feat: use vllm multiprocessing in multinode scenario (#6191)

Signed-off-by: Julien Mancuso <jmancuso@nvidia.com>

deploy/operator/internal/consts/consts.go

@@ -90,6 +90,17 @@ const (
 	// Records which operator version created the resource, enabling version-gated behavior changes.
 	KubeAnnotationDynamoOperatorOriginVersion = "nvidia.com/dynamo-operator-origin-version"
 
+	// vLLM distributed executor backend override annotation.
+	// Users can set this on a DGD to explicitly choose "mp" or "ray" for multi-node vLLM deployments.
+	// When present, takes priority over the version-based default.
+	KubeAnnotationVLLMDistributedExecutorBackend = "nvidia.com/vllm-distributed-executor-backend"
+
+	// VLLMMpMasterPort is the default port for vLLM multiprocessing coordination between nodes.
+	VLLMMpMasterPort = "29500"
+
+	// VLLMNixlSideChannelHostEnvVar is the env var that tells vLLM which host IP to use for the NIXL side channel.
+	VLLMNixlSideChannelHostEnvVar = "VLLM_NIXL_SIDE_CHANNEL_HOST"
+
 	// Metrics related constants
 	KubeAnnotationEnableMetrics  = "nvidia.com/enable-metrics"  // User-provided annotation to control metrics
 	KubeLabelMetricsEnabled      = "nvidia.com/metrics-enabled" // Controller-managed label for pod selection

deploy/operator/internal/dynamo/backend_sglang.go

@@ -62,7 +62,7 @@ func (b *SGLangBackend) UpdateContainer(container *corev1.Container, numberOfNod
 	injectFlagsIntoContainerCommand(container, flags, needsShell, "sglang")
 }
 
-func (b *SGLangBackend) UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string) {
+func (b *SGLangBackend) UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string, multinodeDeployer MultinodeDeployer) {
 	// do nothing
 }
 

deploy/operator/internal/dynamo/backend_trtllm.go

@@ -17,24 +17,6 @@ type TRTLLMBackend struct {
 	MpiRunSecretName string
 }
 
-// shellQuoteForBashC quotes a string for safe use inside a single-quoted bash -c '...' command.
-// Since the outer context is already single-quoted, we CANNOT use single quotes here.
-// Instead, we use double quotes and escape characters that are special inside double quotes:
-// backslash, double quote, dollar sign, and backtick.
-// Any embedded single quotes are also escaped (\') since the result lives inside bash -c '...'.
-func shellQuoteForBashC(s string) string {
-	if strings.ContainsAny(s, " \t\n'\"\\{}[]$`!") {
-		escaped := s
-		escaped = strings.ReplaceAll(escaped, `\`, `\\`) // must be first
-		escaped = strings.ReplaceAll(escaped, `"`, `\"`)
-		escaped = strings.ReplaceAll(escaped, `$`, `\$`)
-		escaped = strings.ReplaceAll(escaped, "`", "\\`")
-		escaped = strings.ReplaceAll(escaped, "'", `'"'"'`) // end single-quote, literal ', restart single-quote
-		return `"` + escaped + `"`
-	}
-	return s
-}
-
 func (b *TRTLLMBackend) UpdateContainer(container *corev1.Container, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string, multinodeDeployer MultinodeDeployer) {
 	// Check for volumeMounts with useAsCompilationCache=true
 	for _, volumeMount := range component.VolumeMounts {
@@ -92,7 +74,7 @@ func (b *TRTLLMBackend) UpdateContainer(container *corev1.Container, numberOfNod
 	}
 }
 
-func (b *TRTLLMBackend) UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string) {
+func (b *TRTLLMBackend) UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string, multinodeDeployer MultinodeDeployer) {
 	// Add SSH keypair volume for TRTLLM multinode deployments
 	if numberOfNodes > 1 {
 		sshVolume := corev1.Volume{

deploy/operator/internal/dynamo/backend_trtllm_test.go

@@ -448,7 +448,7 @@ func TestTRTLLMBackend_UpdatePodSpec(t *testing.T) {
 			component := &v1alpha1.DynamoComponentDeploymentSharedSpec{}
 
 			// Call UpdatePodSpec
-			backend.UpdatePodSpec(podSpec, tt.numberOfNodes, tt.role, component, "test-service")
+			backend.UpdatePodSpec(podSpec, tt.numberOfNodes, tt.role, component, "test-service", tt.multinodeDeployer)
 
 			// Check volume count
 			if len(podSpec.Volumes) != tt.expectedVolumeCount {

deploy/operator/internal/dynamo/backend_vllm.go

@@ -6,6 +6,8 @@ import (
 	"strings"
 
 	"github.com/ai-dynamo/dynamo/deploy/operator/api/v1alpha1"
+	commonconsts "github.com/ai-dynamo/dynamo/deploy/operator/internal/consts"
+	"github.com/ai-dynamo/dynamo/deploy/operator/internal/featuregate"
 	corev1 "k8s.io/api/core/v1"
 	"sigs.k8s.io/controller-runtime/pkg/log"
 )
@@ -25,7 +27,18 @@ func (b *VLLMBackend) UpdateContainer(container *corev1.Container, numberOfNodes
 
 	if isMultinode {
 		// Apply multinode-specific argument modifications
-		updateVLLMMultinodeArgs(container, role, serviceName, multinodeDeployer, component.Resources, numberOfNodes)
+		updateVLLMMultinodeArgs(container, role, serviceName, multinodeDeployer, component.Resources, numberOfNodes, component.Annotations)
+
+		if shouldUseMpBackend(component.Annotations) {
+			container.Env = append(container.Env, corev1.EnvVar{
+				Name: commonconsts.VLLMNixlSideChannelHostEnvVar,
+				ValueFrom: &corev1.EnvVarSource{
+					FieldRef: &corev1.ObjectFieldSelector{
+						FieldPath: "status.podIP",
+					},
+				},
+			})
+		}
 
 		// Remove probes for multinode worker and leader
 		if role == RoleWorker {
@@ -65,21 +78,64 @@ func (b *VLLMBackend) UpdateContainer(container *corev1.Container, numberOfNodes
 	}
 }
 
-func (b *VLLMBackend) UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string) {
-	// do nothing
+func (b *VLLMBackend) UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string, multinodeDeployer MultinodeDeployer) {
+	if numberOfNodes <= 1 || role != RoleWorker || !shouldUseMpBackend(component.Annotations) {
+		return
+	}
+
+	if len(podSpec.Containers) == 0 {
+		return
+	}
+
+	leaderHostname := multinodeDeployer.GetLeaderHostname(serviceName)
+	mainImage := podSpec.Containers[0].Image
+
+	waitScript := fmt.Sprintf(`import socket, time
+host, port = "%s", %s
+print(f"Waiting for leader master port at {host}:{port}...", flush=True)
+start = time.monotonic()
+last_status = start
+last_err = ""
+while True:
+    try:
+        s = socket.create_connection((host, port), timeout=2)
+        s.close()
+        elapsed = time.monotonic() - start
+        print(f"Leader master port ready (waited {elapsed:.1f}s)", flush=True)
+        break
+    except Exception as e:
+        last_err = f"{type(e).__name__}: {e}"
+    now = time.monotonic()
+    if now - last_status >= 30:
+        print(f"Still waiting for {host}:{port}... ({now - start:.0f}s elapsed, last error: {last_err})", flush=True)
+        last_status = now
+    time.sleep(2)
+`, leaderHostname, commonconsts.VLLMMpMasterPort)
+
+	initContainer := corev1.Container{
+		Name:    "wait-for-leader-mp",
+		Image:   mainImage,
+		Command: []string{"python3", "-c", waitScript},
+	}
+
+	podSpec.InitContainers = append(podSpec.InitContainers, initContainer)
 }
 
-// updateVLLMMultinodeArgs will inject Ray-specific flags for tensor parallel multinode deployments
-// OR data parallel flags for data parallel multinode deployments
-func updateVLLMMultinodeArgs(container *corev1.Container, role Role, serviceName string, multinodeDeployer MultinodeDeployer, resources *v1alpha1.Resources, numberOfNodes int32) {
+// updateVLLMMultinodeArgs dispatches to the appropriate injection function based on
+// parallelism strategy (TP/PP distributed vs data-parallel) and executor backend (mp vs ray).
+func updateVLLMMultinodeArgs(container *corev1.Container, role Role, serviceName string, multinodeDeployer MultinodeDeployer, resources *v1alpha1.Resources, numberOfNodes int32, annotations map[string]string) {
 	expandedArgs := getExpandedArgs(container)
-	if needsRayDistributedLaunch(expandedArgs, resources) {
+	needsDistributed := needsTensorParallelMultinodeLaunch(expandedArgs, resources)
+
+	if needsDistributed && shouldUseMpBackend(annotations) {
+		injectMpDistributedLaunchFlags(container, role, serviceName, multinodeDeployer, numberOfNodes)
+	} else if needsDistributed {
 		injectRayDistributedLaunchFlags(container, role, serviceName, multinodeDeployer)
-	} else if needsDataParallelLaunch(expandedArgs, resources) {
+	} else if needsDataParallelMultinodeLaunch(expandedArgs, resources) {
 		injectDataParallelLaunchFlags(container, role, serviceName, multinodeDeployer, resources, numberOfNodes)
 	} else {
 		logger := log.Log.WithName("vllm-backend")
-		logger.Info("No need to inject Ray-specific or data parallel flags for multinode deployments", "args", strings.Join(container.Args, " "))
+		logger.Info("No need to inject tensor or data parallel flags for multinode deployments", "args", strings.Join(container.Args, " "))
 	}
 }
 
@@ -93,6 +149,61 @@ func getExpandedArgs(container *corev1.Container) []string {
 	return expandedArgs
 }
 
+// shouldUseMpBackend determines whether to use multiprocessing (mp) or Ray for vLLM
+// multi-node distributed launches.
+//
+// Decision logic:
+//  1. Explicit override annotation takes priority (user set "mp" or "ray")
+//  2. Operator origin version feature gate: uses featuregate.VLLMMultiprocessing
+func shouldUseMpBackend(annotations map[string]string) bool {
+	logger := log.Log.WithName("vllm-backend")
+
+	// Step 1: Check explicit override
+	if override, exists := annotations[commonconsts.KubeAnnotationVLLMDistributedExecutorBackend]; exists {
+		switch strings.ToLower(override) {
+		case "mp":
+			logger.Info("Using mp backend (explicit override)")
+			return true
+		case "ray":
+			logger.Info("Using ray backend (explicit override)")
+			return false
+		default:
+			logger.Info("Ignoring invalid vllm-distributed-executor-backend annotation value, falling through to version check",
+				"value", override)
+		}
+	}
+
+	// Step 2: Check operator origin version gate
+	return featuregate.VLLMMultiprocessing.IsEnabled(annotations)
+}
+
+// injectMpDistributedLaunchFlags injects vLLM multiprocessing flags for multi-node TP/PP deployments.
+//
+// Leader: runs the original vLLM command with --distributed-executor-backend mp,
+// --nnodes, --node-rank 0, --master-addr, --master-port
+//
+// Worker: runs the same vLLM command with --headless, --node-rank <rank>, and the same
+// coordination flags. An init container (injected via UpdatePodSpec) handles waiting for
+// the leader's master port before the worker's main container starts.
+func injectMpDistributedLaunchFlags(container *corev1.Container, role Role, serviceName string, multinodeDeployer MultinodeDeployer, numberOfNodes int32) {
+	leaderHostname := multinodeDeployer.GetLeaderHostname(serviceName)
+	mpFlags := fmt.Sprintf("--distributed-executor-backend mp --nnodes %d --master-addr %s --master-port %s",
+		numberOfNodes, leaderHostname, commonconsts.VLLMMpMasterPort)
+
+	needsShell := false
+
+	switch role {
+	case RoleLeader:
+		mpFlags += " --node-rank 0"
+	case RoleWorker:
+		nodeRank, needsShellForRank := multinodeDeployer.GetNodeRank()
+		needsShell = needsShellForRank
+		mpFlags += fmt.Sprintf(" --node-rank %s --headless", nodeRank)
+	}
+
+	injectFlagsIntoContainerCommand(container, mpFlags, needsShell, "vllm")
+}
+
 func injectRayDistributedLaunchFlags(container *corev1.Container, role Role, serviceName string, multinodeDeployer MultinodeDeployer) {
 	switch role {
 	case RoleLeader:
@@ -179,9 +290,9 @@ func injectDataParallelLaunchFlags(container *corev1.Container, role Role, servi
 	injectFlagsIntoContainerCommand(container, strings.Join(flags, " "), needsShell, "vllm")
 }
 
-// if world size (within DP rank) > GPU count, then we need to inject ray
-// world size = tensor parallel size * pipeline parallel size
-func needsRayDistributedLaunch(expandedArgs []string, resources *v1alpha1.Resources) bool {
+// needsMultinodeDistributedLaunch returns true when the model's world size (TP * PP)
+// exceeds the GPU count of a single node, requiring multi-node distribution (via mp or ray).
+func needsTensorParallelMultinodeLaunch(expandedArgs []string, resources *v1alpha1.Resources) bool {
 	containerGPUs := getContainerGPUs(resources)
 	if containerGPUs == 0 {
 		return false
@@ -196,7 +307,7 @@ func getWorldSize(expandedArgs []string) int64 {
 }
 
 // if world size across all DP ranks > GPU count, then we need to inject data parallel multinode coordination
-func needsDataParallelLaunch(expandedArgs []string, resources *v1alpha1.Resources) bool {
+func needsDataParallelMultinodeLaunch(expandedArgs []string, resources *v1alpha1.Resources) bool {
 	dataParallelSize := getFlagValue(expandedArgs, dataParallelSizeFlag)
 	containerGPUs := getContainerGPUs(resources)
 	if containerGPUs == 0 {

deploy/operator/internal/dynamo/graph.go

@@ -867,7 +867,7 @@ func ParseBackendFramework(framework string) (BackendFramework, error) {
 // Each backend (SGLang, VLLM, etc.) implements this interface
 type Backend interface {
 	UpdateContainer(container *corev1.Container, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string, multinodeDeployer MultinodeDeployer)
-	UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string)
+	UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string, multinodeDeployer MultinodeDeployer)
 }
 
 // NoopBackend does no processing - used for non-worker components like frontend, planner, router
@@ -877,7 +877,7 @@ func (b *NoopBackend) UpdateContainer(container *corev1.Container, numberOfNodes
 	// No-op: frontend, planner, router, etc. don't need backend-specific processing
 }
 
-func (b *NoopBackend) UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string) {
+func (b *NoopBackend) UpdatePodSpec(podSpec *corev1.PodSpec, numberOfNodes int32, role Role, component *v1alpha1.DynamoComponentDeploymentSharedSpec, serviceName string, multinodeDeployer MultinodeDeployer) {
 	// No-op: frontend, planner, router, etc. don't need backend-specific processing
 }
 
@@ -1166,7 +1166,7 @@ func GenerateBasePodSpec(
 	podSpec.Volumes = append(podSpec.Volumes, volumes...)
 	podSpec.ImagePullSecrets = controller_common.AppendUniqueImagePullSecrets(podSpec.ImagePullSecrets, imagePullSecrets)
 
-	backend.UpdatePodSpec(&podSpec, numberOfNodes, role, component, serviceName)
+	backend.UpdatePodSpec(&podSpec, numberOfNodes, role, component, serviceName, multinodeDeployer)
 
 	// Inject checkpoint configuration if enabled
 	// This handles ALL checkpoint-related modifications:

deploy/operator/internal/dynamo/utils.go

@@ -27,19 +27,39 @@ import (
  *
  * The needsShell flag indicates when environment variables require shell interpretation
  */
+// shellQuoteForBashC quotes a string so it survives shell interpretation inside sh -c.
+// Simple args (flags, paths) pass through unchanged; args containing special characters
+// (JSON, env vars, spaces, quotes) are wrapped in double quotes with inner escaping.
+func shellQuoteForBashC(s string) string {
+	if strings.ContainsAny(s, " \t\n'\"\\{}[]$`!") {
+		escaped := s
+		escaped = strings.ReplaceAll(escaped, `\`, `\\`) // must be first
+		escaped = strings.ReplaceAll(escaped, `"`, `\"`)
+		escaped = strings.ReplaceAll(escaped, `$`, `\$`)
+		escaped = strings.ReplaceAll(escaped, "`", "\\`")
+		escaped = strings.ReplaceAll(escaped, "'", `'"'"'`)
+		return `"` + escaped + `"`
+	}
+	return s
+}
+
 func injectFlagsIntoContainerCommand(container *corev1.Container, flags string, needsShell bool, framework string) {
 	if len(container.Command) > 0 && isPythonCommand(container.Command[0]) {
 		// Direct python command case
 		if needsShell {
-			// Transform to shell wrapper for env var interpretation
+			// Transform to shell wrapper for env var interpretation.
+			// Quote each original arg individually so JSON and other special
+			// characters survive shell interpretation.
 			fullCommand := strings.Join(container.Command, " ")
-			originalArgs := strings.Join(container.Args, " ")
+			quotedArgs := make([]string, len(container.Args))
+			for i, arg := range container.Args {
+				quotedArgs[i] = shellQuoteForBashC(arg)
+			}
+			originalArgs := strings.Join(quotedArgs, " ")
 			var shellCommand string
 			if len(container.Args) > 0 {
-				// Use exec to ensure PID 1 is given to the python command
 				shellCommand = fmt.Sprintf("exec %s %s %s", fullCommand, originalArgs, flags)
 			} else {
-				// Use exec to ensure PID 1 is given to the python command
 				shellCommand = fmt.Sprintf("exec %s %s", fullCommand, flags)
 			}
 			container.Command = []string{"sh", "-c"}

deploy/operator/internal/featuregate/gates.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 featuregate
+
+import semver "github.com/Masterminds/semver/v3"
+
+// Feature gates gated on the operator origin version (the operator version that
+// first reconciled / created the DGD resource).
+
+var (
+	// VLLMMultiprocessing gates the use of vLLM native multiprocessing (mp)
+	// instead of Ray for multi-node deployments. Enabled for DGDs originally
+	// created by operator >= 1.0.0-dev.
+	VLLMMultiprocessing = OperatorOriginFeatureGate{
+		Name:             "VLLMMultiprocessing",
+		MinOriginVersion: *semver.MustParse("1.0.0-dev"),
+	}
+)

deploy/operator/internal/featuregate/operator_origin.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 featuregate
+
+import (
+	semver "github.com/Masterminds/semver/v3"
+	"github.com/ai-dynamo/dynamo/deploy/operator/internal/consts"
+	"sigs.k8s.io/controller-runtime/pkg/log"
+)
+
+// OperatorOriginFeatureGate represents a feature gated on the operator version
+// that originally created the DGD (nvidia.com/dynamo-operator-origin-version).
+//
+// The origin version is stamped by the mutating webhook at CREATE time and never
+// changes afterwards. This allows the operator to introduce new default behaviors
+// for newly created resources while preserving backward compatibility for existing ones.
+//
+// When the annotation is absent (pre-upgrade DGD), IsEnabled returns false
+// to preserve backward compatibility.
+type OperatorOriginFeatureGate struct {
+	Name             string         // Human-readable feature name (for logging)
+	MinOriginVersion semver.Version // Minimum origin version required (semver)
+}
+
+// IsEnabled returns true if the origin version in annotations meets or exceeds
+// the gate's MinOriginVersion threshold.
+//
+// Returns false when:
+//   - annotations is nil (no metadata)
+//   - origin version annotation is absent (pre-upgrade DGD)
+//   - origin version is not valid semver
+//   - origin version < MinOriginVersion
+func (fg OperatorOriginFeatureGate) IsEnabled(annotations map[string]string) bool {
+	logger := log.Log.WithName("featuregate").WithValues("feature", fg.Name)
+
+	originVersion, exists := annotations[consts.KubeAnnotationDynamoOperatorOriginVersion]
+	if !exists {
+		logger.V(1).Info("No operator origin version annotation, feature disabled (backward compat)")
+		return false
+	}
+
+	version, err := semver.NewVersion(originVersion)
+	if err != nil {
+		logger.Info("Invalid origin version, feature disabled",
+			"version", originVersion, "error", err.Error())
+		return false
+	}
+
+	enabled := version.Compare(&fg.MinOriginVersion) >= 0
+
+	logger.V(1).Info("Feature gate evaluated",
+		"originVersion", originVersion,
+		"threshold", fg.MinOriginVersion,
+		"enabled", enabled)
+
+	return enabled
+}

deploy/operator/internal/featuregate/operator_origin_test.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 featuregate
+
+import (
+	"testing"
+
+	semver "github.com/Masterminds/semver/v3"
+	"github.com/ai-dynamo/dynamo/deploy/operator/internal/consts"
+)
+
+func TestOperatorOriginFeatureGate_IsEnabled(t *testing.T) {
+	gate := OperatorOriginFeatureGate{
+		Name:             "TestFeature",
+		MinOriginVersion: *semver.MustParse("1.0.0-dev"),
+	}
+
+	tests := []struct {
+		name        string
+		annotations map[string]string
+		want        bool
+	}{
+		{
+			name:        "nil annotations = disabled (backward compat)",
+			annotations: nil,
+			want:        false,
+		},
+		{
+			name:        "empty annotations = disabled (backward compat)",
+			annotations: map[string]string{},
+			want:        false,
+		},
+		{
+			name: "origin version below threshold = disabled",
+			annotations: map[string]string{
+				consts.KubeAnnotationDynamoOperatorOriginVersion: "0.9.0",
+			},
+			want: false,
+		},
+		{
+			name: "origin version at threshold = enabled",
+			annotations: map[string]string{
+				consts.KubeAnnotationDynamoOperatorOriginVersion: "1.0.0-dev",
+			},
+			want: true,
+		},
+		{
+			name: "origin version above threshold (release > pre-release) = enabled",
+			annotations: map[string]string{
+				consts.KubeAnnotationDynamoOperatorOriginVersion: "1.0.0",
+			},
+			want: true,
+		},
+		{
+			name: "origin version well above threshold = enabled",
+			annotations: map[string]string{
+				consts.KubeAnnotationDynamoOperatorOriginVersion: "2.0.0",
+			},
+			want: true,
+		},
+		{
+			name: "pre-release below threshold = disabled",
+			annotations: map[string]string{
+				consts.KubeAnnotationDynamoOperatorOriginVersion: "0.9.0-dev",
+			},
+			want: false,
+		},
+		{
+			name: "invalid origin version = disabled (graceful fallback)",
+			annotations: map[string]string{
+				consts.KubeAnnotationDynamoOperatorOriginVersion: "not-a-version",
+			},
+			want: false,
+		},
+		{
+			name: "fallback version 0.0.0-unknown = disabled",
+			annotations: map[string]string{
+				consts.KubeAnnotationDynamoOperatorOriginVersion: "0.0.0-unknown",
+			},
+			want: false,
+		},
+		{
+			name: "unrelated annotations without origin version = disabled",
+			annotations: map[string]string{
+				"some.other/annotation": "value",
+			},
+			want: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := gate.IsEnabled(tt.annotations)
+			if got != tt.want {
+				t.Errorf("IsEnabled() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestOperatorOriginFeatureGate_DifferentThresholds(t *testing.T) {
+	annotations := map[string]string{
+		consts.KubeAnnotationDynamoOperatorOriginVersion: "0.9.0",
+	}
+
+	tests := []struct {
+		name             string
+		minOriginVersion semver.Version
+		want             bool
+	}{
+		{
+			name:             "threshold below origin = enabled",
+			minOriginVersion: *semver.MustParse("0.8.0"),
+			want:             true,
+		},
+		{
+			name:             "threshold equal to origin = enabled",
+			minOriginVersion: *semver.MustParse("0.9.0"),
+			want:             true,
+		},
+		{
+			name:             "threshold above origin = disabled",
+			minOriginVersion: *semver.MustParse("1.0.0"),
+			want:             false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			gate := OperatorOriginFeatureGate{
+				Name:             "TestFeature",
+				MinOriginVersion: tt.minOriginVersion,
+			}
+			got := gate.IsEnabled(annotations)
+			if got != tt.want {
+				t.Errorf("IsEnabled() with threshold %v = %v, want %v", tt.minOriginVersion, got, tt.want)
+			}
+		})
+	}
+}

deploy/operator/internal/webhook/validation/dynamographdeployment.go

@@ -441,6 +441,17 @@ func (v *DynamoGraphDeploymentValidator) validateAnnotations() error {
 		}
 	}
 
+	// Validate vLLM distributed executor backend override
+	if value, exists := annotations[consts.KubeAnnotationVLLMDistributedExecutorBackend]; exists {
+		switch strings.ToLower(value) {
+		case "mp", "ray":
+			// valid
+		default:
+			errs = append(errs, fmt.Errorf("annotation %s has invalid value %q: must be \"mp\" or \"ray\"",
+				consts.KubeAnnotationVLLMDistributedExecutorBackend, value))
+		}
+	}
+
 	return errors.Join(errs...)
 }
 

deploy/operator/internal/webhook/validation/dynamographdeployment_test.go

@@ -677,6 +677,79 @@ func TestDynamoGraphDeploymentValidator_Validate(t *testing.T) {
 			wantErr: false,
 		},
 		// Annotation validation test cases
+		{
+			name: "valid annotation vllm-distributed-executor-backend=mp",
+			deployment: &nvidiacomv1alpha1.DynamoGraphDeployment{
+				ObjectMeta: metav1.ObjectMeta{
+					Name:      "test-graph",
+					Namespace: "default",
+					Annotations: map[string]string{
+						consts.KubeAnnotationVLLMDistributedExecutorBackend: "mp",
+					},
+				},
+				Spec: nvidiacomv1alpha1.DynamoGraphDeploymentSpec{
+					Services: map[string]*nvidiacomv1alpha1.DynamoComponentDeploymentSharedSpec{
+						"main": {},
+					},
+				},
+			},
+			wantErr: false,
+		},
+		{
+			name: "valid annotation vllm-distributed-executor-backend=ray",
+			deployment: &nvidiacomv1alpha1.DynamoGraphDeployment{
+				ObjectMeta: metav1.ObjectMeta{
+					Name:      "test-graph",
+					Namespace: "default",
+					Annotations: map[string]string{
+						consts.KubeAnnotationVLLMDistributedExecutorBackend: "ray",
+					},
+				},
+				Spec: nvidiacomv1alpha1.DynamoGraphDeploymentSpec{
+					Services: map[string]*nvidiacomv1alpha1.DynamoComponentDeploymentSharedSpec{
+						"main": {},
+					},
+				},
+			},
+			wantErr: false,
+		},
+		{
+			name: "valid annotation vllm-distributed-executor-backend case insensitive MP",
+			deployment: &nvidiacomv1alpha1.DynamoGraphDeployment{
+				ObjectMeta: metav1.ObjectMeta{
+					Name:      "test-graph",
+					Namespace: "default",
+					Annotations: map[string]string{
+						consts.KubeAnnotationVLLMDistributedExecutorBackend: "MP",
+					},
+				},
+				Spec: nvidiacomv1alpha1.DynamoGraphDeploymentSpec{
+					Services: map[string]*nvidiacomv1alpha1.DynamoComponentDeploymentSharedSpec{
+						"main": {},
+					},
+				},
+			},
+			wantErr: false,
+		},
+		{
+			name: "invalid annotation vllm-distributed-executor-backend",
+			deployment: &nvidiacomv1alpha1.DynamoGraphDeployment{
+				ObjectMeta: metav1.ObjectMeta{
+					Name:      "test-graph",
+					Namespace: "default",
+					Annotations: map[string]string{
+						consts.KubeAnnotationVLLMDistributedExecutorBackend: "invalid",
+					},
+				},
+				Spec: nvidiacomv1alpha1.DynamoGraphDeploymentSpec{
+					Services: map[string]*nvidiacomv1alpha1.DynamoComponentDeploymentSharedSpec{
+						"main": {},
+					},
+				},
+			},
+			wantErr: true,
+			errMsg:  `annotation nvidia.com/vllm-distributed-executor-backend has invalid value "invalid": must be "mp" or "ray"`,
+		},
 		{
 			name: "no annotations is valid",
 			deployment: &nvidiacomv1alpha1.DynamoGraphDeployment{
@@ -765,6 +838,46 @@ func TestDynamoGraphDeploymentValidator_Validate(t *testing.T) {
 			wantErr: true,
 			errMsg:  `annotation nvidia.com/dynamo-operator-origin-version has invalid value "not-a-version": must be valid semver`,
 		},
+		{
+			name: "both annotations valid",
+			deployment: &nvidiacomv1alpha1.DynamoGraphDeployment{
+				ObjectMeta: metav1.ObjectMeta{
+					Name:      "test-graph",
+					Namespace: "default",
+					Annotations: map[string]string{
+						consts.KubeAnnotationDynamoOperatorOriginVersion:    "1.0.0",
+						consts.KubeAnnotationVLLMDistributedExecutorBackend: "mp",
+					},
+				},
+				Spec: nvidiacomv1alpha1.DynamoGraphDeploymentSpec{
+					Services: map[string]*nvidiacomv1alpha1.DynamoComponentDeploymentSharedSpec{
+						"main": {},
+					},
+				},
+			},
+			wantErr: false,
+		},
+		{
+			name: "both annotations invalid",
+			deployment: &nvidiacomv1alpha1.DynamoGraphDeployment{
+				ObjectMeta: metav1.ObjectMeta{
+					Name:      "test-graph",
+					Namespace: "default",
+					Annotations: map[string]string{
+						consts.KubeAnnotationDynamoOperatorOriginVersion:    "bad",
+						consts.KubeAnnotationVLLMDistributedExecutorBackend: "invalid",
+					},
+				},
+				Spec: nvidiacomv1alpha1.DynamoGraphDeploymentSpec{
+					Services: map[string]*nvidiacomv1alpha1.DynamoComponentDeploymentSharedSpec{
+						"main": {},
+					},
+				},
+			},
+			wantErr:     true,
+			errMsg:      "annotation nvidia.com/dynamo-operator-origin-version has invalid value \"bad\": must be valid semver\nannotation nvidia.com/vllm-distributed-executor-backend has invalid value \"invalid\": must be \"mp\" or \"ray\"",
+			errContains: true,
+		},
 	}
 
 	for _, tt := range tests {

examples/backends/vllm/deploy/disagg-multinode.yaml

@@ -39,10 +39,16 @@ spec:
           - -m
           - dynamo.vllm
           args:
-            - --model
-            - Qwen/Qwen3-0.6B
-            - --tensor-parallel-size
-            - "2"
+          - --model
+          - Qwen/Qwen3-0.6B
+          - --tensor-parallel-size
+          - "2"
+          - --is-decode-worker
+          - --connector
+          - none
+          - --kv-transfer-config
+          - '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id":
+            "vllm-disagg-decode-engine-0abc123"}'
     prefill:
       envFromSecret: hf-token-secret
       componentType: worker
@@ -61,8 +67,14 @@ spec:
           - -m
           - dynamo.vllm
           args:
-            - --model
-            - Qwen/Qwen3-0.6B
-            - --is-prefill-worker
-            - --tensor-parallel-size
-            - "2"
+          - --model
+          - Qwen/Qwen3-0.6B
+          - --tensor-parallel-size
+          - "2"
+          - --is-prefill-worker
+          - --connector
+          - none
+          - --kv-transfer-config
+          - '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id":
+            "vllm-disagg-prefill-engine-0abc123"}'
+