Just install this along with Backend.AI agents, using the same virtual environment. This will allow the agents to detect CUDA devices on their hosts and make them available to Backend.AI kernel sessions.
$ pip install backend.ai-accelerator-cudaThis open-source edition of CUDA plugins support allocation of one or more CUDA devices to a container, slot-by-slot.
| Backend.AI Agent | CUDA Plugin |
|---|---|
| 20.09.x | 2.0.0 |
| 20.03.9 ~ | 2.0.0 |
| 20.03.0 ~ 20.03.8 | 0.14.x |
| 19.09.17 ~ | 0.13.x |
| 19.06.x, 19.09.0 ~ 19.09.16 | 0.11.x, 0.12.x |
| 19.03.x | 0.10.x |
In the versions released after the above matrix, the agent will set the required version range of this plugin as an extra requirements set "cuda".