From a physical GPU node to a running workload.
Build the physical cluster, bring up Kubernetes, advertise its GPUs and explore Run:ai scheduling. Follow the same server through each layer when a workload queues, fails or waits for data.
Independent behavior model. No real Kubernetes cluster, Run:ai license, GPU time or vendor binary is supplied. Read the model and exercise guide · All labs
Choose a reference stack, then launch when ready.
Five minutes to explore. More time when you need it.
Try a hosted lab free for five minutes per verified account. After that, buy an hour for $20 USD. No subscription or automatic renewal.
Use your Wirewalk Institute account. Create an account, confirm your email or reset your password, then return here to sign in. No course purchase is required.
Sign in before launching a lab. Your free time starts only when you launch or resume the lab.
Pause stops time and freezes lab controls. Closing or losing the connection stops time within four seconds. One active lab per account. Progress is saved automatically; viewing it and exporting a session use no time. Access and prepaid-hour details.
Physical stack and scheduler
One inventory, different views: physical GPUs, registered capacity, allocatable resources and current allocations describe the same hardware. Do not add these figures together.
The server underneath the Node
Selected physical server / Kubernetes Node
Bring the stack up in layers
These are teaching stages, not commands run on a host. Physical provisioning uses the BCM-style image model. Kubernetes registration, GPU resource advertisement and PVC readiness are additional gates.
Research volume
A bound PVC does not prove that the storage data path is healthy. An existing mount can keep working through a CSI-controller outage while a new Pod waits to mount.
Submit and investigate training work
Distributed workers require different nodes in this reference plan. GPU IDs are exclusive; MIG, fractional GPUs and time slicing are outside this lab. Native mode can hold a partial group; Run:ai mode waits for a complete placement.
Worker Pod placement
Deserved quota, borrowing and reclamation
This two-project, single-pool policy is a teaching approximation. Priority is compared within a project; it is not a global right to evict another project's protected workload. Deserved quotas cannot oversubscribe physical capacity in this model. Application checkpoints are stored by the modeled training program, not by saving GPU memory.
Find the layer that stopped progress
API or kubelet loss does not automatically stop an existing application container. A missing device plugin blocks fresh GPU placement. RDMA or assigned-device failure can abort a running collective. Storage loss can leave a workload Running while it waits at required I/O.
Build evidence before changing a setting
Investigation sequence
- Complete setup and submit a small validation workload.
- Start one exercise and inspect its broken layer through the console or layer summary.
- Repair that dependency or policy. Failed jobs need an explicit restart.
- Submit a new validation workload and wait for final output publication.
For quota and PDB exercises, first finish or cancel other work. Quota starts a full-cluster preemptible borrower: let it run, inspect it, then submit an eight-GPU beta workload to reclaim deserved resources. The PDB exercise creates protected training: find its actual worker node, drain it, inspect the budget, then deliberately lower the budget or finish/cancel the workload before retrying maintenance.
Follow the same state through the console
Finite, documented commands produce synthetic summaries. No host shell, Kubernetes API server or Run:ai binary is invoked. Run:ai-style listing aliases are teaching shortcuts, not a promise of compatibility with every CLI release.
WIREWALK KUBERNETES / RUN:AI MODEL — synthetic outputs only.
Watch resources and progress separately
Capacity is physical inventory represented to Kubernetes. Allocatable is healthy advertised capacity, not free capacity. Allocations consume it. These are model counts, not DCGM measurements or performance predictions.
Reconciliation and application events
Account state saves automatically, including workloads. The downloadable recipe preserves configuration and policy only; importing it provisions a fresh model and clears workload/evidence progress. Account credit is never exported or imported.