afcuda to use same devices order than nvidia-smi
WilliamTambellini opened this issue · comments
William Tambellini commented
As today on a multi gpu machine, the order of gpus is different in AFcuda compared to nvidia-smi
The env var
CUDA_DEVICE_ORDER
does not seem to change anything.
Description
- What problem are you trying to solve?
selecting the gpu to run on - (Optional) API of new function
none, just internally making afcuda follow the same order of gpu than nvidia-smi - (Optional)Are there other libraries that implement this feature?
pytorch
ref
https://stackoverflow.com/questions/52815708/order-of-cuda-devices
syurkevi commented
You may be able to set the desired devices corresponding to the native order by using the cuda interface function setNativeId. ArrayFire does not currently base device order from the cuda environment variables, but instead chooses the order based on predicted hardware performance.