Project · Stand up a GPU pod and prove the full stack works

Starting from a cluster with at least one NVIDIA GPU, install the device plugin (or confirm it's running), schedule a pod that requests a GPU, prove access with nvidia-smi, and run a real GPU workload while observing utilization. Produce a short writeup that walks each layer of the stack and how you verified it.

• A working DaemonSet (device plugin) and a node showing nvidia.com/gpu under Allocatable.
• A GPU pod spec plus captured nvidia-smi output proving in-container GPU access.
• A real GPU workload (PyTorch or CUDA sample) with logs and a screenshot of non-idle GPU-Util.
• A one-page writeup mapping node -> device plugin -> request -> runtime -> verification.

• nvidia.com/gpu appears under Allocatable on the GPU node and the device plugin pod is healthy.
• The GPU pod's nvidia-smi output lists the device, driver, and CUDA version.
• The workload demonstrably uses the GPU (GPU-Util > 0 captured during the run).
• The writeup correctly attributes responsibility to each layer (plugin advertises, runtime injects).