CUDA virtual memory management with HOST NUMA might require special system capabilities.
trivialfis opened this issue · 1 comments
trivialfis commented
@hcho3 noticed that some virtual memory-related calls might fail in a CI container environment.
trivialfis commented
Closing this as I haven't seen any reproducible example yet. I have been using docker with GPUs for a while. Also, host numa is used by nccl as well.