Logging for cuda usage could be a little confused when specifying gpu-id
Vikho opened this issue · comments
Vikho commented
When I use the training script and specify export CUDA_VISIBLE_DEVICES=1
, the logging shows _cuda:0_
, which is inconsistent with the actual usage (indeed using cuda:1
). This may cause some confusion for those who are new to GPU deep learning.
Ross Wightman commented
using CUDA_VISIBLE_DEVICES is a low level alteration, it changes the device enumeration visible to the application so 0 is what the application sees.