dynamo/deploy/operator/internal/controller/dynamographdeploymentreques...

2029 lines
77 KiB
Go

/*
* SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: Apache-2.0
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package controller
import (
"bytes"
"context"
"encoding/json"
"errors"
"fmt"
"io"
"strings"
"text/template"
batchv1 "k8s.io/api/batch/v1"
corev1 "k8s.io/api/core/v1"
apierrors "k8s.io/apimachinery/pkg/api/errors"
"k8s.io/apimachinery/pkg/api/meta"
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
"k8s.io/apimachinery/pkg/apis/meta/v1/unstructured"
"k8s.io/apimachinery/pkg/runtime"
"k8s.io/apimachinery/pkg/types"
"k8s.io/apimachinery/pkg/util/yaml"
"k8s.io/client-go/tools/record"
"k8s.io/utils/ptr"
ctrl "sigs.k8s.io/controller-runtime"
"sigs.k8s.io/controller-runtime/pkg/builder"
"sigs.k8s.io/controller-runtime/pkg/client"
"sigs.k8s.io/controller-runtime/pkg/event"
"sigs.k8s.io/controller-runtime/pkg/handler"
"sigs.k8s.io/controller-runtime/pkg/log"
"sigs.k8s.io/controller-runtime/pkg/predicate"
sigsyaml "sigs.k8s.io/yaml"
configv1alpha1 "github.com/ai-dynamo/dynamo/deploy/operator/api/config/v1alpha1"
dgdv1alpha1 "github.com/ai-dynamo/dynamo/deploy/operator/api/v1alpha1"
nvidiacomv1beta1 "github.com/ai-dynamo/dynamo/deploy/operator/api/v1beta1"
"github.com/ai-dynamo/dynamo/deploy/operator/internal/consts"
commonController "github.com/ai-dynamo/dynamo/deploy/operator/internal/controller_common"
"github.com/ai-dynamo/dynamo/deploy/operator/internal/gpu"
"github.com/ai-dynamo/dynamo/deploy/operator/internal/observability"
)
const (
// Job naming
JobNamePrefixOnline = "profile-online-"
JobNamePrefixAIC = "profile-aic-"
// Container names
ContainerNameProfiler = "profiler"
ContainerNameOutputCopier = "output-copier"
// ServiceAccount
ServiceAccountProfilingJob = "dgdr-profiling-job"
// ConfigMap naming
ConfigMapOutputPrefix = "dgdr-output-"
// Annotation keys
AnnotationAdditionalResources = "dgdr.nvidia.com/additional-resources"
// Annotation keys for v1alpha1 round-trip compatibility.
// The conversion layer stores v1alpha1 fields that have no v1beta1 spec equivalent
// as annotations so the controller can still honour them for converted resources.
AnnotationConfigMapRef = "nvidia.com/dgdr-config-map-ref"
AnnotationOutputPVC = "nvidia.com/dgdr-output-pvc"
// Size limits
MaxAnnotationSize = 250000 // ~250KB, below K8s 256KB limit
// Sidecar image
SidecarImage = "bitnami/kubectl:latest"
// Volume names
VolumeNameProfilingOutput = "profiling-output"
VolumeNameProfilingConfig = "profiling-config"
VolumeNameModelCache = "model-cache"
// Volume paths
ProfilingOutputPath = "/data"
ProfilingOutputFile = "final_config.yaml"
ProfilingConfigMountPath = "/config"
ProfilingConfigDefaultKey = "disagg.yaml"
DefaultModelCacheMountPath = "/opt/model-cache"
// Command line arguments
ArgModel = "--model"
ArgBackend = "--backend"
ArgTTFT = "--ttft"
ArgITL = "--itl"
ArgConfig = "--config"
// Messages
MessageInitialized = "DGDR initialized successfully"
MessageDiscoveringHardware = "Discovering GPU hardware and preparing profiling job"
MessageProfilingJobCreated = "Profiling job created"
MessageAICProfilingJobCreated = "AIC profiling job created"
MessageProfilingInProgress = "Profiling is in progress"
MessageSpecGenerated = "DynamoGraphDeployment spec generated successfully"
MessageSpecAvailable = "Generated spec is available in annotation nvidia.com/generated-dgd-spec"
MessageDeploymentCreated = "DynamoGraphDeployment %s created successfully"
MessageDeploymentReady = "DynamoGraphDeployment %s is ready"
MessageDeploymentDegraded = "DynamoGraphDeployment %s degraded from Ready to %s"
MessageDeploymentDeleted = "DGD %s was deleted. DGDR will not recreate it. Delete this DGDR and create a new one to redeploy."
MessageInvalidState = "Invalid state"
MessageSpecChangeRejected = "Cannot modify spec in phase '%s'. DynamoGraphDeploymentRequest is immutable once profiling starts. Create a new resource with a different name instead."
MessageJobCreationFailed = "JobCreationFailed"
MessageDeploymentCreationFailed = "DeploymentCreationFailed"
MessageResultsRetrievalFailed = "ResultsRetrievalFailed"
MessageGenerationFailed = "GenerationFailed"
MessageAIConfiguratorCheckFailed = "AIConfiguratorCheckFailed"
MessageProfilingCheckFailed = "ProfilingCheckFailed"
MessageConfigMapNotFound = "ConfigMap %s not found in namespace %s"
MessageConfigMapKeyNotFound = "key %s not found in ConfigMap %s"
MessageModelCachePVCNotFound = "model cache PVC %s not found in namespace %s"
)
// shell script template for the output copier sidecar.
//
// The sidecar is a continuous poller that:
// 1. During profiling: polls profiler_status.yaml every 10s, relays phase+message
// to the output ConfigMap so the controller can track sub-phase progress.
// 2. After profiler terminates: writes the final profiling output (final_config.yaml
// + profiler_status.yaml) to the same ConfigMap, preserving the phase+message keys.
const sidecarScriptTemplate = `
set -e
set -o pipefail
STATUS_FILE="{{.OutputPath}}/profiler_status.yaml"
LAST_PHASE=""
START_TIME=$(date +%s)
LAST_PROGRESS_LOG=$START_TIME
PROGRESS_INTERVAL=300
# relay_phase: read phase+message from profiler_status.yaml and write to ConfigMap.
# Only writes when the phase changes (debounce).
relay_phase() {
if [ ! -f "$STATUS_FILE" ]; then
return
fi
PHASE=$(grep "^phase:" "$STATUS_FILE" 2>/dev/null | awk '{print $2}' | tr -d '"' | tr -d "'" || true)
MESSAGE=$(grep "^message:" "$STATUS_FILE" 2>/dev/null | sed 's/^message: *//' | tr -d '"' | tr -d "'" || true)
if [ -z "$PHASE" ] || [ "$PHASE" = "$LAST_PHASE" ]; then
return
fi
echo "Phase update: $PHASE - $MESSAGE"
cat >/tmp/progress.yaml <<PEOF
apiVersion: v1
kind: ConfigMap
metadata:
name: {{.ConfigMapName}}
namespace: {{.Namespace}}
labels:
dgdr.nvidia.com/name: {{.DGDRName}}
dgdr.nvidia.com/namespace: {{.Namespace}}
nvidia.com/managed-by: dynamo-operator
data:
phase: "$PHASE"
message: "$MESSAGE"
PEOF
kubectl apply -f /tmp/progress.yaml 2>/dev/null && LAST_PHASE="$PHASE" || echo "Warning: failed to update progress ConfigMap"
}
# Main loop: poll profiler_status.yaml and wait for profiler to terminate
echo "Waiting for profiler to complete..."
while true; do
CURRENT_TIME=$(date +%s)
ELAPSED=$((CURRENT_TIME - START_TIME))
# Relay phase updates to ConfigMap
relay_phase
# Log progress every 5 minutes
if [ $((CURRENT_TIME - LAST_PROGRESS_LOG)) -ge $PROGRESS_INTERVAL ]; then
echo "Still waiting... ($(($ELAPSED / 60)) minutes elapsed)"
LAST_PROGRESS_LOG=$CURRENT_TIME
fi
# Check if profiler container terminated
CONTAINER_STATUS=$(kubectl get pod $HOSTNAME -n {{.Namespace}} -o jsonpath='{.status.containerStatuses[?(@.name=="profiler")].state}' 2>/dev/null || echo "")
if echo "$CONTAINER_STATUS" | grep -q "terminated"; then
echo "Profiler terminated (ran for $(($ELAPSED / 60)) minutes)"
break
fi
sleep 10
done
# Final relay: pick up any last phase change written just before termination
relay_phase
# Check profiler status file (2 minute timeout)
echo "Checking profiler status..."
TIMEOUT=120
CHECK_START=$(date +%s)
# Wait for status file to exist
while [ ! -f "$STATUS_FILE" ]; do
ELAPSED=$(($(date +%s) - CHECK_START))
if [ $ELAPSED -ge $TIMEOUT ]; then
echo "ERROR: Status file not found after ${TIMEOUT}s"
exit 1
fi
sleep 2
done
# Read and parse status from YAML file
STATUS=$(grep "^status:" "$STATUS_FILE" | awk '{print $2}' | tr -d '"' | tr -d "'")
if [ -z "$STATUS" ]; then
echo "ERROR: Invalid status file format"
exit 1
fi
# Check status value
case "$STATUS" in
success)
MESSAGE=$(grep "^message:" "$STATUS_FILE" | sed 's/^message: *//' | tr -d '"' | tr -d "'")
echo "Profiler succeeded: $MESSAGE"
;;
failed)
ERROR=$(grep "^error:" "$STATUS_FILE" | sed 's/^error: *//' | tr -d '"' | tr -d "'")
MESSAGE=$(grep "^message:" "$STATUS_FILE" | sed 's/^message: *//' | tr -d '"' | tr -d "'")
echo "ERROR: Profiler failed: ${ERROR:-$MESSAGE}"
exit 1
;;
running)
echo "ERROR: Profiler still running (unexpected)"
exit 1
;;
*)
echo "ERROR: Unknown status: $STATUS"
exit 1
;;
esac
echo "Writing profiling output to ConfigMap..."
# Read final phase+message to preserve them alongside the profiling output
FINAL_PHASE=$(grep "^phase:" "$STATUS_FILE" 2>/dev/null | awk '{print $2}' | tr -d '"' | tr -d "'" || true)
FINAL_MESSAGE=$(grep "^message:" "$STATUS_FILE" 2>/dev/null | sed 's/^message: *//' | tr -d '"' | tr -d "'" || true)
# Start building ConfigMap YAML with DGD spec + preserved phase/message
cat >/tmp/cm.yaml <<EOF
apiVersion: v1
kind: ConfigMap
metadata:
name: {{.ConfigMapName}}
namespace: {{.Namespace}}
labels:
dgdr.nvidia.com/name: {{.DGDRName}}
dgdr.nvidia.com/namespace: {{.Namespace}}
nvidia.com/managed-by: dynamo-operator
data:
phase: "$FINAL_PHASE"
message: "$FINAL_MESSAGE"
{{.OutputFile}}: |
EOF
sed 's/^/ /' {{.OutputPath}}/{{.OutputFile}} >> /tmp/cm.yaml
# Add profiler status file for debugging
if [ -f {{.OutputPath}}/profiler_status.yaml ]; then
echo " profiler_status.yaml: |" >> /tmp/cm.yaml
sed 's/^/ /' {{.OutputPath}}/profiler_status.yaml >> /tmp/cm.yaml
fi
# Add webui_data.json for pareto curve data (used by operator to populate status.profilingResults.pareto)
if [ -f {{.OutputPath}}/webui_data.json ]; then
echo " webui_data.json: |" >> /tmp/cm.yaml
sed 's/^/ /' {{.OutputPath}}/webui_data.json >> /tmp/cm.yaml
fi
# Note: Profiling data (raw_data.npz converted to JSON) is included in the
# generated DGD YAML as a separate ConfigMap by the profiler, no need to add it here
kubectl apply -f /tmp/cm.yaml
echo "Saved profiling output to ConfigMap {{.ConfigMapName}}"
`
// profilingPhaseReason returns the condition Reason for a profiling sub-phase.
// By design, the ProfilingPhase string values are identical to the Reason values
// (e.g., ProfilingPhaseSweepingDecode = "SweepingDecode" = ProfilingReasonSweepingDecode).
func profilingPhaseReason(phase nvidiacomv1beta1.ProfilingPhase) string {
if phase == nvidiacomv1beta1.ProfilingPhaseDone {
return nvidiacomv1beta1.ProfilingReasonCompleted
}
return string(phase)
}
// profilingPhaseFailureReason returns the condition Reason for a failed profiling sub-phase.
// By convention, failure reasons are "<Phase>Failed" (e.g., "SweepingDecodeFailed").
// An empty phase yields the generic "ProfilingFailed".
func profilingPhaseFailureReason(phase nvidiacomv1beta1.ProfilingPhase) string {
if phase == "" {
return "ProfilingFailed"
}
return string(phase) + "Failed"
}
// validProfilingPhases is the set of phases the profiler sidecar may report.
var validProfilingPhases = map[nvidiacomv1beta1.ProfilingPhase]struct{}{
nvidiacomv1beta1.ProfilingPhaseInitializing: {},
nvidiacomv1beta1.ProfilingPhaseSweepingPrefill: {},
nvidiacomv1beta1.ProfilingPhaseSweepingDecode: {},
nvidiacomv1beta1.ProfilingPhaseSelectingConfig: {},
nvidiacomv1beta1.ProfilingPhaseBuildingCurves: {},
nvidiacomv1beta1.ProfilingPhaseGeneratingDGD: {},
nvidiacomv1beta1.ProfilingPhaseDone: {},
}
// isValidProfilingPhase returns true if phase is a recognized ProfilingPhase value.
func isValidProfilingPhase(phase string) bool {
_, ok := validProfilingPhases[nvidiacomv1beta1.ProfilingPhase(phase)]
return ok
}
// DynamoGraphDeploymentRequestReconciler reconciles a DynamoGraphDeploymentRequest object
type DynamoGraphDeploymentRequestReconciler struct {
client.Client
APIReader client.Reader
Recorder record.EventRecorder
Config *configv1alpha1.OperatorConfiguration
RuntimeConfig *commonController.RuntimeConfig
GPUDiscoveryCache *gpu.GPUDiscoveryCache
GPUDiscovery *gpu.GPUDiscovery
// RBACMgr handles RBAC setup for profiling jobs
RBACManager RBACManager
}
// RBACManager interface for managing RBAC resources
type RBACManager interface {
EnsureServiceAccountWithRBAC(ctx context.Context, targetNamespace, serviceAccountName, clusterRoleName string) error
}
// GetRecorder implements commonController.Reconciler interface
func (r *DynamoGraphDeploymentRequestReconciler) GetRecorder() record.EventRecorder {
return r.Recorder
}
// FinalizeResource implements commonController.Finalizer interface
func (r *DynamoGraphDeploymentRequestReconciler) FinalizeResource(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) error {
logger := log.FromContext(ctx)
logger.Info("DGDR finalized successfully", "name", dgdr.Name)
return nil
}
// +kubebuilder:rbac:groups=nvidia.com,resources=dynamographdeploymentrequests,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=nvidia.com,resources=dynamographdeploymentrequests/status,verbs=get;update;patch
// +kubebuilder:rbac:groups=nvidia.com,resources=dynamographdeploymentrequests/finalizers,verbs=update
// +kubebuilder:rbac:groups=nvidia.com,resources=dynamographdeployments,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=nvidia.com,resources=dynamographdeployments/status,verbs=get;update;patch
// +kubebuilder:rbac:groups=nvidia.com,resources=dynamographdeployments/finalizers,verbs=update
// +kubebuilder:rbac:groups=batch,resources=jobs,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=core,resources=pods,verbs=get;list;watch
// +kubebuilder:rbac:groups=core,resources=configmaps,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=core,resources=events,verbs=create;patch
// Reconcile handles the reconciliation loop for DynamoGraphDeploymentRequest
func (r *DynamoGraphDeploymentRequestReconciler) Reconcile(ctx context.Context, req ctrl.Request) (ctrl.Result, error) {
logger := log.FromContext(ctx)
logger.Info("Reconciling DynamoGraphDeploymentRequest", "name", req.Name, "namespace", req.Namespace)
// Fetch the DGDR instance
dgdr := &nvidiacomv1beta1.DynamoGraphDeploymentRequest{}
if err := r.Get(ctx, req.NamespacedName, dgdr); err != nil {
if apierrors.IsNotFound(err) {
logger.Info("DGDR resource not found, ignoring since object must be deleted")
return ctrl.Result{}, nil
}
logger.Error(err, "Failed to get DGDR")
return ctrl.Result{}, err
}
// Handle finalizer using common function
finalized, err := commonController.HandleFinalizer(ctx, dgdr, r.Client, r)
if err != nil {
return ctrl.Result{}, err
}
if finalized {
// Resource was deleted and finalized
return ctrl.Result{}, nil
}
// Check for spec changes (immutability enforcement)
if dgdr.Status.ObservedGeneration > 0 && dgdr.Status.ObservedGeneration != dgdr.Generation {
// Spec changed after initial processing
if dgdr.Status.Phase == nvidiacomv1beta1.DGDRPhaseProfiling || dgdr.Status.Phase == nvidiacomv1beta1.DGDRPhaseDeploying ||
dgdr.Status.Phase == nvidiacomv1beta1.DGDRPhaseReady || dgdr.Status.Phase == nvidiacomv1beta1.DGDRPhaseDeployed {
logger.Info("Spec change detected in immutable phase",
"phase", dgdr.Status.Phase,
"observedGeneration", dgdr.Status.ObservedGeneration,
"currentGeneration", dgdr.Generation)
r.Recorder.Event(dgdr, corev1.EventTypeWarning, nvidiacomv1beta1.EventReasonSpecChangeRejected,
fmt.Sprintf(MessageSpecChangeRejected, dgdr.Status.Phase))
// Keep the old observedGeneration to continue rejecting changes
// No phase transition - stay in current phase with old spec
return ctrl.Result{}, nil
}
}
// Phase machine: handle different phases
switch dgdr.Status.Phase {
case nvidiacomv1beta1.DGDRPhasePending, "":
return r.handlePendingPhase(ctx, dgdr)
case nvidiacomv1beta1.DGDRPhaseProfiling:
return r.handleProfilingPhase(ctx, dgdr)
case nvidiacomv1beta1.DGDRPhaseDeploying:
return r.handleDeployingPhase(ctx, dgdr)
case nvidiacomv1beta1.DGDRPhaseReady:
return r.handleReadyPhase(ctx, dgdr)
case nvidiacomv1beta1.DGDRPhaseDeployed:
return r.handleDeployedPhase(ctx, dgdr)
case nvidiacomv1beta1.DGDRPhaseFailed:
return r.handleFailedPhase(ctx, dgdr)
default:
logger.Info("Unknown phase", "phase", dgdr.Status.Phase)
return r.updatePhaseAndRequeue(ctx, dgdr, nvidiacomv1beta1.DGDRPhaseFailed, MessageInvalidState)
}
}
// handlePendingPhase processes newly created or pending DGDR resources.
// When ObservedGeneration == 0, performs initial validation (merged from v1alpha1 Initializing state).
// Otherwise, starts the profiling process.
func (r *DynamoGraphDeploymentRequestReconciler) handlePendingPhase(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (ctrl.Result, error) {
logger := log.FromContext(ctx)
// First-time processing: validate spec (merged from handleInitialState)
if dgdr.Status.ObservedGeneration == 0 {
logger.Info("Handling initial validation", "name", dgdr.Name)
// Validate the spec
if err := r.validateSpec(ctx, dgdr); err != nil {
r.Recorder.Event(dgdr, corev1.EventTypeWarning, nvidiacomv1beta1.EventReasonValidationFailed, err.Error())
return r.updatePhaseWithCondition(ctx, dgdr, nvidiacomv1beta1.DGDRPhaseFailed, nvidiacomv1beta1.ConditionTypeValidation, metav1.ConditionFalse, nvidiacomv1beta1.EventReasonValidationFailed, err.Error())
}
// Set observedGeneration to track the spec we're processing
dgdr.Status.ObservedGeneration = dgdr.Generation
// Initialize status — next reconcile will discover hardware and create the profiling job.
r.Recorder.Event(dgdr, corev1.EventTypeNormal, nvidiacomv1beta1.EventReasonInitialized, MessageInitialized)
return r.updatePhaseWithCondition(ctx, dgdr, nvidiacomv1beta1.DGDRPhasePending,
nvidiacomv1beta1.ConditionTypeProfiling, metav1.ConditionFalse,
"DiscoveringHardware", MessageDiscoveringHardware)
}
logger.Info("Handling pending phase", "name", dgdr.Name)
// Create profiling job (online or AIC)
if err := r.createProfilingJob(ctx, dgdr); err != nil {
r.Recorder.Event(dgdr, corev1.EventTypeWarning, nvidiacomv1beta1.EventReasonProfilingJobFailed, err.Error())
return r.updatePhaseWithCondition(ctx, dgdr, nvidiacomv1beta1.DGDRPhaseFailed, nvidiacomv1beta1.ConditionTypeProfiling, metav1.ConditionFalse, MessageJobCreationFailed, err.Error())
}
// Record event with appropriate message
if isOnlineProfiling(dgdr) {
r.Recorder.Event(dgdr, corev1.EventTypeNormal, nvidiacomv1beta1.EventReasonProfilingJobCreated, MessageProfilingJobCreated)
} else {
r.Recorder.Event(dgdr, corev1.EventTypeNormal, nvidiacomv1beta1.EventReasonProfilingJobCreated, MessageAICProfilingJobCreated)
}
// Update to Profiling phase — use Initializing reason to indicate the profiler is loading.
dgdr.SetProfilingPhase(nvidiacomv1beta1.ProfilingPhaseInitializing)
return r.updatePhaseWithCondition(ctx, dgdr, nvidiacomv1beta1.DGDRPhaseProfiling, nvidiacomv1beta1.ConditionTypeProfiling, metav1.ConditionFalse, nvidiacomv1beta1.ProfilingReasonInitializing, MessageDiscoveringHardware)
}
// updateProfilingSubPhase reads the output ConfigMap and updates status.profilingPhase
// and the Profiling/Succeeded conditions. The sidecar continuously polls profiler_status.yaml
// and writes phase+message to the output ConfigMap (dgdr-output-<name>). This function
// reads those keys and copies them verbatim into the DGDR status.
func (r *DynamoGraphDeploymentRequestReconciler) updateProfilingSubPhase(
ctx context.Context,
dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest,
) error {
logger := log.FromContext(ctx)
outputCMName := getOutputConfigMapName(dgdr)
cm := &corev1.ConfigMap{}
if err := r.Get(ctx, types.NamespacedName{
Name: outputCMName, Namespace: dgdr.Namespace,
}, cm); err != nil {
return nil // No output ConfigMap yet — skip
}
phase, exists := cm.Data["phase"]
if !exists || phase == "" {
return nil
}
if !isValidProfilingPhase(phase) {
return fmt.Errorf("invalid profiling phase %q in ConfigMap %s", phase, outputCMName)
}
profilingPhase := nvidiacomv1beta1.ProfilingPhase(phase)
if dgdr.Status.ProfilingPhase == profilingPhase {
return nil // No change
}
logger.Info("Profiling sub-phase updated", "phase", phase)
dgdr.SetProfilingPhase(profilingPhase)
// Reason is derived from phase; message comes from the profiler via ConfigMap.
reason := profilingPhaseReason(profilingPhase)
message := cm.Data["message"] // written by profiler, relayed by sidecar
meta.SetStatusCondition(&dgdr.Status.Conditions, metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeProfiling,
Status: metav1.ConditionFalse,
ObservedGeneration: dgdr.Generation,
Reason: reason,
Message: message,
})
meta.SetStatusCondition(&dgdr.Status.Conditions, metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeSucceeded,
Status: metav1.ConditionFalse,
ObservedGeneration: dgdr.Generation,
Reason: reason,
Message: message,
})
return r.Status().Update(ctx, dgdr)
}
// handleProfilingPhase monitors profiling progress and generates spec when complete
func (r *DynamoGraphDeploymentRequestReconciler) handleProfilingPhase(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (ctrl.Result, error) {
logger := log.FromContext(ctx)
logger.Info("Handling profiling phase", "name", dgdr.Name)
// Check for sub-phase updates from output ConfigMap (populated by sidecar poller)
if err := r.updateProfilingSubPhase(ctx, dgdr); err != nil {
return ctrl.Result{}, err
}
// Check profiling job status (both online and offline/AIC run as Jobs)
// Note: We watch the Job via Owns(), so we'll be triggered automatically on Job changes
completed, err := r.checkProfilingJobStatus(ctx, dgdr)
if err != nil {
r.Recorder.Event(dgdr, corev1.EventTypeWarning, MessageProfilingCheckFailed, err.Error())
// Job failed - keep profilingPhase set so users can see where it died.
// profilingPhase is already current: set to Initializing on entry,
// then updated by updateProfilingSubPhase() above (reads output ConfigMap).
failureReason := "ProfilingFailed"
failureMessage := err.Error()
if dgdr.Status.ProfilingPhase != "" {
failureReason = profilingPhaseFailureReason(dgdr.Status.ProfilingPhase)
}
// Set phase and conditions directly so we can use sub-phase-specific failure
// reason on both Profiling and Succeeded conditions. (updatePhaseWithCondition
// would hardcode Succeeded reason to generic "Failed".)
dgdr.Status.Phase = nvidiacomv1beta1.DGDRPhaseFailed
meta.SetStatusCondition(&dgdr.Status.Conditions, metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeSucceeded,
Status: metav1.ConditionFalse,
ObservedGeneration: dgdr.Generation,
Reason: failureReason,
Message: failureMessage,
})
dgdr.AddStatusCondition(metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeProfiling,
Status: metav1.ConditionFalse,
ObservedGeneration: dgdr.Generation,
Reason: failureReason,
Message: failureMessage,
})
if err := r.Status().Update(ctx, dgdr); err != nil {
return ctrl.Result{}, err
}
return ctrl.Result{Requeue: true}, nil
}
if !completed {
logger.Info("Profiling job still running", "name", dgdr.Name)
// Transition from Initializing to ProfilingRunning once the job is confirmed active.
cond := meta.FindStatusCondition(dgdr.Status.Conditions, nvidiacomv1beta1.ConditionTypeProfiling)
if cond != nil && cond.Reason == nvidiacomv1beta1.ProfilingReasonInitializing {
return r.updatePhaseWithCondition(ctx, dgdr, nvidiacomv1beta1.DGDRPhaseProfiling, nvidiacomv1beta1.ConditionTypeProfiling, metav1.ConditionFalse, "ProfilingRunning", MessageProfilingInProgress)
}
// Don't requeue - we'll be triggered when the Job completes/fails
return ctrl.Result{}, nil
}
profilingResults, dgdName, err := r.generateDGDSpec(ctx, dgdr)
if err != nil {
dgdr.ClearProfilingPhase()
r.Recorder.Event(dgdr, corev1.EventTypeWarning, MessageGenerationFailed, err.Error())
return r.updatePhaseWithCondition(ctx, dgdr, nvidiacomv1beta1.DGDRPhaseFailed, nvidiacomv1beta1.ConditionTypeSpecGenerated, metav1.ConditionFalse, MessageGenerationFailed, err.Error())
}
if err := r.Get(ctx, types.NamespacedName{Name: dgdr.Name, Namespace: dgdr.Namespace}, dgdr); err != nil {
return ctrl.Result{}, fmt.Errorf("failed to refetch DGDR after generateDGDSpec: %w", err)
}
dgdr.ClearProfilingPhase()
meta.SetStatusCondition(&dgdr.Status.Conditions, metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeProfiling,
Status: metav1.ConditionTrue,
ObservedGeneration: dgdr.Generation,
Reason: "ProfilingCompleted",
Message: "Profiling job completed successfully",
})
dgdr.Status.DGDName = dgdName
dgdr.Status.ProfilingResults = profilingResults
r.Recorder.Event(dgdr, corev1.EventTypeNormal, nvidiacomv1beta1.EventReasonSpecGenerated, MessageSpecGenerated)
// Create additional resources (ConfigMaps) immediately after profiling
// This ensures that the `planner-profile-data` ConfigMap is available for both auto and manual deployment
// v1beta1 uses the DGDR namespace for additional resources.
targetNamespace := dgdr.Namespace
if err := r.createAdditionalResources(ctx, dgdr, targetNamespace); err != nil {
logger.Error(err, "Failed to create additional resources after profiling")
// Don't fail the DGDR, just log the error - ConfigMaps can be created manually
r.Recorder.Event(dgdr, corev1.EventTypeWarning, "ConfigMapCreationFailed",
fmt.Sprintf("Failed to create ConfigMaps from profiling output: %v", err))
}
// If autoApply is enabled, transition to Deploying phase
if dgdr.Spec.AutoApply == nil || *dgdr.Spec.AutoApply {
logger.Info("AutoApply enabled, transitioning to Deploying phase")
return r.updatePhaseWithCondition(ctx, dgdr, nvidiacomv1beta1.DGDRPhaseDeploying, nvidiacomv1beta1.ConditionTypeSpecGenerated, metav1.ConditionTrue, nvidiacomv1beta1.EventReasonSpecGenerated, MessageSpecGenerated)
}
// Otherwise, transition to Ready phase
return r.updatePhaseWithCondition(ctx, dgdr, nvidiacomv1beta1.DGDRPhaseReady, nvidiacomv1beta1.ConditionTypeSpecGenerated, metav1.ConditionTrue, nvidiacomv1beta1.EventReasonSpecGenerated, MessageSpecAvailable)
}
// handleReadyPhase handles DGDR in Ready phase (profiling complete, spec available)
func (r *DynamoGraphDeploymentRequestReconciler) handleReadyPhase(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (ctrl.Result, error) {
logger := log.FromContext(ctx)
logger.Info("DGDR is ready", "name", dgdr.Name)
// Nothing to monitor in Ready phase - spec is available for manual application
return ctrl.Result{}, nil
}
// handleDeployingPhase handles DGD creation and monitors deployment
func (r *DynamoGraphDeploymentRequestReconciler) handleDeployingPhase(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (ctrl.Result, error) {
logger := log.FromContext(ctx)
logger.Info("Handling deploying phase", "name", dgdr.Name)
if dgdr.Spec.AutoApply != nil && !*dgdr.Spec.AutoApply {
// Shouldn't be in this phase without autoApply
logger.Info("AutoApply not enabled, transitioning to Ready")
dgdr.Status.Phase = nvidiacomv1beta1.DGDRPhaseReady
setSucceededCondition(dgdr, nvidiacomv1beta1.DGDRPhaseReady)
return ctrl.Result{}, r.Status().Update(ctx, dgdr)
}
if dgdr.Status.DGDName == "" {
return r.createDGD(ctx, dgdr)
}
dgd := &dgdv1alpha1.DynamoGraphDeployment{}
err := r.Get(ctx, types.NamespacedName{
Name: dgdr.Status.DGDName,
Namespace: dgdr.Namespace,
}, dgd)
if apierrors.IsNotFound(err) {
// Annotation present means DGD was never created (spec ready but create not yet called).
// Annotation absent means DGD was previously created and then manually deleted.
if _, hasSpec := dgdr.Annotations["nvidia.com/generated-dgd-spec"]; hasSpec {
return r.createDGD(ctx, dgdr)
}
return r.handleDGDDeleted(ctx, dgdr)
}
if err != nil {
return ctrl.Result{}, err
}
// Check if DGD is Ready
var condStatus metav1.ConditionStatus
var condReason, condMessage string
if dgd.Status.State == dgdv1alpha1.DGDStateSuccessful {
logger.Info("DGD is Ready, transitioning to Deployed phase")
dgdr.Status.Phase = nvidiacomv1beta1.DGDRPhaseDeployed
setSucceededCondition(dgdr, nvidiacomv1beta1.DGDRPhaseDeployed)
r.Recorder.Event(dgdr, corev1.EventTypeNormal, nvidiacomv1beta1.EventReasonDeploymentReady,
fmt.Sprintf(MessageDeploymentReady, dgd.Name))
condStatus = metav1.ConditionTrue
condReason = nvidiacomv1beta1.EventReasonDeploymentReady
condMessage = fmt.Sprintf(MessageDeploymentReady, dgd.Name)
} else {
logger.Info("DGD not yet ready", "name", dgd.Name, "state", dgd.Status.State)
condStatus = metav1.ConditionFalse
condReason = "DeploymentInProgress"
condMessage = fmt.Sprintf("DGD %s is in %s state", dgd.Name, string(dgd.Status.State))
}
updateDeploymentInfo(dgdr, dgd)
meta.SetStatusCondition(&dgdr.Status.Conditions, metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeDeploymentReady,
Status: condStatus,
Reason: condReason,
Message: condMessage,
})
return ctrl.Result{}, r.Status().Update(ctx, dgdr)
}
// handleDeployedPhase monitors a healthy DGD and detects degradation or deletion
func (r *DynamoGraphDeploymentRequestReconciler) handleDeployedPhase(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (ctrl.Result, error) {
logger := log.FromContext(ctx)
logger.Info("DGDR is deployed", "name", dgdr.Name)
// Check if DGD still exists and monitor its status
dgd := &dgdv1alpha1.DynamoGraphDeployment{}
err := r.Get(ctx, types.NamespacedName{
Name: dgdr.Status.DGDName,
Namespace: dgdr.Namespace,
}, dgd)
if apierrors.IsNotFound(err) {
// DGD was deleted by user
return r.handleDGDDeleted(ctx, dgdr)
}
if err != nil {
return ctrl.Result{}, err
}
// Check if DGD degraded from Ready
if dgd.Status.State != dgdv1alpha1.DGDStateSuccessful {
logger.Info("DGD degraded, transitioning back to Deploying",
"dgdState", dgd.Status.State)
dgdr.Status.Phase = nvidiacomv1beta1.DGDRPhaseDeploying
setSucceededCondition(dgdr, nvidiacomv1beta1.DGDRPhaseDeploying)
updateDeploymentInfo(dgdr, dgd)
r.Recorder.Event(dgdr, corev1.EventTypeWarning, nvidiacomv1beta1.EventReasonDeploymentDegraded,
fmt.Sprintf(MessageDeploymentDegraded, dgd.Name, string(dgd.Status.State)))
meta.SetStatusCondition(&dgdr.Status.Conditions, metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeDeploymentReady,
Status: metav1.ConditionFalse,
Reason: nvidiacomv1beta1.EventReasonDeploymentDegraded,
Message: fmt.Sprintf("Deployment degraded to %s", string(dgd.Status.State)),
})
} else {
// DGD is healthy — update replica info only if changed
if !updateDeploymentInfo(dgdr, dgd) {
// Nothing changed, skip the status write
return ctrl.Result{}, nil
}
}
return ctrl.Result{}, r.Status().Update(ctx, dgdr)
}
// handleDGDDeleted handles the case when auto-created DGD is deleted by user.
// In v1beta1, this transitions to Failed (DeploymentDeleted phase was removed).
func (r *DynamoGraphDeploymentRequestReconciler) handleDGDDeleted(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (ctrl.Result, error) {
logger := log.FromContext(ctx)
logger.Info("DGD was deleted by user, transitioning to Failed phase")
dgdr.Status.Phase = nvidiacomv1beta1.DGDRPhaseFailed
setSucceededCondition(dgdr, nvidiacomv1beta1.DGDRPhaseFailed)
r.Recorder.Event(dgdr, corev1.EventTypeWarning, nvidiacomv1beta1.EventReasonDeploymentDeleted,
fmt.Sprintf(MessageDeploymentDeleted, dgdr.Status.DGDName))
dgdr.Status.DGDName = ""
dgdr.Status.DeploymentInfo = nil
meta.SetStatusCondition(&dgdr.Status.Conditions, metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeDeploymentReady,
Status: metav1.ConditionFalse,
Reason: nvidiacomv1beta1.EventReasonDeploymentDeleted,
Message: "Deployment was deleted by user. Create a new DGDR to redeploy.",
})
return ctrl.Result{}, r.Status().Update(ctx, dgdr)
}
// createDGD creates a DynamoGraphDeployment with the generated spec
func (r *DynamoGraphDeploymentRequestReconciler) createDGD(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (ctrl.Result, error) {
logger := log.FromContext(ctx)
// Extract DGD spec from annotation (stored by generateDGDSpec)
dgdSpecYAML, ok := dgdr.Annotations["nvidia.com/generated-dgd-spec"]
if !ok || dgdSpecYAML == "" {
return ctrl.Result{}, fmt.Errorf("generated DGD spec not found in annotation nvidia.com/generated-dgd-spec")
}
generatedDGD := &dgdv1alpha1.DynamoGraphDeployment{}
if err := yaml.Unmarshal([]byte(dgdSpecYAML), generatedDGD); err != nil {
return ctrl.Result{}, fmt.Errorf("failed to unmarshal generated deployment from annotation: %w", err)
}
// Determine DGD name and namespace from generated deployment
dgdName := generatedDGD.Name
dgdNamespace := dgdr.Namespace
// Build labels (start with generated DGD's labels)
labels := make(map[string]string)
if generatedDGD.Labels != nil {
for k, v := range generatedDGD.Labels {
labels[k] = v
}
}
// Add/override with managed labels
labels[nvidiacomv1beta1.LabelDGDRName] = dgdr.Name
labels[nvidiacomv1beta1.LabelDGDRNamespace] = dgdr.Namespace
labels[nvidiacomv1beta1.LabelManagedBy] = nvidiacomv1beta1.LabelValueDynamoOperator
// Build annotations (start with generated DGD's annotations)
annotations := make(map[string]string)
if generatedDGD.Annotations != nil {
for k, v := range generatedDGD.Annotations {
annotations[k] = v
}
}
// Create DGD from generated deployment
dgd := &dgdv1alpha1.DynamoGraphDeployment{
ObjectMeta: metav1.ObjectMeta{
Name: dgdName,
Namespace: dgdNamespace,
Labels: labels,
Annotations: annotations,
},
Spec: generatedDGD.Spec,
}
// Note: We don't set owner reference on DGD
// If a DGDR is deleted, the DGD may be serving traffic and should persist independently.
// We use labels (LabelDGDRName) to track the relationship.
logger.Info("Creating DynamoGraphDeployment", "name", dgdName, "namespace", dgdNamespace)
if err := r.Create(ctx, dgd); err != nil {
if apierrors.IsAlreadyExists(err) {
logger.Info("DGD already exists, updating status")
delete(dgdr.Annotations, "nvidia.com/generated-dgd-spec")
if updateErr := r.Update(ctx, dgdr); updateErr != nil {
logger.Error(updateErr, "Failed to remove generated-dgd-spec annotation on IsAlreadyExists path")
return ctrl.Result{}, updateErr
}
dgdr.Status.DGDName = dgdName
return ctrl.Result{}, r.Status().Update(ctx, dgdr)
}
r.Recorder.Event(dgdr, corev1.EventTypeWarning, MessageDeploymentCreationFailed, err.Error())
return ctrl.Result{}, err
}
delete(dgdr.Annotations, "nvidia.com/generated-dgd-spec")
if err := r.Update(ctx, dgdr); err != nil {
// Return the error to force a retry. The DGD was created successfully, so a
// retry will hit the IsAlreadyExists path above and attempt cleanup again.
return ctrl.Result{}, fmt.Errorf("failed to remove generated-dgd-spec annotation after DGD creation: %w", err)
}
// Update status
dgdr.Status.DGDName = dgdName
r.Recorder.Event(dgdr, corev1.EventTypeNormal, nvidiacomv1beta1.EventReasonDeploymentCreated,
fmt.Sprintf(MessageDeploymentCreated, dgdName))
meta.SetStatusCondition(&dgdr.Status.Conditions, metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeDeploymentReady,
Status: metav1.ConditionFalse,
Reason: nvidiacomv1beta1.EventReasonDeploymentCreated,
Message: fmt.Sprintf("DGD %s created, waiting for Ready", dgdName),
})
logger.Info("DynamoGraphDeployment created successfully", "name", dgdName)
return ctrl.Result{}, r.Status().Update(ctx, dgdr)
}
// createAdditionalResources creates ConfigMaps from the profiling output that should be deployed alongside the DGD
func (r *DynamoGraphDeploymentRequestReconciler) createAdditionalResources(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest, targetNamespace string) error {
logger := log.FromContext(ctx)
// Check if there are additional resources stored in annotations
if dgdr.Annotations == nil {
return nil
}
resourcesYAML, exists := dgdr.Annotations[AnnotationAdditionalResources]
if !exists || resourcesYAML == "" {
return nil
}
// Parse using standard Kubernetes YAML decoder
decoder := yaml.NewYAMLOrJSONDecoder(bytes.NewReader([]byte(resourcesYAML)), 4096)
resourceCount := 0
for {
obj := &unstructured.Unstructured{}
if err := decoder.Decode(obj); err != nil {
if err == io.EOF {
break
}
logger.Error(err, "Failed to decode resource, skipping")
continue
}
if obj.GetKind() == "" {
continue
}
resourceCount++
// Only support ConfigMap for now (what profiler actually generates)
if obj.GetKind() != "ConfigMap" {
logger.Info("Skipping non-ConfigMap resource from profiling output", "kind", obj.GetKind(), "name", obj.GetName())
continue
}
cm := &corev1.ConfigMap{}
if err := runtime.DefaultUnstructuredConverter.FromUnstructured(obj.Object, cm); err != nil {
logger.Error(err, "Failed to convert to ConfigMap", "name", obj.GetName())
continue
}
// Override namespace and add tracking labels
cm.Namespace = targetNamespace
if cm.Labels == nil {
cm.Labels = make(map[string]string)
}
cm.Labels[nvidiacomv1beta1.LabelDGDRName] = dgdr.Name
cm.Labels[nvidiacomv1beta1.LabelDGDRNamespace] = dgdr.Namespace
cm.Labels[nvidiacomv1beta1.LabelManagedBy] = nvidiacomv1beta1.LabelValueDynamoOperator
// Create the ConfigMap
if err := r.Create(ctx, cm); err != nil {
if apierrors.IsAlreadyExists(err) {
logger.Info("ConfigMap already exists, skipping", "name", cm.Name)
} else {
return fmt.Errorf("failed to create ConfigMap %s: %w", cm.Name, err)
}
} else {
logger.Info("Created ConfigMap from profiling output", "name", cm.Name, "namespace", targetNamespace)
}
}
if resourceCount > 0 {
logger.Info("Deploying additional resources from profiling output", "count", resourceCount)
}
return nil
}
// handleFailedPhase handles DGDR in Failed phase
func (r *DynamoGraphDeploymentRequestReconciler) handleFailedPhase(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (ctrl.Result, error) {
logger := log.FromContext(ctx)
logger.Info("DGDR is in failed phase", "name", dgdr.Name)
// Could implement retry logic here if desired
return ctrl.Result{}, nil
}
// getProfilingJobName returns the job name for a DGDR
func getProfilingJobName(dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) string {
// Use "profile-" prefix for all profiling jobs
return fmt.Sprintf("profile-%s", dgdr.Name)
}
// getOutputConfigMapName returns the ConfigMap name for profiling output
func getOutputConfigMapName(dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) string {
return fmt.Sprintf("%s%s", ConfigMapOutputPrefix, dgdr.Name)
}
// isOnlineProfiling returns true. In v1beta1, the profiler decides online vs AIC
// mode internally based on its config. The controller always uses the same label.
func isOnlineProfiling(_ *nvidiacomv1beta1.DynamoGraphDeploymentRequest) bool {
return true
}
// validateSpec validates the DGDR spec
func (r *DynamoGraphDeploymentRequestReconciler) validateSpec(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) error {
var errs []error
// Disallow searchStrategy: thorough with backend: auto.
if dgdr.Spec.SearchStrategy == nvidiacomv1beta1.SearchStrategyThorough &&
dgdr.Spec.Backend == nvidiacomv1beta1.BackendTypeAuto {
errs = append(errs, fmt.Errorf(
"spec.searchStrategy %q is incompatible with spec.backend %q: set spec.backend to a specific backend (sglang, trtllm, or vllm)",
nvidiacomv1beta1.SearchStrategyThorough,
nvidiacomv1beta1.BackendTypeAuto,
))
}
// Validate model cache PVC if provided
if dgdr.Spec.ModelCache != nil && dgdr.Spec.ModelCache.PVCName != "" {
pvc := &corev1.PersistentVolumeClaim{}
err := r.Get(ctx, types.NamespacedName{
Name: dgdr.Spec.ModelCache.PVCName,
Namespace: dgdr.Namespace,
}, pvc)
if err != nil {
if apierrors.IsNotFound(err) {
errs = append(errs, fmt.Errorf(MessageModelCachePVCNotFound, dgdr.Spec.ModelCache.PVCName, dgdr.Namespace))
} else {
return err
}
}
}
if err := r.validateGPUHardwareInfo(ctx, dgdr); err != nil {
errs = append(errs, err)
}
// The profiler will validate the rest of the configuration
return errors.Join(errs...)
}
// validateGPUHardwareInfo ensures GPU hardware information is available when required for profiling
func (r *DynamoGraphDeploymentRequestReconciler) validateGPUHardwareInfo(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) error {
logger := log.FromContext(ctx)
// Check if user provided hardware info in the typed spec
hasManualConfig := dgdr.Spec.Hardware != nil && (dgdr.Spec.Hardware.GPUSKU != "" ||
dgdr.Spec.Hardware.VRAMMB != nil ||
dgdr.Spec.Hardware.NumGPUsPerNode != nil)
// If manual config is provided, validation passes
if hasManualConfig {
return nil
}
isNamespaceScoped := r.Config.Namespace.Restricted != ""
if isNamespaceScoped {
return fmt.Errorf(
"GPU hardware info required but cannot be auto-discovered." +
"\n\nOptions to resolve:" +
"\n\n1. Re-enable GPU discovery (if it was disabled during Helm install):" +
"\n helm upgrade ... --set dynamo-operator.gpuDiscovery.enabled=true" +
"\n\n2. Add hardware config to spec.hardware:" +
"\n numGpusPerNode: 8" +
"\n gpuSku: \"H100-SXM5-80GB\"" +
"\n vramMb: 81920")
}
_, err := r.GPUDiscovery.DiscoverGPUsFromDCGM(ctx, r.APIReader, r.GPUDiscoveryCache)
if err == nil {
// GPU discovery is available, validation passes
return nil
}
// Refine the logger message
reason := GetGPUDiscoveryFailureReason(err)
logger.Info("GPU discovery not available", "reason", reason, "error", err.Error())
return fmt.Errorf("GPU hardware info required but auto-discovery failed. Add spec.hardware.gpuSku, spec.hardware.vramMb, spec.hardware.numGpusPerNode")
}
// GetGPUDiscoveryFailureReason classifies a GPU discovery error and
// returns a stable, actionable reason string suitable for structured logging.
//
// The classification is based on known error message patterns produced during:
// - DCGM exporter pod discovery
// - Helm-based GPU operator and DCGM discovery
// - Metrics scraping
// - Prometheus parsing
//
// If the error does not match any known category, "unknown" is returned.
func GetGPUDiscoveryFailureReason(err error) string {
if err == nil {
return "unknown"
}
errMsg := strings.ToLower(err.Error())
switch {
case strings.Contains(errMsg, "list pods"):
return "failed to list DCGM exporter pods (RBAC/cluster connectivity issue)"
case strings.Contains(errMsg, "gpu operator is not installed"):
return "GPU Operator not installed in expected namespace"
case strings.Contains(errMsg, "helm init failed"):
return "failed to initialize Helm client (RBAC, kubeconfig, or Helm driver issue)"
case strings.Contains(errMsg, "timeout waiting for dcgm exporter pods"):
return "timeout while waiting for DCGM exporter pods to become ready"
case strings.Contains(errMsg, "http get"):
return "failed to reach DCGM metrics endpoint on pod (network/port issue)"
case strings.Contains(errMsg, "metrics endpoint") &&
strings.Contains(errMsg, "status"):
return "DCGM pod metrics endpoint returned non-200 status"
case strings.Contains(errMsg, "parse prometheus metrics"):
return "failed to parse dcgm Prometheus metrics (invalid format)"
case strings.Contains(errMsg, "no gpus detected"):
return "no GPUs detected in dcgm metrics (GPU model or metrics missing)"
case strings.Contains(errMsg, "dcgm is not enabled in the GPU Operator"):
return "DCGM is not enabled in the GPU Operator (check GPU Operator configuration and permissions)"
case strings.Contains(errMsg, "failed to scrape any dcgm exporter pod"):
return "failed to scrape any dcgm exporter pod (check DCGM exporter pod status and network connectivity)"
case strings.Contains(errMsg, "no gpu metrics could be parsed from any dcgm pod"):
return "no GPU metrics could be parsed from any DCGM pod (check DCGM exporter pod status and network connectivity)"
case strings.Contains(errMsg, "failed to create helm path"):
return "failed to initialize Helm client (RBAC, kubeconfig, or Helm driver issue)"
}
return "unknown"
}
// createProfilingJob creates a Kubernetes Job for profiling using SyncResource
func (r *DynamoGraphDeploymentRequestReconciler) createProfilingJob(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) error {
logger := log.FromContext(ctx)
// Delete any existing output ConfigMap to ensure fresh profiling results
// This prevents using stale data from previous profiling runs
outputConfigMapName := getOutputConfigMapName(dgdr)
existingCM := &corev1.ConfigMap{}
err := r.Get(ctx, types.NamespacedName{
Name: outputConfigMapName,
Namespace: dgdr.Namespace,
}, existingCM)
if err == nil {
// ConfigMap exists, delete it
logger.Info("Deleting existing output ConfigMap to ensure fresh profiling results", "configMap", outputConfigMapName)
if err := r.Delete(ctx, existingCM); err != nil && !apierrors.IsNotFound(err) {
logger.Error(err, "Failed to delete existing output ConfigMap", "configMap", outputConfigMapName)
return fmt.Errorf("failed to delete existing output ConfigMap: %w", err)
}
logger.Info("Successfully deleted old output ConfigMap", "configMap", outputConfigMapName)
} else if !apierrors.IsNotFound(err) {
// Unexpected error checking for ConfigMap
logger.Error(err, "Failed to check for existing output ConfigMap", "configMap", outputConfigMapName)
return fmt.Errorf("failed to check for existing output ConfigMap: %w", err)
}
// Ensure profiling job RBAC exists (only for cluster-wide installation)
if r.Config.Namespace.Restricted == "" {
if err := r.RBACManager.EnsureServiceAccountWithRBAC(
ctx,
dgdr.Namespace,
ServiceAccountProfilingJob,
r.Config.RBAC.DGDRProfilingClusterRoleName,
); err != nil {
logger.Error(err, "Failed to ensure profiling job RBAC")
return fmt.Errorf("failed to ensure profiling job RBAC: %w", err)
}
}
// Enrich hardware from GPU discovery before marshalling the spec.
// This fills in gpuSku, vramMb, numGpusPerNode if the user didn't set them.
if err := r.enrichHardwareFromDiscovery(ctx, dgdr); err != nil {
logger.Info("GPU discovery not available, proceeding without enrichment", "reason", err.Error())
}
// Use SyncResource to create/update the job
modified, job, err := commonController.SyncResource(ctx, r, dgdr, func(ctx context.Context) (*batchv1.Job, bool, error) {
jobName := getProfilingJobName(dgdr)
outputConfigMapName := getOutputConfigMapName(dgdr)
// Marshal the DGDR spec to JSON — the profiler receives the spec verbatim
specJSON, err := marshalDGDRSpec(dgdr)
if err != nil {
return nil, false, err
}
// Common environment variables
profilerEnv := []corev1.EnvVar{
{
Name: "HUGGING_FACE_HUB_TOKEN",
ValueFrom: &corev1.EnvVarSource{
SecretKeyRef: &corev1.SecretKeySelector{
LocalObjectReference: corev1.LocalObjectReference{
Name: "hf-token-secret",
},
Key: "HF_TOKEN",
},
},
},
{
Name: "NATS_SERVER",
Value: fmt.Sprintf("nats://%s-nats:4222", dgdr.Namespace),
},
{
Name: "ETCD_ENDPOINTS",
Value: fmt.Sprintf("%s-etcd:2379", dgdr.Namespace),
},
// DGDR metadata for setting ownerReferences
{
Name: "DGDR_NAME",
Value: dgdr.Name,
},
{
Name: "DGDR_NAMESPACE",
Value: dgdr.Namespace,
},
{
Name: "DGDR_UID",
Value: string(dgdr.UID),
},
}
// Build volume mounts
volumeMounts := []corev1.VolumeMount{
{
Name: VolumeNameProfilingOutput,
MountPath: ProfilingOutputPath,
},
}
// Add model cache PVC mount if configured
modelCachePVC, modelCacheMountPath := extractModelCachePVCConfig(dgdr)
if modelCachePVC != "" {
logger.Info("Mounting model cache PVC to profiler pod", "pvc", modelCachePVC, "mountPath", modelCacheMountPath)
volumeMounts = append(volumeMounts, corev1.VolumeMount{
Name: VolumeNameModelCache,
MountPath: modelCacheMountPath,
ReadOnly: true,
})
}
// v1alpha1 round-trip: mount ConfigMap if referenced via annotation
cmRef := configMapRefFromAnnotation(dgdr)
if cmRef != nil {
volumeMounts = append(volumeMounts, corev1.VolumeMount{
Name: VolumeNameProfilingConfig,
MountPath: ProfilingConfigMountPath,
ReadOnly: true,
})
}
// Profiler args: pass the DGDR spec as JSON via --config
// --output-dir must match ProfilingOutputPath so the sidecar can find profiler_status.yaml
profilerArgs := []string{"--config", specJSON, "--output-dir", ProfilingOutputPath}
// Use image from spec; the defaulting webhook fills this in for production builds.
// Guard against empty image in case the webhook didn't run (e.g. local dev builds).
imageName := dgdr.Spec.Image
if imageName == "" {
return nil, false, fmt.Errorf("spec.image is required but not set; ensure the defaulting webhook ran or set spec.image explicitly")
}
logger.Info("Using profiler image", "image", imageName)
profilerContainer := corev1.Container{
Name: ContainerNameProfiler,
Image: imageName,
Command: []string{"python", "-m", "dynamo.profiler"},
Args: profilerArgs,
Env: profilerEnv,
VolumeMounts: volumeMounts,
WorkingDir: "/workspace",
}
// Generate sidecar script from template
tmpl, err := template.New("sidecar").Parse(sidecarScriptTemplate)
if err != nil {
return nil, false, fmt.Errorf("failed to parse sidecar script template: %w", err)
}
var scriptBuf bytes.Buffer
err = tmpl.Execute(&scriptBuf, map[string]string{
"OutputPath": ProfilingOutputPath,
"OutputFile": ProfilingOutputFile,
"ConfigMapName": outputConfigMapName,
"Namespace": dgdr.Namespace,
"DGDRName": dgdr.Name,
})
if err != nil {
return nil, false, fmt.Errorf("failed to execute sidecar script template: %w", err)
}
sidecarContainer := corev1.Container{
Name: ContainerNameOutputCopier,
Image: SidecarImage,
Command: []string{"/bin/sh", "-c"},
Args: []string{scriptBuf.String()},
VolumeMounts: []corev1.VolumeMount{{
Name: VolumeNameProfilingOutput,
MountPath: ProfilingOutputPath,
ReadOnly: true,
}},
}
// Use PVC for profiling output if round-tripped v1alpha1 annotation is present,
// otherwise use emptyDir (v1beta1 default).
var profilingOutputVolume corev1.Volume
if outputPVC := outputPVCFromAnnotation(dgdr); outputPVC != "" {
logger.Info("Using PVC for profiling output (from v1alpha1 annotation)", "pvc", outputPVC)
profilingOutputVolume = corev1.Volume{
Name: VolumeNameProfilingOutput,
VolumeSource: corev1.VolumeSource{
PersistentVolumeClaim: &corev1.PersistentVolumeClaimVolumeSource{
ClaimName: outputPVC,
},
},
}
} else {
profilingOutputVolume = corev1.Volume{
Name: VolumeNameProfilingOutput,
VolumeSource: corev1.VolumeSource{
EmptyDir: &corev1.EmptyDirVolumeSource{},
},
}
}
volumes := []corev1.Volume{profilingOutputVolume}
// Add model cache PVC volume if configured
if modelCachePVC != "" {
volumes = append(volumes, corev1.Volume{
Name: VolumeNameModelCache,
VolumeSource: corev1.VolumeSource{
PersistentVolumeClaim: &corev1.PersistentVolumeClaimVolumeSource{
ClaimName: modelCachePVC,
ReadOnly: true,
},
},
})
}
// v1alpha1 round-trip: add ConfigMap volume if referenced via annotation
if cmRef != nil {
cmKey := cmRef.Key
if cmKey == "" {
cmKey = ProfilingConfigDefaultKey
}
volumes = append(volumes, corev1.Volume{
Name: VolumeNameProfilingConfig,
VolumeSource: corev1.VolumeSource{
ConfigMap: &corev1.ConfigMapVolumeSource{
LocalObjectReference: corev1.LocalObjectReference{
Name: cmRef.Name,
},
Items: []corev1.KeyToPath{{
Key: cmKey,
Path: ProfilingConfigDefaultKey,
}},
},
},
})
}
// Limit retries to prevent infinite loop
backoffLimit := int32(3)
podSpec := corev1.PodSpec{
ServiceAccountName: ServiceAccountProfilingJob,
RestartPolicy: corev1.RestartPolicyNever,
SecurityContext: &corev1.PodSecurityContext{
RunAsNonRoot: ptr.To(true),
RunAsUser: ptr.To[int64](1000),
RunAsGroup: ptr.To[int64](1000),
FSGroup: ptr.To[int64](1000),
},
Containers: []corev1.Container{profilerContainer, sidecarContainer},
Volumes: volumes,
ImagePullSecrets: []corev1.LocalObjectReference{
{Name: "nvcr-imagepullsecret"},
},
}
job := &batchv1.Job{
ObjectMeta: metav1.ObjectMeta{
Name: jobName,
Namespace: dgdr.Namespace,
Labels: map[string]string{
nvidiacomv1beta1.LabelApp: nvidiacomv1beta1.LabelValueDynamoProfiler,
nvidiacomv1beta1.LabelDGDR: dgdr.Name,
nvidiacomv1beta1.LabelManagedBy: nvidiacomv1beta1.LabelValueDynamoOperator,
},
},
Spec: batchv1.JobSpec{
BackoffLimit: &backoffLimit,
Template: corev1.PodTemplateSpec{
Spec: podSpec,
},
},
}
var jobOverrides *batchv1.JobSpec
if dgdr.Spec.Overrides != nil {
jobOverrides = dgdr.Spec.Overrides.ProfilingJob
}
applyProfilingJobOverrides(job, jobOverrides)
return job, false, nil
})
if err != nil {
return err
}
if modified {
logger.Info("Profiling job created/updated", "job", job.Name)
}
// Store the job name in status for observability
dgdr.Status.ProfilingJobName = job.Name
return nil
}
// marshalDGDRSpec produces the JSON string passed to the profiler via --config.
// The profiler receives the DGDR spec verbatim — no bespoke key mapping needed.
func marshalDGDRSpec(dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (string, error) {
specJSON, err := json.Marshal(dgdr.Spec)
if err != nil {
return "", fmt.Errorf("failed to marshal DGDR spec to JSON: %w", err)
}
return string(specJSON), nil
}
// enrichHardwareFromDiscovery fills in hardware fields that the user didn't set.
// Called before marshalDGDRSpec(). Mutates dgdr.Spec.Hardware in-place (memory only, not persisted).
func (r *DynamoGraphDeploymentRequestReconciler) enrichHardwareFromDiscovery(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) error {
if dgdr.Spec.Hardware == nil {
dgdr.Spec.Hardware = &nvidiacomv1beta1.HardwareSpec{}
}
hw := dgdr.Spec.Hardware
if hw.GPUSKU != "" && hw.VRAMMB != nil && hw.NumGPUsPerNode != nil {
return nil // all fields already set by user; TotalGPUs is filled below when discovery runs
}
var gpuInfo *gpu.GPUInfo
logger := log.FromContext(ctx)
// Check if user provided hardware info in the typed spec
hasManualConfig := dgdr.Spec.Hardware != nil && (dgdr.Spec.Hardware.GPUSKU != "" ||
dgdr.Spec.Hardware.VRAMMB != nil ||
dgdr.Spec.Hardware.NumGPUsPerNode != nil)
if !hasManualConfig {
logger.Info("Attempting GPU discovery for profiling job")
discoveredInfo, err := r.GPUDiscovery.DiscoverGPUsFromDCGM(ctx, r.APIReader, r.GPUDiscoveryCache)
if err != nil {
// This path is expected for namespace-restricted operators without node read permissions
// Refine the logger message
reason := GetGPUDiscoveryFailureReason(err)
logger.Info("GPU discovery not available, using manual hardware configuration from profiling config",
"reason", reason, "error", err.Error())
return err
} else {
gpuInfo = discoveredInfo
logger.Info("GPU discovery completed successfully",
"gpusPerNode", gpuInfo.GPUsPerNode,
"nodesWithGPUs", gpuInfo.NodesWithGPUs,
"totalGpus", gpuInfo.GPUsPerNode*gpuInfo.NodesWithGPUs,
"model", gpuInfo.Model,
"vramMiB", gpuInfo.VRAMPerGPU,
"system", gpuInfo.System,
"cloudprovider", gpuInfo.CloudProvider)
}
}
if hw.GPUSKU == "" {
if gpuInfo.System != "" {
hw.GPUSKU = gpuInfo.System
} else {
// Unknown GPU type: use raw model name; profiler will attempt naive config generation.
hw.GPUSKU = nvidiacomv1beta1.GPUSKUType(gpuInfo.Model)
}
}
if hw.VRAMMB == nil {
vram := float64(gpuInfo.VRAMPerGPU)
hw.VRAMMB = &vram
}
if hw.NumGPUsPerNode == nil {
n := int32(gpuInfo.GPUsPerNode)
hw.NumGPUsPerNode = &n
}
if hw.TotalGPUs == nil {
// TODO: This is a temporary limit to prevent the profiler from using too many GPUs.
// Will be removed once a fix is in the Profiler/AIC.
const defaultMaxAutoGPUs = int32(32)
total := int32(gpuInfo.GPUsPerNode * gpuInfo.NodesWithGPUs)
if total > defaultMaxAutoGPUs {
logger.Info("Capping auto-discovered TotalGPUs at default limit; set hardware.totalGpus to override",
"discovered", total, "cap", defaultMaxAutoGPUs)
total = defaultMaxAutoGPUs
}
hw.TotalGPUs = &total
}
return nil
}
// extractModelCachePVCConfig reads model cache PVC settings from the typed v1beta1 spec.
// Returns (pvcName, mountPath) — both empty if not configured.
func extractModelCachePVCConfig(dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (string, string) {
if dgdr.Spec.ModelCache == nil || dgdr.Spec.ModelCache.PVCName == "" {
return "", ""
}
mountPath := dgdr.Spec.ModelCache.PVCMountPath
if mountPath == "" {
mountPath = DefaultModelCacheMountPath
}
return dgdr.Spec.ModelCache.PVCName, mountPath
}
// configMapKeySelector mirrors v1alpha1.ConfigMapKeySelector for annotation deserialization.
type configMapKeySelector struct {
Name string `json:"name"`
Key string `json:"key,omitempty"`
}
// configMapRefFromAnnotation reads the ConfigMap reference from the round-trip annotation.
// Returns nil for native v1beta1 resources (no annotation present).
func configMapRefFromAnnotation(dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) *configMapKeySelector {
if dgdr.Annotations == nil {
return nil
}
raw, ok := dgdr.Annotations[AnnotationConfigMapRef]
if !ok || raw == "" {
return nil
}
var ref configMapKeySelector
if err := json.Unmarshal([]byte(raw), &ref); err != nil {
return nil
}
return &ref
}
// outputPVCFromAnnotation reads the output PVC name from the round-trip annotation.
// Returns "" for native v1beta1 resources (always emptyDir).
func outputPVCFromAnnotation(dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) string {
if dgdr.Annotations == nil {
return ""
}
return dgdr.Annotations[AnnotationOutputPVC]
}
// checkProfilingJobStatus checks if the profiling job has completed
func (r *DynamoGraphDeploymentRequestReconciler) checkProfilingJobStatus(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (bool, error) {
logger := log.FromContext(ctx)
jobName := getProfilingJobName(dgdr)
job := &batchv1.Job{}
if err := r.Get(ctx, types.NamespacedName{Name: jobName, Namespace: dgdr.Namespace}, job); err != nil {
return false, err
}
// Check job conditions
for _, condition := range job.Status.Conditions {
if condition.Type == batchv1.JobComplete && condition.Status == corev1.ConditionTrue {
logger.Info("Profiling job completed", "job", jobName)
return true, nil
}
if condition.Type == batchv1.JobFailed && condition.Status == corev1.ConditionTrue {
// Get detailed error from pod logs
detailedError := r.getProfilingJobErrorDetails(ctx, dgdr, job)
if detailedError != "" {
return false, fmt.Errorf("profiling job failed: %s. Details: %s", condition.Message, detailedError)
}
return false, fmt.Errorf("profiling job failed: %s", condition.Message)
}
}
return false, nil
}
// getProfilingJobErrorDetails retrieves detailed error information from failed profiling job pods
func (r *DynamoGraphDeploymentRequestReconciler) getProfilingJobErrorDetails(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest, job *batchv1.Job) string {
logger := log.FromContext(ctx)
// List pods owned by this job
podList := &corev1.PodList{}
labelSelector := client.MatchingLabels{
"job-name": job.Name,
}
if err := r.List(ctx, podList, client.InNamespace(dgdr.Namespace), labelSelector); err != nil {
logger.Error(err, "Failed to list pods for profiling job")
return ""
}
// Look for failed pods and extract error details
for _, pod := range podList.Items {
// Check pod phase and container statuses
if pod.Status.Phase == corev1.PodFailed {
// Get profiler container status (first container)
for _, containerStatus := range pod.Status.ContainerStatuses {
if containerStatus.Name == ContainerNameProfiler && containerStatus.State.Terminated != nil {
terminated := containerStatus.State.Terminated
// Construct detailed error message
errorMsg := fmt.Sprintf("Pod: %s, Container: %s, ExitCode: %d, Reason: %s",
pod.Name, containerStatus.Name, terminated.ExitCode, terminated.Reason)
if terminated.Message != "" {
errorMsg += fmt.Sprintf(", Message: %s", terminated.Message)
}
logger.Info("Retrieved profiling job error details", "error", errorMsg)
return errorMsg
}
}
// If no terminated state found, check waiting state
for _, containerStatus := range pod.Status.ContainerStatuses {
if containerStatus.Name == ContainerNameProfiler && containerStatus.State.Waiting != nil {
waiting := containerStatus.State.Waiting
errorMsg := fmt.Sprintf("Pod: %s, Container: %s, Waiting - Reason: %s, Message: %s",
pod.Name, containerStatus.Name, waiting.Reason, waiting.Message)
logger.Info("Retrieved profiling job waiting details", "error", errorMsg)
return errorMsg
}
}
}
}
return ""
}
// generateDGDSpec reads profiling output from the sidecar ConfigMap, extracts the
// DynamoGraphDeployment spec and pareto configs, stores the spec in an annotation via
// r.Update, and returns the ProfilingResultsStatus and DGD name.
func (r *DynamoGraphDeploymentRequestReconciler) generateDGDSpec(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest) (*nvidiacomv1beta1.ProfilingResultsStatus, string, error) {
logger := log.FromContext(ctx)
logger.Info("Generating DGD spec from profiling results", "name", dgdr.Name, "backend", dgdr.Spec.Backend)
// Read the generated spec from ConfigMap (created by sidecar)
outputConfigMapName := getOutputConfigMapName(dgdr)
cm := &corev1.ConfigMap{}
err := r.Get(ctx, types.NamespacedName{
Name: outputConfigMapName,
Namespace: dgdr.Namespace,
}, cm)
if err != nil {
if apierrors.IsNotFound(err) {
return nil, "", fmt.Errorf("output ConfigMap %s not found - profiling may not have completed yet", outputConfigMapName)
}
return nil, "", fmt.Errorf("failed to get output ConfigMap: %w", err)
}
// Select the right config file based on mocker feature flag
// Profiler writes the selected config (real or mocker) to a single output file
outputFile := ProfilingOutputFile
// Get YAML content from ConfigMap
yamlContent, exists := cm.Data[outputFile]
if !exists {
return nil, "", fmt.Errorf("key %s not found in ConfigMap %s", outputFile, outputConfigMapName)
}
logger.Info("Found profiling output in ConfigMap", "configMap", outputConfigMapName, "outputFile", outputFile, "size", len(yamlContent))
// Extract DGD and any supporting resources from potentially multi-document YAML (ConfigMap + DGD)
dgd, additionalResources, err := r.extractResourcesFromYAML([]byte(yamlContent))
if err != nil {
return nil, "", fmt.Errorf("failed to extract DGD from %s: %w", outputFile, err)
}
logger.Info("Parsed profiling output", "dgdName", dgd.Name, "additionalResources", len(additionalResources))
if len(additionalResources) > 0 {
if err := r.storeAdditionalResources(ctx, dgdr, additionalResources); err != nil {
logger.Error(err, "Failed to store additional resources")
return nil, "", err
}
// storeAdditionalResources calls r.Update internally, bumping resourceVersion.
// Refetch so the subsequent r.Update for the spec annotation doesn't 409.
if err := r.Get(ctx, types.NamespacedName{Name: dgdr.Name, Namespace: dgdr.Namespace}, dgdr); err != nil {
return nil, "", fmt.Errorf("failed to refetch DGDR after storing additional resources: %w", err)
}
}
profilingResults := &nvidiacomv1beta1.ProfilingResultsStatus{}
if webUIData, ok := cm.Data["webui_data.json"]; ok {
pareto, err := extractParetoFromWebUIData([]byte(webUIData))
if err != nil {
logger.Error(err, "Failed to parse webui_data.json; skipping pareto population")
} else {
profilingResults.Pareto = pareto
logger.Info("Populated ProfilingResults.Pareto", "count", len(pareto))
}
}
// Store the generated DGD in ProfilingResults.SelectedConfig
dgdJSON, err := json.Marshal(dgd)
if err != nil {
return nil, "", fmt.Errorf("failed to marshal generated DGD to JSON: %w", err)
}
profilingResults.SelectedConfig = &runtime.RawExtension{Raw: dgdJSON}
// Serialize the DGD spec to an annotation so createDGD can retrieve it
dgdBytes, err := sigsyaml.Marshal(dgd)
if err != nil {
return nil, "", fmt.Errorf("failed to marshal generated DGD: %w", err)
}
if dgdr.Annotations == nil {
dgdr.Annotations = make(map[string]string)
}
dgdr.Annotations["nvidia.com/generated-dgd-spec"] = string(dgdBytes)
if err := r.Update(ctx, dgdr); err != nil {
return nil, "", fmt.Errorf("failed to update DGDR with generated DGD annotation: %w", err)
}
return profilingResults, dgd.Name, nil
}
// extractParetoFromWebUIData parses webui_data.json and returns all Pareto-optimal
// deployment configurations from the cost table. Each row's last column ("Action")
// is a partial DynamoGraphDeployment YAML snippet.
func extractParetoFromWebUIData(data []byte) ([]nvidiacomv1beta1.ParetoConfig, error) {
var parsed struct {
Cost struct {
Table struct {
Data [][]json.RawMessage `json:"data"`
} `json:"table"`
} `json:"cost"`
}
if err := json.Unmarshal(data, &parsed); err != nil {
return nil, fmt.Errorf("failed to unmarshal webui_data.json: %w", err)
}
rows := parsed.Cost.Table.Data
if len(rows) == 0 {
return nil, nil
}
// Schema: [TTFT(ms), PrefillThpt, ITL(ms), DecodeThpt, TokensPerUser, GPUHours, ActionYAML]
const minColumns = 7
const actionColumnIndex = 6
pareto := make([]nvidiacomv1beta1.ParetoConfig, 0, len(rows))
for _, row := range rows {
if len(row) < minColumns {
continue
}
var actionYAML string
if err := json.Unmarshal(row[actionColumnIndex], &actionYAML); err != nil {
continue
}
var configObj map[string]interface{}
if err := sigsyaml.Unmarshal([]byte(stripYAMLComments(actionYAML)), &configObj); err != nil {
continue
}
if len(configObj) == 0 {
continue
}
configJSON, err := json.Marshal(configObj)
if err != nil {
continue
}
pareto = append(pareto, nvidiacomv1beta1.ParetoConfig{
Config: runtime.RawExtension{Raw: configJSON},
})
}
return pareto, nil
}
// stripYAMLComments removes comment lines (lines whose first non-whitespace character
// is '#') from a YAML string. The profiler prefixes action snippets with comment lines.
func stripYAMLComments(s string) string {
lines := strings.Split(s, "\n")
out := lines[:0] // reuse backing array; write index always <= range read index
for _, line := range lines {
if !strings.HasPrefix(strings.TrimLeft(line, " \t"), "#") {
out = append(out, line)
}
}
return strings.Join(out, "\n")
}
// storeAdditionalResources marshals additional resources to YAML and stores them in DGDR annotations.
// Validates annotation size and fails gracefully if too large.
func (r *DynamoGraphDeploymentRequestReconciler) storeAdditionalResources(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest, resources []*unstructured.Unstructured) error {
if len(resources) == 0 {
return nil
}
var resourcesYAML []byte
for i, res := range resources {
resYAML, err := sigsyaml.Marshal(res.Object)
if err != nil {
return fmt.Errorf("failed to marshal resource %s/%s: %w", res.GetKind(), res.GetName(), err)
}
if i > 0 {
resourcesYAML = append(resourcesYAML, []byte("\n---\n")...)
}
resourcesYAML = append(resourcesYAML, resYAML...)
}
// Validate size before storing
if len(resourcesYAML) > MaxAnnotationSize {
return fmt.Errorf("additional resources YAML size (%d bytes) exceeds maximum annotation size (%d bytes); "+
"consider reducing the number of resources or storing them separately",
len(resourcesYAML), MaxAnnotationSize)
}
if dgdr.Annotations == nil {
dgdr.Annotations = make(map[string]string)
}
dgdr.Annotations[AnnotationAdditionalResources] = string(resourcesYAML)
return r.Update(ctx, dgdr)
}
// extractResourcesFromYAML parses multi-document YAML from profiling output,
// extracting the DynamoGraphDeployment and any ConfigMaps that should be deployed with it.
func (r *DynamoGraphDeploymentRequestReconciler) extractResourcesFromYAML(yamlContent []byte) (*dgdv1alpha1.DynamoGraphDeployment, []*unstructured.Unstructured, error) {
decoder := yaml.NewYAMLOrJSONDecoder(bytes.NewReader(yamlContent), 4096)
var dgd *dgdv1alpha1.DynamoGraphDeployment
var additionalResources []*unstructured.Unstructured
for {
obj := &unstructured.Unstructured{}
if err := decoder.Decode(obj); err != nil {
if err == io.EOF {
break
}
// Skip invalid documents and continue
continue
}
// Skip empty objects
if obj.GetKind() == "" {
continue
}
if obj.GetKind() == "DynamoGraphDeployment" {
dgd = &dgdv1alpha1.DynamoGraphDeployment{}
if err := runtime.DefaultUnstructuredConverter.FromUnstructured(obj.Object, dgd); err != nil {
return nil, nil, fmt.Errorf("failed to convert to DynamoGraphDeployment: %w", err)
}
} else {
// Store ConfigMaps or other resources for deployment
additionalResources = append(additionalResources, obj)
}
}
if dgd == nil {
return nil, nil, fmt.Errorf("no DynamoGraphDeployment found in YAML content")
}
return dgd, additionalResources, nil
}
// extractDGDFromYAML is a convenience wrapper that extracts only the DGD (used by tests)
func (r *DynamoGraphDeploymentRequestReconciler) extractDGDFromYAML(yamlContent []byte) (*dgdv1alpha1.DynamoGraphDeployment, error) {
dgd, _, err := r.extractResourcesFromYAML(yamlContent)
return dgd, err
}
// updateDeploymentInfo populates status.deploymentInfo from DGD service replica counts.
func updateDeploymentInfo(dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest, dgd *dgdv1alpha1.DynamoGraphDeployment) bool {
var totalReplicas, totalAvailable int32
for _, svc := range dgd.Status.Services {
totalReplicas += svc.Replicas
if svc.AvailableReplicas != nil {
totalAvailable += *svc.AvailableReplicas
}
}
// Short-circuit if nothing changed
if cur := dgdr.Status.DeploymentInfo; cur != nil &&
cur.Replicas != nil && *cur.Replicas == totalReplicas &&
cur.AvailableReplicas != nil && *cur.AvailableReplicas == totalAvailable {
return false
}
dgdr.Status.DeploymentInfo = &nvidiacomv1beta1.DeploymentInfoStatus{
Replicas: &totalReplicas,
AvailableReplicas: &totalAvailable,
}
return true
}
// setSucceededCondition sets the aggregate Succeeded condition based on the current phase.
func setSucceededCondition(dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest, phase nvidiacomv1beta1.DGDRPhase) {
var status metav1.ConditionStatus
var reason, message string
switch phase {
case nvidiacomv1beta1.DGDRPhasePending, "":
status, reason, message = metav1.ConditionFalse, "Pending", "DGDR is pending"
case nvidiacomv1beta1.DGDRPhaseProfiling:
status, reason, message = metav1.ConditionFalse, "Profiling", "Profiling is in progress"
case nvidiacomv1beta1.DGDRPhaseReady:
status, reason, message = metav1.ConditionTrue, "SpecGenerated", "Profiling complete, spec available"
case nvidiacomv1beta1.DGDRPhaseDeploying:
status, reason, message = metav1.ConditionFalse, "Deploying", "Deployment is in progress"
case nvidiacomv1beta1.DGDRPhaseDeployed:
status, reason, message = metav1.ConditionTrue, "Deployed", "Deployment is healthy"
case nvidiacomv1beta1.DGDRPhaseFailed:
status, reason, message = metav1.ConditionFalse, "Failed", "DGDR has failed"
default:
status, reason, message = metav1.ConditionFalse, "Unknown", "Unknown phase"
}
meta.SetStatusCondition(&dgdr.Status.Conditions, metav1.Condition{
Type: nvidiacomv1beta1.ConditionTypeSucceeded,
Status: status,
ObservedGeneration: dgdr.Generation,
Reason: reason,
Message: message,
})
}
// updatePhaseAndRequeue updates the DGDR phase and requeues
func (r *DynamoGraphDeploymentRequestReconciler) updatePhaseAndRequeue(ctx context.Context, dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest, phase nvidiacomv1beta1.DGDRPhase, message string) (ctrl.Result, error) {
logger := log.FromContext(ctx)
logger.Info("Updating DGDR phase", "name", dgdr.Name, "phase", phase, "message", message)
dgdr.Status.Phase = phase
setSucceededCondition(dgdr, phase)
if err := r.Status().Update(ctx, dgdr); err != nil {
return ctrl.Result{}, err
}
return ctrl.Result{Requeue: true}, nil
}
// updatePhaseWithCondition updates phase and adds/updates a condition
func (r *DynamoGraphDeploymentRequestReconciler) updatePhaseWithCondition(
ctx context.Context,
dgdr *nvidiacomv1beta1.DynamoGraphDeploymentRequest,
phase nvidiacomv1beta1.DGDRPhase,
conditionType string,
status metav1.ConditionStatus,
reason string,
message string,
) (ctrl.Result, error) {
dgdr.Status.Phase = phase
setSucceededCondition(dgdr, phase)
condition := metav1.Condition{
Type: conditionType,
Status: status,
ObservedGeneration: dgdr.Generation,
LastTransitionTime: metav1.Now(),
Reason: reason,
Message: message,
}
dgdr.AddStatusCondition(condition)
if err := r.Status().Update(ctx, dgdr); err != nil {
return ctrl.Result{}, err
}
return ctrl.Result{Requeue: true}, nil
}
// SetupWithManager sets up the controller with the Manager
func (r *DynamoGraphDeploymentRequestReconciler) SetupWithManager(mgr ctrl.Manager) error {
return ctrl.NewControllerManagedBy(mgr).
For(&nvidiacomv1beta1.DynamoGraphDeploymentRequest{}).
Named(consts.ResourceTypeDynamoGraphDeploymentRequest).
Owns(&batchv1.Job{}, builder.WithPredicates(predicate.Funcs{
// ignore creation cause we don't want to be called again after we create the job
CreateFunc: func(ce event.CreateEvent) bool { return false },
DeleteFunc: func(de event.DeleteEvent) bool { return true },
UpdateFunc: func(de event.UpdateEvent) bool { return true },
GenericFunc: func(ge event.GenericEvent) bool { return true },
})). // Watch Jobs created by this controller (via ownerReference)
// Watch DGDs created by this controller (via label)
Watches(
&dgdv1alpha1.DynamoGraphDeployment{},
handler.EnqueueRequestsFromMapFunc(func(ctx context.Context, obj client.Object) []ctrl.Request {
// Find DGDR by label instead of owner reference
dgd := obj.(*dgdv1alpha1.DynamoGraphDeployment)
dgdrName, hasName := dgd.Labels[nvidiacomv1beta1.LabelDGDRName]
dgdrNamespace, hasNamespace := dgd.Labels[nvidiacomv1beta1.LabelDGDRNamespace]
if !hasName || !hasNamespace {
return nil
}
return []ctrl.Request{{
NamespacedName: types.NamespacedName{
Name: dgdrName,
Namespace: dgdrNamespace,
},
}}
}),
builder.WithPredicates(predicate.Funcs{
// ignore creation cause we don't want to be called again after we create the DGD
CreateFunc: func(ce event.CreateEvent) bool { return false },
DeleteFunc: func(de event.DeleteEvent) bool { return true },
UpdateFunc: func(ue event.UpdateEvent) bool { return true },
GenericFunc: func(ge event.GenericEvent) bool { return true },
}),
).
// Watch output ConfigMaps for profiling sub-phase updates (via label)
Watches(
&corev1.ConfigMap{},
handler.EnqueueRequestsFromMapFunc(func(ctx context.Context, obj client.Object) []ctrl.Request {
// Only trigger for ConfigMaps with DGDR labels (written by the sidecar)
cm := obj.(*corev1.ConfigMap)
dgdrName, hasName := cm.Labels[nvidiacomv1beta1.LabelDGDRName]
dgdrNamespace, hasNamespace := cm.Labels[nvidiacomv1beta1.LabelDGDRNamespace]
if !hasName || !hasNamespace {
return nil
}
return []ctrl.Request{{
NamespacedName: types.NamespacedName{
Name: dgdrName,
Namespace: dgdrNamespace,
},
}}
}),
builder.WithPredicates(predicate.Funcs{
CreateFunc: func(ce event.CreateEvent) bool {
labels := ce.Object.GetLabels()
_, hasName := labels[nvidiacomv1beta1.LabelDGDRName]
_, hasNamespace := labels[nvidiacomv1beta1.LabelDGDRNamespace]
return hasName && hasNamespace
},
UpdateFunc: func(ue event.UpdateEvent) bool {
labels := ue.ObjectNew.GetLabels()
_, hasName := labels[nvidiacomv1beta1.LabelDGDRName]
_, hasNamespace := labels[nvidiacomv1beta1.LabelDGDRNamespace]
return hasName && hasNamespace
},
DeleteFunc: func(de event.DeleteEvent) bool { return false },
GenericFunc: func(ge event.GenericEvent) bool { return false },
}),
).
// Set the event filter to ignore resources handled by other controllers in namespace-restricted mode
WithEventFilter(commonController.EphemeralDeploymentEventFilter(r.Config, r.RuntimeConfig)).
Complete(observability.NewObservedReconciler(r, consts.ResourceTypeDynamoGraphDeploymentRequest))
}