docs
docs copied to clipboard
[DOC] Add NCCL notice for Docker containers running 12.8 in the Release Selector
Report needed documentation
Report needed documentation
@dantegd and team discovered that 25.02 ARM containers on CUDA 12.8 have a docker permission challenge, requiring them to run the flag NCCL_CUMEM_HOST_ENABLE=0 inside the container before running some multi GPU algorithms.
Describe the documentation you'd like
Can we add a note between the selector and the code command, if the user selects Stable, Docker, and CUDA 12.8 options, detailing the need to add the flag above?
@dantegd @bdice @aravenel for awareness
Linking PR https://github.com/rapidsai/docker/pull/735 in case we prefer to enable the variable in the container and not the instructions (or both)