Running multiple instances of pyflex on different GPU's

[alissamueller]

Hi!
The bottleneck of my experiments with softgym is sampling. I have a few environments in an async vector env, however, I would like to increase the amount by using multiple GPUs. I tried setting the desc.deviceIndex [1] to one of the available GPU's, but it always seems to reset to the default during the NvFlexInit.

When I set the EGL_GPU environment variable to the available GPUs I get the message that multiple devices are found, but there is still only one being used.

Is there a way to utilize multiple GPUs for sampling?
Thank you for any help you can offer.

[¹]

softgym/PyFlex/bindings/pyflex.cpp

Line 144 in d0a9d7d

desc.deviceIndex = g_device;