GPU Server Usage (`GPU_SERVER`)

This page documents practical usage of the lab GPU server for ADAMACS workflows.

Server and connection

Host role: GPU compute server for model-heavy workloads (for example DLC, denoising, cascade)
Permanent IP: <GPU_SERVER_IP>
Legacy note: ignore older references to <OLD_GPU_SERVER_IP>

Connect via SSH:

ssh <your_username>@<GPU_SERVER_IP>

Account and security basics

Keep credentials private and do not store passwords in notebooks, repos, or docs.
Prefer SSH keys after first login.
Do not modify system GPU drivers or CUDA installation.
If credentials were shared by email, move them to a password manager and rotate when needed.

Usage policy (important)

Use GPU, CPU, and RAM responsibly (no automatic quota management is assumed).
Stop jobs and close Jupyter kernels when done.
Do not run uncontrolled 24/7 jobs.
Do not occupy all GPUs without coordination.
Treat this host as compute-first, not storage-first.

Storage policy

Home directories are limited and should not hold large datasets.
Use /mnt/data (about 7 TB NVMe) only for temporary, active compute data.
Store canonical datasets on mounted network shares.
Clean temporary outputs after runs complete.

Recommended workflow (headless-first)

The server is intended as a headless compute machine.

Recommended:

Develop GUI-heavy steps locally.
Use VS Code Remote SSH or terminal SSH for execution on <GPU_SERVER_IP>.
Run Jupyter on the server only as needed, then shut it down.

Official VS Code Remote SSH docs:

https://code.visualstudio.com/docs/remote/ssh

Optional GUI path (when required)

Preferred default is local GUI + remote compute. If remote GUI is unavoidable, use SSH X-forwarding.

macOS X-forwarding

Install XQuartz: https://www.xquartz.org/
Connect with X-forwarding:

ssh -X <your_username>@<GPU_SERVER_IP>

Run your GUI app on the server (example shown for DLC).

For GUI mode in this setup, a lab example is:

pip install "deeplabcut[gui]==2.3.8"
python -m deeplabcut

DeepLabCut docs:

https://deeplabcut.github.io/DeepLabCut/

DeepLabCut lab note (internal)

Lab-specific walkthrough (with and without remote GUI):

<INTERNAL_LAB_NOTE_URL>

Note:

request the current internal note URL from lab ops.

Copy-paste email templates

Use these as quick templates for future communication.

Access request template

Subject: GPU server account request (<GPU_SERVER_IP>)

Hello,

Could you please create GPU server access for me on <GPU_SERVER_IP>?
Preferred username: <username>
Use case: <short use case, e.g. DeepLabCut training/inference for ADAMACS>

I confirm I will follow server policy:
- no driver/CUDA changes
- responsible shared GPU usage
- temporary data only on /mnt/data

Thank you.

Acknowledgement template

Subject: Re: GPU server access

Dear <Name>,

Thank you very much for setting up the account and for the detailed instructions.

Best regards,
<Your Name>

Resource coordination template

Subject: Planned long GPU run on <GPU_SERVER_IP>

Hello,

I plan to run a longer GPU job:
- start: <date/time>
- expected duration: <hours>
- expected GPU usage: <N GPUs>
- workflow: <DLC/denoising/cascade/etc.>

Please let me know if this conflicts with other planned usage.

GPU Server Usage (GPU_SERVER)