First public commit as part of preview announcement
AWS OFI NCCL supports NCCL v2.3.7+ and requires libfabric v1.6.x+. Please note that current master of libfabric is broken for rxm providers and would require PR-4641.
The plugin has been tested with following libfabric providers:
- tcp;ofi_rxm
- sockets
- verbs;ofi_rxm