Modern Nvidia GPUs support subdivision and direct allocation, however their specification via `CUDA_VISIBLE_DEVICES` is more complex. - [ ] verify that `canary` indeed doesn't properly support multi-device workstations with MIG enabled - [ ] develop scheme for identifying MIG devices - scope to driver version >= 570 - [ ] properly manage allocation/return of MIG GPU resources - [ ] ensure full device is not allocatable resource if using MIG - [ ] option to only use full device, i.e., ignore MIG? --- [NVIDIA docs](https://docs.nvidia.com/datacenter/tesla/mig-user-guide/mig-device-names.html)
Modern Nvidia GPUs support subdivision and direct allocation, however their specification via
CUDA_VISIBLE_DEVICESis more complex.canaryindeed doesn't properly support multi-device workstations with MIG enabledNVIDIA docs