joeasycompute/infra ? reverse-engineered prompt

Build me a bash based GPU node provisioning toolkit for Ubuntu 22.04, 24.04, and 26.04. I want it to handle both NVIDIA CUDA machines and AMD ROCm machines, mainly for fresh compute nodes where the goal is to get the GPU stack installed, get Docker working, run a solid validation pass before the node goes into service, and keep useful operator scripts around for inventory, network checks, power settings, and monitoring.

Please make the main experience an end to end provisioning flow that can survive reboots and pick up where it left off automatically, with logs and state saved in a clear place under opt. I also want the same steps available as separate manual scripts for people who want more control.

Include acceptance tests for GPU, disk, CPU, RAM, and network, plus small utilities for things like GPU power limits, resetting GPU settings, and optionally disabling selected NVIDIA GPUs at boot. Keep it practical and ops friendly for real machines. If you need exact install details, look up current Ubuntu docs online.