Documentation
¶
Overview ¶
workflow_render implements the "nvcrectl workflow render" subcommand.
Index ¶
- func LoadEmbeddedNodes(platform, gpuArch string) ([]corev1.Node, error)
- func NewK8sClient(cf *kubeconfig.ConfigFlags) (client.Client, error)
- func NewK8sWatchClient(cf *kubeconfig.ConfigFlags) (client.WithWatch, error)
- func NewWorkflowCommand() *cobra.Command
- func PrintDryRunSummary(header string, results []DryRunResult)
- func ResolveWorkflow(workflow *nvcrev1alpha1.Workflow, nodes []corev1.Node) (*renderMetadata, error)
- func ResolveWorkflowForPlatform(workflow *nvcrev1alpha1.Workflow, nodes []corev1.Node, platformName string) (*renderMetadata, error)
- func SetRenderAnnotations(workflow *nvcrev1alpha1.Workflow, meta *renderMetadata)
- func SyntheticProviderID(platform string) string
- type DryRunResult
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func LoadEmbeddedNodes ¶
LoadEmbeddedNodes loads a mock node from the embedded YAML templates for the given platform and GPU architecture combination.
func NewK8sClient ¶
func NewK8sClient(cf *kubeconfig.ConfigFlags) (client.Client, error)
NewK8sClient builds a controller-runtime client from the resolved ConfigFlags (kubeconfig, context, and every other kubectl-standard connection/auth flag).
func NewK8sWatchClient ¶
func NewK8sWatchClient(cf *kubeconfig.ConfigFlags) (client.WithWatch, error)
NewK8sWatchClient builds a watch-capable controller-runtime client from the resolved ConfigFlags.
func NewWorkflowCommand ¶
NewWorkflowCommand returns the "workflow" parent cobra command with the "render" subcommand attached. This was previously newWorkflowCommand in root.go.
func PrintDryRunSummary ¶
func PrintDryRunSummary(header string, results []DryRunResult)
PrintDryRunSummary writes a human-readable summary table to stderr. Output goes to stderr so that YAML on stdout remains pipeable.
func ResolveWorkflow ¶
func ResolveWorkflow(workflow *nvcrev1alpha1.Workflow, nodes []corev1.Node) (*renderMetadata, error)
ResolveWorkflow applies overrides to a Workflow in-place and returns detection metadata. The workflow's spec is mutated directly.
func ResolveWorkflowForPlatform ¶ added in v0.3.0
func ResolveWorkflowForPlatform( workflow *nvcrev1alpha1.Workflow, nodes []corev1.Node, platformName string, ) (*renderMetadata, error)
ResolveWorkflowForPlatform is ResolveWorkflow with the platform forced when platformName is non-empty: the dry-run paths pass --platform through here so override matching and the recorded metadata honor the flag instead of silently reverting to node-based detection. GPU architecture is always detected from the nodes. An empty platformName detects the platform too.
func SetRenderAnnotations ¶
func SetRenderAnnotations(workflow *nvcrev1alpha1.Workflow, meta *renderMetadata)
SetRenderAnnotations stores detection metadata as annotations on the Workflow.
func SyntheticProviderID ¶
SyntheticProviderID returns a providerID prefix that detectPlatform maps back to the given platform string. Used for offline rendering.
Types ¶
type DryRunResult ¶
type DryRunResult struct {
Resource string `json:"resource"`
Valid bool `json:"valid"`
Error string `json:"error,omitempty"`
Warning string `json:"warning,omitempty"`
}
DryRunResult records the outcome of validating a single resource via server-side dry-run.
func DryRunCreate ¶
func DryRunCreate(ctx context.Context, c client.Client, namespace string, spec *nvcrev1alpha1.WorkflowSpec, nodes []corev1.Node, synthesizedProducts []string) ([]DryRunResult, error)
DryRunCreate validates the resolved Job, workload, and dependencies via server-side dry-run without persisting anything.
Validation order: Dependencies → Job → Workload. Dependencies are validated first so that we can detect when a workload failure is caused by a missing dependency that would be created at runtime (e.g. a TrainingRuntime referenced by a TrainJob). Such failures are reported as warnings rather than errors.
synthesizedProducts comes from the same discovery call that produced nodes (controller.DiscoverTargetNodesWithSynthesized). It names the nodes whose nvidia.com/gpu.product label was filled in from gpu.nvidia.com ResourceSlices rather than read off the stored Node, which the affinity build below has to know about. Pass nil when the caller has no such list; the only cost is that a DRA-only fleet previews a term the reconcile will not emit.
Source Files