Presentation: GPUs and AI on NRP
- What can a UCLA researcher, student or instructor do with NRP?
- How do I get access, and what is a namespace?
- Which GPUs are there, and how do I run a job on one?
- What AI services does NRP host?
- Know what NRP offers for GPU computing and AI, and what it is not for.
- Get an account, and know who requests a namespace.
- Pick a GPU by the memory a job needs, and know the two ways to use one.
- Name the ways to reach NRP's hosted language models.
Time: 00:00–00:30, then the live demo.
Click the slides, then use the arrow keys or Space to step through. F goes full screen. Open the slides in their own tab.
Download: PowerPoint · PDF
Get access
| Step | Where | Who |
|---|---|---|
| Log in | nrp.ai: Log In → CILogon → UCLA | Everyone |
| Request a namespace | get-started.nrp-nautilus.io | Faculty, staff, postdocs and research scientists. Students: ask your PI or instructor to request one and add you. |
| Chat with the hosted LLMs | nrp-openwebui.nrp-nautilus.io | Anyone with an NRP login |
| Get an LLM API token | nrp.ai/llmtoken | Members of a group with LLM access |
| Run GPU jobs from your laptop | Getting started: kubectl and kubeconfig | Namespace members |
A namespace is your group's workspace on the cluster: a lab, a project or a course. Its admin, usually the PI or instructor, adds members. GPU jobs, data volumes and a JupyterHub all live in a namespace, and what you create there is visible only to its members. Namespaces can be nested under a campus-level group; UCLA does not have a campus-level admin yet, so if your department or group wants that role, get in touch.
NRP suits public, de-identified and synthetic data, simulation, model development and teaching. It is not a place for protected health information, and nothing you send to the hosted LLMs should identify a patient. Keep PHI in UCLA Health's approved environments.
GPUs at a glance
| GPU memory | Examples | Good for | How to ask |
|---|---|---|---|
| Up to 16 GB | RTX 2080 Ti, GTX 1080 Ti, A4000, A100 slices | Teaching, debugging, classic ML, small CNNs | nvidia.com/gpu: 1 |
| 20–24 GB | A10, RTX 3090, L4, RTX 4090 | Most deep learning; models up to ~8B parameters | nvidia.com/gpu: 1, plus memory or model affinity |
| 32–48 GB | L40, V100 32 GB, RTX A6000, A40 | Bigger models and batches | nvidia.com/gpu, or nvidia.com/rtxa6000 / nvidia.com/a40 |
| 80 GB and up | A100 80 GB, H100, H200 | Large training runs and large models | A100: access request; H100/H200 are reserved |
- What's free right now: nrp.ai/viz/resources
- Resource keys and access rules: GPU pods
- Current PyTorch builds need a Turing-or-newer GPU. On a GTX 1080 Ti or V100, use an older image.
AI services
- Hosted LLMs: open-weight models behind
https://ellm.nrp-nautilus.io/v1, an OpenAI-compatible endpoint. The models page says which model suits which task; the catalog rotates. - Four ways in: Open WebUI in the browser; Python or any OpenAI-compatible tool; Jupyter AI in JupyterHub; coding agents, with tested configurations.
- Run your own: vLLM on a GPU in your namespace, training and fine-tuning in GPU jobs, and RAG with NRP's managed Milvus vector database. Worked examples: PEARC26 tutorial, and a physics neural network trained in GPU jobs in CMS HATS.
Labs and classes
A JupyterHub of your own gives every student the same environment in the browser, GPUs from a menu, the AI assistant preconfigured, and shared storage. Instructors: contact us and we'll help you set it up. Students: if your course needs GPUs, ask your professor to get in touch.
Help
- Documentation: nrp.ai/documentation
- Support chat and email: nrp.ai/contact
- Labs, courses, or a UCLA campus group: Daniel Diaz, d4diaz@ucsd.edu
- NRP is a national Kubernetes cluster, free for non-commercial research and education. It is not HIPAA compliant.
- Log in at nrp.ai with your UCLA account. PIs, faculty, staff and postdocs request a namespace; students are added by their PI.
- Ask for the GPU memory a job needs, not a brand name. Explore on a GPU in JupyterHub; run real work as a Job, which releases the GPU when it exits.
- Open-weight LLMs sit behind one OpenAI-compatible endpoint, usable from a browser, Python, Jupyter AI or a coding agent.