FAQ¶
Access & Authentication¶
How do I get access to PAIS?
Access is granted through research groups. Ask your PI to add you, or contact the PAIS team. See Requesting Access.
My Tuakiri login isn't working.
Make sure you're selecting "University of Auckland" at the Tuakiri identity provider discovery page. If your UoA account is active but PAIS login fails, your group membership may not yet be provisioned — contact the PAIS team with your UPI.
I lost my API key.
API keys are only shown once at generation time. If you lose one, revoke it in the Portal and generate a new one. See Getting an API Key.
Inference API¶
Is the PAIS API compatible with OpenAI's API?
Yes — it implements the OpenAI Chat Completions and Embeddings API spec. Any client or library that supports OpenAI's API works with PAIS by changing the base_url and api_key.
Which models are available?
llama-3-1-8b (chat) and qwen3-vl-embedding-8b (embeddings). Use client.models.list() to confirm available models programmatically. See Available Models.
The API is returning 503 / timing out.
The vLLM inference process may be warming up or the GPU may be fully loaded. Wait 30 seconds and retry. If the issue persists, check the PAIS status page or contact the team.
Can I use GPT-4o or Claude through the PAIS API?
Not directly, but the platform can broker requests to cloud-hosted models (GCP Vertex AI, AWS Bedrock) via LiteLLM. This is configured at the platform level — contact the PAIS team.
Notebooks & GPU¶
My notebook server won't start.
Check the research group ResourceQuota — you may be over the GPU or memory limit. Try reducing the requested resources, or coordinate with your group. See GPU Access & Scheduling.
Files I wrote in my notebook are nobody-owned when I access them elsewhere.
This is the VAST root-squash behaviour. Run chmod -R 755 /path/to/your/files from a terminal in your notebook, or contact the PAIS team for systematic cases. See Storage.
How do I use a custom Docker image for my notebook?
The cluster uses buildah (not Docker) for container builds. Contact the PAIS team to build and register a custom notebook image.
ML Pipelines¶
The KFP pipeline fails with "MLMD connection refused".
This is likely the known Kubeflow / Istio DestinationRule issue. The ml-pipeline-mysql DestinationRule must use tls.mode: DISABLE. Contact the PAIS team to re-apply the patch. See the CLAUDE.md in the repo root for details.
Where are MLflow artifacts stored?
Artifacts are stored on VAST via the mlflow-artifacts PVC. They persist indefinitely (Retain policy) and are accessible to all members of the experiment's namespace.
Platform & Costs¶
What does it cost to use PAIS?
There is no per-token or per-hour charge for researchers. The platform is funded through the university and GPU hardware contributions from research groups. OpenCost tracks resource usage for internal reporting.
Can my research group add GPUs to PAIS?
Yes — the condominium model allows groups to contribute GPU nodes and receive priority access to their contributed capacity. Contact the PAIS team to discuss hardware requirements.
Where is my data stored? Is it private?
All data stays on university-controlled hardware (no data leaves to cloud providers for standard API calls). Each research group's VAST storage is isolated by Kubernetes RBAC and NetworkPolicy. Contact the PAIS team if you have specific data sovereignty requirements.