backend.ai-accelerator-cuda-mock 23.3.0a1

Backend.AI Accelerator Plugin for CUDA (Mockup)

backend.ai-accelerator-cuda-mock

A mockup plugin for CUDA accelerators

This plugin deceives the agent and manager to think as if there are CUDA devices. The configuration follows cuda-mock.toml placed in the same location of agent.toml. Please refer the sample configurations in the configs/accelerator directory and copy one of them as a starting point.

The statistics are randomly generated in reasonable ranges, but it may seem like "jumping around" because there is no smoothing mechanism of generated values. The configurations for fractional/discrete mode, fraction size, and device masks in etcd are exactly same as the original plugin.

The containers are created without any real CUDA device mounts but with BACKENDAI_MOCK_CUDA_DEVICES and BACKENDAI_MOCK_CUDA_DEVICE_COUNT environment variables. Since the manager does not know if the reported devices are real or not, you can start any CUDA-only containers (but of course they won't work as expected).