Windows contributors

Windows devices are part of the broker's compute pool; native Slurm is not a prerequisite. CPU workers use the same Python protocol as Linux and macOS. The CI matrix includes Windows protocol tests; these do not certify every Windows GPU, driver or inference engine.

PowerShell setup

Install Python 3.11+ and clone this repository, then run from its root:

py -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e .
$join = Read-Host "Community enrollment key" -AsSecureString
$env:OPEN_COMPUTE_JOIN_KEY = [System.Net.NetworkCredential]::new('', $join).Password
.\.venv\Scripts\open-compute.exe worker --broker https://YOUR-BROKER-HOST --name windows-desktop

Replace the broker host with your community's HTTPS endpoint. Do not use the administrator token as the enrollment key. Credentials and a completed-result outbox persist beside the worker; protect them with your user's NTFS permissions. Press Ctrl+C to stop. If Windows sleeps, the lease expires and work may be retried. You can also use the browser over HTTPS without installing Python.

Inference

The worker can call an administrator-configured OpenAI-compatible server, including vLLM or SGLang running on a Linux GPU host. This is not a claim that those serving engines run natively on every Windows GPU. Endpoint drivers and engine compatibility are separate from worker protocol support.

Can Windows join Slurm itself?

The supported path in this project is Slurm on Linux. A WSL2 Linux environment is an experimental possibility, not a verified Slurm deployment here. It needs slurmd, cluster authentication, stable addressing, compatible Slurm versions, resource/cgroup configuration and connectivity to the controller and storage. NVIDIA CUDA availability in WSL2 does not establish Slurm GPU/GRES compatibility. Sleep, networking changes and Windows/WSL restarts require explicit testing.

A Linux VM is another CPU option; GPU passthrough depends on the hypervisor and hardware. For predictable managed Slurm GPU nodes, use native Linux. Keep Windows contributors on the native broker worker unless there is a tested operational reason to add the complexity of a Linux VM or WSL2 Slurm node.