For Developers
Kubernetes Resource Schema
TensorFusion defines 12 custom resources in the API group tensor-fusion.ai, version v1. This page says what each one is for, who normally writes it, and where its field reference is.
How the resources fit together
- Resource hierarchy:
TensorFusionCluster→GPUPool→GPUNode→GPU. You write the cluster and its pools; the controller and the hypervisors create and update nodes and GPUs. - Node provisioning: a
GPUPoolthat provisions its own nodes createsGPUNodeClaimobjects, each built from aGPUNodeClass. - Workloads: a
TensorFusionWorkloaddescribes the vGPU resources of an application and may reference a reusableWorkloadProfile. ATensorFusionConnectionpairs a client Pod with a remote worker. - Policy:
SchedulingConfigTemplate,GPUResourceQuotaandProviderConfighold scheduling policy, namespace quotas and vendor-specific settings.
Infrastructure resources
| Kind | Scope | Written by | Purpose | Field reference |
|---|---|---|---|---|
TensorFusionCluster | Cluster | Admin (the Helm chart creates one) | Top-level object. Holds the GPU pools and the cluster-wide component settings. | TensorFusionCluster |
GPUPool | Cluster | Admin, usually through TensorFusionCluster.spec.gpuPools | A group of GPU nodes with shared capacity, provisioning, oversubscription, QoS pricing and component settings. | GPUPool |
GPUNode | Cluster | Controller | One GPU node: its capacity, allocation and phase. | GPUNode |
GPU | Cluster | Hypervisor | One physical GPU: model, capacity, available resources and the Pods running on it. | GPU |
GPUNodeClass | Cluster | Admin | Cloud instance template for provisioned nodes: OS image, subnets, security groups, disks, tags and user data. | GPUNodeClass |
GPUNodeClaim | Cluster | Controller | A request for one new node: instance type, zone, capacity type and the TFLOPS, VRAM and GPU count it will add. | CRD YAML |
Policy and configuration resources
| Kind | Scope | Written by | Purpose | Field reference |
|---|---|---|---|---|
SchedulingConfigTemplate | Cluster | Admin | Placement strategy, vertical scaling rules, rebalancing and hypervisor rate-limit parameters. Referenced by a pool. | SchedulingConfigTemplate |
GPUResourceQuota | Namespaced | Admin | Namespace-wide totals (TFLOPS, VRAM, worker count) and per-workload maximums and defaults. | GPUResourceQuota |
ProviderConfig | Cluster | Admin (the Helm chart creates the NVIDIA one) | Per-vendor settings: component images, GPU model catalog, partition templates, and the native resource names TensorFusion takes over. | CRD YAML |
Workload resources
| Kind | Scope | Written by | Purpose | Field reference |
|---|---|---|---|---|
TensorFusionWorkload | Namespaced | User, or the webhook from Pod annotations | vGPU requests and limits, QoS, isolation mode and autoscaling of one application. | TensorFusionWorkload |
WorkloadProfile | Namespaced | User or admin | A reusable set of workload settings that workloads and Pod annotations can reference by name. | WorkloadProfile |
TensorFusionConnection | Namespaced | Controller | Binds one client Pod to one remote worker and carries the worker's connection URL in its status. | TensorFusionConnection |
Most users never write these workload resources by hand. Adding workload annotations to a Pod is enough; the webhook creates the rest.
Get the exact schema of your version
The field reference pages on this site are generated from the CRDs and can lag behind a release. GPUNodeClaim and ProviderConfig do not have a page yet. The schema installed in your cluster is always authoritative:
# List the resources
kubectl api-resources --api-group=tensor-fusion.ai
# Describe a field, with type and documentation
kubectl explain gpupool.spec.capacityConfig
kubectl explain tensorfusionworkload.spec --recursive
# Dump the full OpenAPI schema of one CRD
kubectl get crd gpupools.tensor-fusion.ai -o yamlThe source of the CRDs is in the tensor-fusion repository: Go types under api/v1/ and generated YAML under config/crd/bases/.
Further reading
- Helm chart values
- TensorFusion DeepWiki, an auto-generated walkthrough of the codebase
TensorFusion Docs