Custom Catalog Entries | NVIDIA Cluster Readiness Engine

The catalog is extensible — add a YAML file to pkg/catalog/entries/ to register a custom certification category.

File layout

pkg/catalog/entries/
  <domain>/
    <variant>.yaml     ← new file

The catalog loader discovers entries by scanning this directory tree at startup. No registration step or code change is required.

YAML structure

A catalog entry YAML has up to four top-level sections:

1 # Dependencies: Kubernetes resources applied before the job runs
2 # (e.g., a TrainingRuntime defining the MPI/PyTorch topology)
3 dependencies:
4   - apiVersion: trainer.kubeflow.org/v1alpha1
5     kind: TrainingRuntime
6     metadata:
7       name: my-variant-runtime
8     spec:
9       # ...
10 
11 # Job template: defines the workload
12 jobTemplate:
13   spec:
14     workload:
15       trainJob:
16         runtimeRef:
17           kind: TrainingRuntime
18           name: my-variant-runtime
19         trainer:
20           image: nvcr.io/nvidia/pytorch:26.01-py3
21           args:
22             - my-benchmark-command
23           numNodes: {{ .NodesPerJob }}
24           numProcPerNode: {{ .GpusPerNode }}
25 
26 # Orchestration: controls how jobs are grouped and scheduled
27 orchestration:
28   execution:
29     timeoutPerJob: 30m
30   iterations: 1
31 
32 # Overrides: platform- or GPU-specific patches (optional)
33 overrides:
34   - when:
35       platform:
36         equals: aws
37     jobTemplate:
38       spec:
39         workload:
40           trainJob:
41             trainer:
42               image: public.ecr.aws/my-org/my-benchmark:latest

Template variables like {{ .NodesPerJob }} and {{ .GpusPerNode }} are resolved at render time from the detected cluster and any flag overrides.

Minimal example

1 # pkg/catalog/entries/my-domain/my-variant.yaml
2 dependencies:
3   - apiVersion: trainer.kubeflow.org/v1alpha1
4     kind: TrainingRuntime
5     metadata:
6       name: my-variant-runtime
7     spec:
8       mlPolicy:
9         numNodes: 1
10         torch:
11           numProcPerNode: 1
12       template:
13         spec:
14           replicatedJobs:
15             - name: node
16               template:
17                 spec:
18                   template:
19                     spec:
20                       containers:
21                         - name: node
22                           image: nvcr.io/nvidia/pytorch:26.01-py3
23                           resources:
24                             limits:
25                               nvidia.com/gpu: "{{ .GpusPerNode }}"
26 
27 jobTemplate:
28   spec:
29     workload:
30       trainJob:
31         runtimeRef:
32           kind: TrainingRuntime
33           name: my-variant-runtime
34         trainer:
35           args:
36             - my-command
37 
38 orchestration:
39   execution:
40     timeoutPerJob: 30m
41   iterations: 1

Verify

After adding the file, verify it appears in the catalog and renders correctly:

$ nvcrectl certification list-categories
$ 
$ nvcrectl certification render \
>   --platform aws \
>   /tmp/my-cert.yaml

Check the rendered Workflow for correct resource requests, env vars, and override annotations.