Skip to content
Weld Weld Docs

Windows PC as a build and GPU node

Most teams have one Windows PC with a strong NVIDIA card. Joined as a node with the build and gpu roles, it packages Windows builds for everyone (including people on Macs) and runs concept and image-to-3D generation for the asset pipeline.

Time: about 30 minutes of your attention; model and engine downloads take longer. You need: Windows 10 or 11, an NVIDIA GPU with 12 GB or more of VRAM (16 GB recommended), roughly 100 GB free for Unreal, packaging and the derived data cache, plus about 70 GB for the AI models, and an Epic Games account.

  1. Join with roles.

    Open PowerShell and run weld node with the roles this PC should have:

    Terminal window
    weld node --hub https://<your-hub> --roles build,gpu

    Already joined from your Mac? Then this is a second machine on the same account: run weld node --hub https://<your-hub> --roles build,gpu, answer yes to “already have an account?” and approve the PC in the browser with your passkey.

    In a terminal, join also offers gpu and build by itself when it sees an NVIDIA GPU or an Unreal install.

  2. Install the prerequisites.

    Join lists what’s missing and the exact command for each: winget for Git, Git LFS, uv and Blender, and the official Claude Code installer. It asks before installing, and gives large downloads such as Blender their own question. Every command is shown as it runs, with the full output in <data dir>\logs\prereqs-<time>.log.

    Two things Weld won’t do for you:

    • The NVIDIA driver is never installed by Weld, only checked and pointed to. Keep it current with the NVIDIA app.
    • Unreal Engine 5.8 needs your Epic login: Weld installs and opens the Epic Games Launcher and prints the steps. Install UE 5.8 from the launcher, then run weld setup again.

  3. Set up GPU generation.

    Terminal window
    weld setup --roles gpu-gen

    This installs what the asset pipeline needs on the GPU: the CUDA 12.4 toolkit (toolkit only; your display driver is left alone), Visual Studio Build Tools, the TRELLIS.2 environment and model weights, BiRefNet and the FLUX.1 [schnell] concept model.

    FLUX.1 [schnell] and DINOv3 are gated on Hugging Face: accept their licenses on their Hugging Face pages, sign in to Hugging Face as the message shows, and run weld setup --roles gpu-gen again. Until then you’ll see WELD-E4213. When everything is in place, the command reports “Already installed” for each part.

  4. Check what the hub sees.

    Terminal window
    weld node caps
    weld node role list
    weld game engines

    caps should list your GPU and VRAM, Unreal 5.8, Blender, git and Claude. role list should include build (add routing roles with weld node role add build, and weld node role add gpu-gen if your team routes asset jobs by that role).

  5. Run the node in the background.

    Terminal window
    weld node install-service

    On Windows this adds a per-user entry that starts the node when you sign in and restarts it if it stops. The PC must be signed in for jobs to run.

  6. Keep it awake.

    Weld keeps the PC awake while a job runs, but it can’t wake a sleeping PC to start one. Set Windows’ power plan so the PC doesn’t sleep while plugged in, at least on build days, and leave it signed in.

  7. Trust your teammates’ jobs.

    Typed jobs (builds, smoke tests, asset jobs) from a teammate run on this PC once you trust their key. Compare fingerprints with them over a channel you trust (a call, a chat you know is them):

    Terminal window
    weld node keys          # on both machines; compare the fingerprints
    weld node trust sam

    shell jobs from others still wait for you to run weld jobs approve <id> on this PC.

  8. Smoke-test it.

    From your Mac (or here):

    Terminal window
    weld asset make "a wooden crate" --gpu remote --yes   # --yes approves every gate, for this test only
    weld jobs list

    And, inside a game repo, a packaging dry run on this PC:

    Terminal window
    weld ship --channel windows --dry-run

Optional: run your own asset requests from the web

Section titled “Optional: run your own asset requests from the web”

If you want “Request asset” on the dashboard to run on this PC without anyone starting it by hand:

Terminal window
weld node policy allow-web-assets --per-hour 2 --per-day 6 --max-tris 50000

Only typed asset jobs for your own requests run this way, never shell or agent runs, within those caps. It’s off by default and only you can turn it on, in your own terminal. weld node policy deny-web-assets turns it off again.

When Unreal is busy

Section titled “When Unreal is busy”

GPU jobs wait while Unreal Editor, UnrealBuildTool, AutomationTool or shader compilers are running on the PC, so generation and a build don’t fight over the card. On 16 GB cards TRELLIS.2 runs its 1024 cascade in low-VRAM mode; if Windows starts paging GPU memory to system RAM, Weld warns and uses lighter settings next time.

Next

Section titled “Next”