feat(gms): operator-managed GMS checkpoint/restore support (#8153)

deploy/operator/internal/gms/runtime.go

+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+package gms
+
+import (
+	"path/filepath"
+
+	corev1 "k8s.io/api/core/v1"
+	"k8s.io/utils/ptr"
+)
+
+const (
+	// ServerContainerName is the name of the GMS server init sidecar.
+	ServerContainerName = "gms-server"
+
+	// SharedVolumeName is the emptyDir volume shared between the GMS server
+	// sidecar and the main workload container for UDS sockets.
+	SharedVolumeName = "gms-shared"
+
+	// SharedMountPath is the mount path for the shared GMS socket directory.
+	SharedMountPath = "/shared"
+
+	// DRAClaimName is the pod-level DRA ResourceClaim name used by both the
+	// main container and GMS sidecars.
+	DRAClaimName = "shared-gpu"
+
+	// ControlVolumeName is the checkpoint-specific control volume name.
+	ControlVolumeName = "gms-control"
+
+	// ControlDir is the mount path for the checkpoint control volume.
+	ControlDir = "/tmp/gms-control"
+
+	readyFile = "gms-ready"
+
+	serverSidecarModule = "gpu_memory_service.cli.server"
+)
+
+// EnsureServerSidecar adds the GMS server as a restartable init sidecar with a
+// startup probe. Used for checkpoint jobs and steady-state pods where the main
+// container needs GMS sockets before starting.
+func EnsureServerSidecar(podSpec *corev1.PodSpec, mainContainer *corev1.Container) {
+	if podSpec == nil || mainContainer == nil {
+		return
+	}
+
+	ensureSharedVolume(podSpec, mainContainer)
+
+	sidecar := serverContainer(mainContainer.Image)
+	sidecar.RestartPolicy = ptr.To(corev1.ContainerRestartPolicyAlways)
+	sidecar.StartupProbe = &corev1.Probe{
+		ProbeHandler: corev1.ProbeHandler{
+			Exec: &corev1.ExecAction{
+				Command: []string{"test", "-f", filepath.Join(SharedMountPath, readyFile)},
+			},
+		},
+		PeriodSeconds:    1,
+		FailureThreshold: 300, // 1s * 300 = 5 min
+	}
+	copyDeviceClaims(mainContainer, &sidecar)
+	// Idempotent — EnsureServerSidecar may be called by both the
+	// steady-state operator path and the checkpoint overlay.
+	for i := range podSpec.InitContainers {
+		if podSpec.InitContainers[i].Name == sidecar.Name {
+			return
+		}
+	}
+	podSpec.InitContainers = append(podSpec.InitContainers, sidecar)
+}
+
+// BuildServerContainer prepares the shared GMS volume/env and returns a GMS
+// server container suitable for use as a regular sidecar. The caller must
+// append the returned container to podSpec.Containers.
+//
+// Used for restore pods where the main container is CRIU-restored and does not
+// need GMS sockets at startup. The gms-loader polls for sockets internally.
+func BuildServerContainer(podSpec *corev1.PodSpec, mainContainer *corev1.Container) corev1.Container {
+	ensureSharedVolume(podSpec, mainContainer)
+	sidecar := serverContainer(mainContainer.Image)
+	copyDeviceClaims(mainContainer, &sidecar)
+	return sidecar
+}
+
+// FindServerContainer returns a pointer to the GMS server container, checking
+// both init containers and regular containers. Returns nil if not present.
+func FindServerContainer(podSpec *corev1.PodSpec) *corev1.Container {
+	if podSpec == nil {
+		return nil
+	}
+	for i := range podSpec.InitContainers {
+		if podSpec.InitContainers[i].Name == ServerContainerName {
+			return &podSpec.InitContainers[i]
+		}
+	}
+	for i := range podSpec.Containers {
+		if podSpec.Containers[i].Name == ServerContainerName {
+			return &podSpec.Containers[i]
+		}
+	}
+	return nil
+}
+
+// ensureSharedVolume adds the shared GMS socket volume, mounts, and env vars.
+// Idempotent — may be called by both steady-state and checkpoint paths.
+func ensureSharedVolume(podSpec *corev1.PodSpec, mainContainer *corev1.Container) {
+	hasVolume := false
+	for _, v := range podSpec.Volumes {
+		if v.Name == SharedVolumeName {
+			hasVolume = true
+			break
+		}
+	}
+	if !hasVolume {
+		podSpec.Volumes = append(podSpec.Volumes, corev1.Volume{
+			Name:         SharedVolumeName,
+			VolumeSource: corev1.VolumeSource{EmptyDir: &corev1.EmptyDirVolumeSource{}},
+		})
+	}
+
+	// Mount and env injection checked independently of volume existence —
+	// another code path may have added the volume without configuring main.
+	hasMount := false
+	for _, m := range mainContainer.VolumeMounts {
+		if m.Name == SharedVolumeName {
+			hasMount = true
+			break
+		}
+	}
+	if !hasMount {
+		mainContainer.VolumeMounts = append(mainContainer.VolumeMounts, corev1.VolumeMount{Name: SharedVolumeName, MountPath: SharedMountPath})
+	}
+
+	hasEnv := false
+	for _, e := range mainContainer.Env {
+		if e.Name == "GMS_SOCKET_DIR" {
+			hasEnv = true
+			break
+		}
+	}
+	if !hasEnv {
+		mainContainer.Env = append(mainContainer.Env,
+			corev1.EnvVar{Name: "TMPDIR", Value: SharedMountPath},
+			corev1.EnvVar{Name: "GMS_SOCKET_DIR", Value: SharedMountPath},
+		)
+	}
+}
+
+// serverContainer builds the base GMS server container without init-specific
+// fields (RestartPolicy, StartupProbe). Callers add those as needed.
+func serverContainer(image string) corev1.Container {
+	return corev1.Container{
+		Name:    ServerContainerName,
+		Image:   image,
+		Command: []string{"python3", "-m", serverSidecarModule},
+		Env: []corev1.EnvVar{
+			{Name: "TMPDIR", Value: SharedMountPath},
+			{Name: "GMS_SOCKET_DIR", Value: SharedMountPath},
+		},
+		VolumeMounts: []corev1.VolumeMount{
+			{Name: SharedVolumeName, MountPath: SharedMountPath},
+		},
+	}
+}
+
+func copyDeviceClaims(src *corev1.Container, dst *corev1.Container) {
+	if src == nil || dst == nil || len(src.Resources.Claims) == 0 {
+		return
+	}
+	claims := make([]corev1.ResourceClaim, len(src.Resources.Claims))
+	copy(claims, src.Resources.Claims)
+	dst.Resources.Claims = append(dst.Resources.Claims, claims...)
+}

deploy/operator/internal/gms/runtime_test.go

+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+package gms
+
+import (
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	corev1 "k8s.io/api/core/v1"
+)
+
+func TestEnsureServerSidecar(t *testing.T) {
+	podSpec := &corev1.PodSpec{
+		Containers: []corev1.Container{{
+			Name:  "main",
+			Image: "test-image:latest",
+			Resources: corev1.ResourceRequirements{
+				Claims: []corev1.ResourceClaim{{Name: DRAClaimName}},
+			},
+		}},
+	}
+
+	EnsureServerSidecar(podSpec, &podSpec.Containers[0])
+
+	require.Len(t, podSpec.Containers, 1)
+	require.Len(t, podSpec.InitContainers, 1)
+
+	main := &podSpec.Containers[0]
+	server := &podSpec.InitContainers[0]
+
+	assert.Equal(t, ServerContainerName, server.Name)
+	assert.Equal(t, []string{"python3", "-m", serverSidecarModule}, server.Command)
+	assert.Equal(t, SharedMountPath, envValue(t, main, "TMPDIR"))
+	assert.Equal(t, SharedMountPath, envValue(t, main, "GMS_SOCKET_DIR"))
+	assert.Equal(t, SharedMountPath, envValue(t, server, "TMPDIR"))
+	assert.Equal(t, SharedMountPath, envValue(t, server, "GMS_SOCKET_DIR"))
+
+	assert.Equal(t, corev1.ContainerRestartPolicyAlways, *server.RestartPolicy)
+	require.NotNil(t, server.StartupProbe)
+	assert.Equal(t, []string{"test", "-f", filepath.Join(SharedMountPath, readyFile)},
+		server.StartupProbe.Exec.Command)
+	assert.Equal(t, int32(1), server.StartupProbe.PeriodSeconds)
+	assert.Equal(t, int32(300), server.StartupProbe.FailureThreshold)
+
+	// DRA claim copied from main
+	assert.Len(t, server.Resources.Claims, 1)
+	assert.Equal(t, DRAClaimName, server.Resources.Claims[0].Name)
+}
+
+func TestBuildServerContainer(t *testing.T) {
+	podSpec := &corev1.PodSpec{
+		Containers: []corev1.Container{{
+			Name:  "main",
+			Image: "test-image:latest",
+			Resources: corev1.ResourceRequirements{
+				Claims: []corev1.ResourceClaim{{Name: DRAClaimName}},
+			},
+		}},
+	}
+
+	server := BuildServerContainer(podSpec, &podSpec.Containers[0])
+
+	// Should not be added to init containers
+	assert.Empty(t, podSpec.InitContainers)
+
+	assert.Equal(t, ServerContainerName, server.Name)
+	assert.Equal(t, []string{"python3", "-m", serverSidecarModule}, server.Command)
+
+	// No init-specific fields
+	assert.Nil(t, server.RestartPolicy)
+	assert.Nil(t, server.StartupProbe)
+
+	// DRA claim copied from main
+	assert.Len(t, server.Resources.Claims, 1)
+	assert.Equal(t, DRAClaimName, server.Resources.Claims[0].Name)
+
+	// Shared volume and env should be set on main
+	main := &podSpec.Containers[0]
+	assert.Equal(t, SharedMountPath, envValue(t, main, "TMPDIR"))
+	assert.Equal(t, SharedMountPath, envValue(t, main, "GMS_SOCKET_DIR"))
+
+	// Shared volume should exist
+	var hasVolume bool
+	for _, v := range podSpec.Volumes {
+		if v.Name == SharedVolumeName {
+			hasVolume = true
+		}
+	}
+	assert.True(t, hasVolume)
+}
+
+func TestEnsureServerSidecarDoesNotAddCheckpointControl(t *testing.T) {
+	podSpec := &corev1.PodSpec{
+		Containers: []corev1.Container{{Name: "main", Image: "test:latest"}},
+	}
+
+	EnsureServerSidecar(podSpec, &podSpec.Containers[0])
+
+	for _, volume := range podSpec.Volumes {
+		if volume.Name == ControlVolumeName {
+			t.Fatal("runtime shaping should not add checkpoint control volume")
+		}
+	}
+	server := FindServerContainer(podSpec)
+	require.NotNil(t, server)
+	for _, env := range server.Env {
+		if env.Name == "GMS_CONTROL_DIR" {
+			t.Fatal("server should not have checkpoint control env")
+		}
+	}
+}
+
+func TestEnsureServerSidecarIdempotent(t *testing.T) {
+	podSpec := &corev1.PodSpec{
+		Containers: []corev1.Container{{Name: "main", Image: "test:latest"}},
+	}
+
+	EnsureServerSidecar(podSpec, &podSpec.Containers[0])
+	EnsureServerSidecar(podSpec, &podSpec.Containers[0])
+
+	assert.Len(t, podSpec.InitContainers, 1)
+	volumeCount := 0
+	for _, v := range podSpec.Volumes {
+		if v.Name == SharedVolumeName {
+			volumeCount++
+		}
+	}
+	assert.Equal(t, 1, volumeCount)
+}
+
+func TestFindServerContainer(t *testing.T) {
+	podSpec := &corev1.PodSpec{
+		Containers: []corev1.Container{{Name: "main", Image: "test:latest"}},
+	}
+	assert.Nil(t, FindServerContainer(podSpec))
+
+	EnsureServerSidecar(podSpec, &podSpec.Containers[0])
+	assert.NotNil(t, FindServerContainer(podSpec))
+	assert.Equal(t, ServerContainerName, FindServerContainer(podSpec).Name)
+}
+
+func envValue(t *testing.T, container *corev1.Container, name string) string {
+	t.Helper()
+	for _, env := range container.Env {
+		if env.Name == name {
+			return env.Value
+		}
+	}
+	t.Fatalf("env %s not found", name)
+	return ""
+}

deploy/snapshot/cmd/snapshotctl/kube.go

@@ -95,14 +95,30 @@ func loadPod(manifestPath string) (*corev1.Pod, error) {
 	if kind := strings.TrimSpace(pod.Kind); kind != "" && kind != "Pod" {
 		return nil, fmt.Errorf("manifest %s is kind %q, expected Pod", manifestPath, kind)
 	}
-	if len(pod.Spec.Containers) != 1 {
+	if len(pod.Spec.Containers) == 0 {
 		return nil, fmt.Errorf(
-			"manifest %s has %d containers; snapshotctl requires exactly one worker container",
+			"manifest %s has no worker containers; snapshotctl requires at least one worker container",
 			manifestPath,
-			len(pod.Spec.Containers),
 		)
 	}
-	if strings.TrimSpace(pod.Spec.Containers[0].Image) == "" {
+	workerContainer := &pod.Spec.Containers[0]
+	if len(pod.Spec.Containers) > 1 {
+		workerContainer = nil
+		for index := range pod.Spec.Containers {
+			if pod.Spec.Containers[index].Name == "main" {
+				workerContainer = &pod.Spec.Containers[index]
+				break
+			}
+		}
+		if workerContainer == nil {
+			return nil, fmt.Errorf(
+				"manifest %s has %d containers; snapshotctl requires a worker container named main",
+				manifestPath,
+				len(pod.Spec.Containers),
+			)
+		}
+	}
+	if strings.TrimSpace(workerContainer.Image) == "" {
 		return nil, fmt.Errorf("manifest %s: worker container image is required", manifestPath)
 	}
 	if strings.TrimSpace(pod.Name) == "" {

deploy/snapshot/protocol/checkpoint.go

@@ -56,15 +56,16 @@ func NewCheckpointJob(podTemplate *corev1.PodTemplateSpec, opts CheckpointJobOpt
 		EnsureLocalhostSeccompProfile(&podTemplate.Spec, opts.SeccompProfile)
 	}
 	if opts.WrapLaunchJob {
-		if len(podTemplate.Spec.Containers) == 0 {
-			return nil, fmt.Errorf("checkpoint job requires one worker container")
+		container, err := ResolveCheckpointWorkerContainer(&podTemplate.Spec)
+		if err != nil {
+			return nil, err
 		}
-		if len(podTemplate.Spec.Containers[0].Command) == 0 {
+		if len(container.Command) == 0 {
 			return nil, fmt.Errorf("checkpoint job requires container.command when cuda-checkpoint launch-job wrapping is enabled")
 		}
-		podTemplate.Spec.Containers[0].Command, podTemplate.Spec.Containers[0].Args = wrapWithCudaCheckpointLaunchJob(
-			podTemplate.Spec.Containers[0].Command,
-			podTemplate.Spec.Containers[0].Args,
+		container.Command, container.Args = wrapWithCudaCheckpointLaunchJob(
+			container.Command,
+			container.Args,
 		)
 	}
 

deploy/snapshot/protocol/checkpoint_container.go

+// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+package protocol
+
+import (
+	"fmt"
+
+	corev1 "k8s.io/api/core/v1"
+)
+
+const checkpointWorkerContainerName = "main"
+
+func ResolveCheckpointWorkerContainer(podSpec *corev1.PodSpec) (*corev1.Container, error) {
+	if podSpec == nil || len(podSpec.Containers) == 0 {
+		return nil, fmt.Errorf("checkpoint job requires at least one container")
+	}
+	if len(podSpec.Containers) == 1 {
+		return &podSpec.Containers[0], nil
+	}
+
+	for i := range podSpec.Containers {
+		if podSpec.Containers[i].Name == checkpointWorkerContainerName {
+			return &podSpec.Containers[i], nil
+		}
+	}
+
+	return nil, fmt.Errorf("checkpoint job requires a container named %q when multiple containers are present", checkpointWorkerContainerName)
+}

deploy/snapshot/protocol/checkpoint_job_test.go

@@ -11,6 +11,17 @@ import (
 	"k8s.io/utils/ptr"
 )
 
+func requireCheckpointContainer(t *testing.T, containers []corev1.Container, name string) *corev1.Container {
+	t.Helper()
+	for i := range containers {
+		if containers[i].Name == name {
+			return &containers[i]
+		}
+	}
+	t.Fatalf("container %q not found", name)
+	return nil
+}
+
 func TestNewCheckpointJob(t *testing.T) {
 	job, err := NewCheckpointJob(&corev1.PodTemplateSpec{
 		ObjectMeta: metav1.ObjectMeta{
@@ -87,6 +98,97 @@ func TestNewCheckpointJob(t *testing.T) {
 	}
 }
 
+func TestNewCheckpointJobPrefersContainerNamedMain(t *testing.T) {
+	job, err := NewCheckpointJob(&corev1.PodTemplateSpec{
+		Spec: corev1.PodSpec{
+			Containers: []corev1.Container{
+				{Name: "sidecar", Command: []string{"sleep"}, Args: []string{"infinity"}},
+				{Name: "main", Command: []string{"python3", "-m", "dynamo.vllm"}, Args: []string{"--model", "Qwen"}},
+			},
+		},
+	}, CheckpointJobOptions{
+		Namespace:             "test-ns",
+		CheckpointID:          "hash",
+		ArtifactVersion:       "2",
+		Name:                  "test-job",
+		TTLSecondsAfterFinish: ptr.To(int32(300)),
+		WrapLaunchJob:         true,
+	})
+	if err != nil {
+		t.Fatalf("expected checkpoint job, got error: %v", err)
+	}
+
+	main := requireCheckpointContainer(t, job.Spec.Template.Spec.Containers, "main")
+	if len(main.Command) != 1 || main.Command[0] != "cuda-checkpoint" {
+		t.Fatalf("expected main container to be wrapped, got %#v", main.Command)
+	}
+	expectedArgs := []string{"--launch-job", "python3", "-m", "dynamo.vllm", "--model", "Qwen"}
+	if len(main.Args) != len(expectedArgs) {
+		t.Fatalf("expected launch-job args %#v, got %#v", expectedArgs, main.Args)
+	}
+	for i := range expectedArgs {
+		if main.Args[i] != expectedArgs[i] {
+			t.Fatalf("expected launch-job args %#v, got %#v", expectedArgs, main.Args)
+		}
+	}
+
+	sidecar := requireCheckpointContainer(t, job.Spec.Template.Spec.Containers, "sidecar")
+	if len(sidecar.Command) != 1 || sidecar.Command[0] != "sleep" {
+		t.Fatalf("expected sidecar command to remain unchanged, got %#v", sidecar.Command)
+	}
+	if len(sidecar.Args) != 1 || sidecar.Args[0] != "infinity" {
+		t.Fatalf("expected sidecar args to remain unchanged, got %#v", sidecar.Args)
+	}
+}
+
+func TestNewCheckpointJobAllowsSingleNonMainContainer(t *testing.T) {
+	job, err := NewCheckpointJob(&corev1.PodTemplateSpec{
+		Spec: corev1.PodSpec{
+			Containers: []corev1.Container{{
+				Name:    "worker",
+				Command: []string{"python3", "-m", "dynamo.vllm"},
+				Args:    []string{"--model", "Qwen"},
+			}},
+		},
+	}, CheckpointJobOptions{
+		Namespace:             "test-ns",
+		CheckpointID:          "hash",
+		ArtifactVersion:       "2",
+		Name:                  "test-job",
+		TTLSecondsAfterFinish: ptr.To(int32(300)),
+		WrapLaunchJob:         true,
+	})
+	if err != nil {
+		t.Fatalf("expected checkpoint job, got error: %v", err)
+	}
+
+	container := &job.Spec.Template.Spec.Containers[0]
+	if len(container.Command) != 1 || container.Command[0] != "cuda-checkpoint" {
+		t.Fatalf("expected single container to be wrapped, got %#v", container.Command)
+	}
+}
+
+func TestNewCheckpointJobRejectsMultiContainerPodWithoutMain(t *testing.T) {
+	_, err := NewCheckpointJob(&corev1.PodTemplateSpec{
+		Spec: corev1.PodSpec{
+			Containers: []corev1.Container{
+				{Name: "sidecar", Command: []string{"sleep"}, Args: []string{"infinity"}},
+				{Name: "worker", Command: []string{"python3", "-m", "dynamo.vllm"}},
+			},
+		},
+	}, CheckpointJobOptions{
+		Namespace:             "test-ns",
+		CheckpointID:          "hash",
+		ArtifactVersion:       "2",
+		Name:                  "test-job",
+		TTLSecondsAfterFinish: ptr.To(int32(300)),
+		WrapLaunchJob:         true,
+	})
+	if err == nil || err.Error() != "checkpoint job requires a container named \"main\" when multiple containers are present" {
+		t.Fatalf("expected missing main container error, got %v", err)
+	}
+}
+
 func TestGetCheckpointJobName(t *testing.T) {
 	name := GetCheckpointJobName("abc123def4567890", "2")
 	if name != "checkpoint-job-abc123def4567890-2" {

deploy/snapshot/protocol/restore.go

@@ -39,12 +39,31 @@ func NewRestorePod(pod *corev1.Pod, opts PodOptions) *corev1.Pod {
 		pod.Annotations = map[string]string{}
 	}
 	ApplyRestoreTargetMetadata(pod.Labels, pod.Annotations, true, opts.CheckpointID, opts.ArtifactVersion)
-	PrepareRestorePodSpec(&pod.Spec, &pod.Spec.Containers[0], opts.Storage, opts.SeccompProfile, true)
+	container := resolveWorkerContainer(&pod.Spec)
+	if container == nil {
+		return nil
+	}
+	PrepareRestorePodSpec(&pod.Spec, container, opts.Storage, opts.SeccompProfile, true)
 	pod.Namespace = opts.Namespace
 	pod.Spec.RestartPolicy = corev1.RestartPolicyNever
 	return pod
 }
 
+func resolveWorkerContainer(podSpec *corev1.PodSpec) *corev1.Container {
+	if podSpec == nil {
+		return nil
+	}
+	if len(podSpec.Containers) == 1 {
+		return &podSpec.Containers[0]
+	}
+	for index := range podSpec.Containers {
+		if podSpec.Containers[index].Name == "main" {
+			return &podSpec.Containers[index]
+		}
+	}
+	return nil
+}
+
 func PrepareRestorePodSpec(
 	podSpec *corev1.PodSpec,
 	container *corev1.Container,
@@ -92,10 +111,10 @@ func ValidateRestorePodSpec(
 	if podSpec == nil {
 		return fmt.Errorf("pod spec is nil")
 	}
-	if len(podSpec.Containers) != 1 {
-		return fmt.Errorf("restore target must have exactly one container, got %d", len(podSpec.Containers))
+	container := resolveWorkerContainer(podSpec)
+	if container == nil {
+		return fmt.Errorf("restore target must include a worker container named main")
 	}
-	container := &podSpec.Containers[0]
 	if storage.PVCName != "" {
 		hasVolume := false
 		for _, volume := range podSpec.Volumes {

deploy/snapshot/protocol/restore_test.go

@@ -187,6 +187,38 @@ func TestPrepareRestorePodSpecSynthesizesStartupProbeFromLiveness(t *testing.T)
 	}
 }
 
+func TestNewRestorePodTargetsMainContainerWhenSidecarsPresent(t *testing.T) {
+	restorePod := NewRestorePod(&corev1.Pod{
+		ObjectMeta: metav1.ObjectMeta{Name: "worker"},
+		Spec: corev1.PodSpec{
+			Containers: []corev1.Container{
+				{Name: "sidecar", Image: "sidecar:latest", Command: []string{"sidecar"}, Args: []string{"run"}},
+				{Name: "main", Image: "test:latest", Command: []string{"python3"}, Args: []string{"-m", "dynamo.vllm"}},
+			},
+		},
+	}, PodOptions{
+		Namespace:       "test-ns",
+		CheckpointID:    "hash",
+		ArtifactVersion: "2",
+		Storage: Storage{
+			Type:     StorageTypePVC,
+			PVCName:  "snapshot-pvc",
+			BasePath: "/checkpoints",
+		},
+		SeccompProfile: DefaultSeccompLocalhostProfile,
+	})
+
+	if got := restorePod.Spec.Containers[0].Command; len(got) != 1 || got[0] != "sidecar" {
+		t.Fatalf("expected sidecar command to remain unchanged, got %#v", got)
+	}
+	if got := restorePod.Spec.Containers[1].Command; len(got) != 2 || got[0] != "sleep" || got[1] != "infinity" {
+		t.Fatalf("expected main container placeholder command, got %#v", got)
+	}
+	if restorePod.Spec.Containers[1].Args != nil {
+		t.Fatalf("expected main container args to be cleared: %#v", restorePod.Spec.Containers[1].Args)
+	}
+}
+
 func TestPrepareRestorePodSpecSynthesizesStartupProbeFromReadiness(t *testing.T) {
 	podSpec := corev1.PodSpec{}
 	readinessProbe := &corev1.Probe{
@@ -279,7 +311,7 @@ func TestValidateRestorePodSpec(t *testing.T) {
 	}
 }
 
-func TestValidateRestorePodSpecRequiresExactlyOneContainer(t *testing.T) {
+func TestValidateRestorePodSpecRequiresMainContainerWhenMultiContainer(t *testing.T) {
 	profile := DefaultSeccompLocalhostProfile
 	podSpec := &corev1.PodSpec{
 		SecurityContext: &corev1.PodSecurityContext{
@@ -314,8 +346,48 @@ func TestValidateRestorePodSpecRequiresExactlyOneContainer(t *testing.T) {
 		BasePath: "/checkpoints",
 	}
 
-	if err := ValidateRestorePodSpec(podSpec, storage, DefaultSeccompLocalhostProfile); err == nil || err.Error() != "restore target must have exactly one container, got 2" {
-		t.Fatalf("expected multi-container restore target to be rejected, got %v", err)
+	if err := ValidateRestorePodSpec(podSpec, storage, DefaultSeccompLocalhostProfile); err == nil || err.Error() != "restore target must include a worker container named main" {
+		t.Fatalf("expected multi-container restore target without main to be rejected, got %v", err)
+	}
+}
+
+func TestValidateRestorePodSpecAllowsMainContainerWithSidecars(t *testing.T) {
+	profile := DefaultSeccompLocalhostProfile
+	podSpec := &corev1.PodSpec{
+		SecurityContext: &corev1.PodSecurityContext{
+			SeccompProfile: &corev1.SeccompProfile{
+				Type:             corev1.SeccompProfileTypeLocalhost,
+				LocalhostProfile: &profile,
+			},
+		},
+		Volumes: []corev1.Volume{{
+			Name: CheckpointVolumeName,
+			VolumeSource: corev1.VolumeSource{
+				PersistentVolumeClaim: &corev1.PersistentVolumeClaimVolumeSource{
+					ClaimName: "snapshot-pvc",
+				},
+			},
+		}},
+		Containers: []corev1.Container{
+			{Name: "sidecar"},
+			{
+				Name: "main",
+				VolumeMounts: []corev1.VolumeMount{{
+					Name:      CheckpointVolumeName,
+					MountPath: "/checkpoints",
+				}},
+			},
+		},
+	}
+
+	storage := Storage{
+		Type:     StorageTypePVC,
+		PVCName:  "snapshot-pvc",
+		BasePath: "/checkpoints",
+	}
+
+	if err := ValidateRestorePodSpec(podSpec, storage, DefaultSeccompLocalhostProfile); err != nil {
+		t.Fatalf("expected main container with sidecars to validate, got %v", err)
 	}
 }
 

docs/kubernetes/api-reference.md

@@ -302,6 +302,7 @@ _Appears in:_
 | Field | Description | Default | Validation |
 | --- | --- | --- | --- |
 | `identity` _[DynamoCheckpointIdentity](#dynamocheckpointidentity)_ | Identity defines the inputs that determine checkpoint equivalence |  | Required: \{\} <br /> |
+| `gpuMemoryService` _[GPUMemoryServiceSpec](#gpumemoryservicespec)_ | GPUMemoryService enables checkpoint-time GPU Memory Service wiring.<br />It is intentionally outside spec.identity, so it does not affect the<br />checkpoint identity hash or deduplication. |  | Optional: \{\} <br /> |
 | `job` _[DynamoCheckpointJobConfig](#dynamocheckpointjobconfig)_ | Job defines the configuration for the checkpoint creation Job |  | Required: \{\} <br /> |
 
 
@@ -852,6 +853,7 @@ via DRA (Dynamic Resource Allocation). The sidecar runs two GMS processes per GP
 
 
 _Appears in:_
+- [DynamoCheckpointSpec](#dynamocheckpointspec)
 - [DynamoComponentDeploymentSharedSpec](#dynamocomponentdeploymentsharedspec)
 - [DynamoComponentDeploymentSpec](#dynamocomponentdeploymentspec)
 

docs/kubernetes/snapshot.md

@@ -142,6 +142,16 @@ spec:
             ...
 ```
 
+If this checkpoint should capture and restore GPU Memory Service helpers, set:
+
+```yaml
+spec:
+  gpuMemoryService:
+    enabled: true
+```
+
+`spec.gpuMemoryService` is outside `spec.identity`, so it does not change the checkpoint identity hash.
+
 For a full working example, see [deploy/operator/config/samples/nvidia.com_v1alpha1_dynamocheckpoint.yaml](https://github.com/ai-dynamo/dynamo/blob/main/deploy/operator/config/samples/nvidia.com_v1alpha1_dynamocheckpoint.yaml).
 
 Apply it:
@@ -262,6 +272,8 @@ spec:
         ...
 ```
 
+Auto mode only hashes `checkpoint.identity`. If you need GMS-specific checkpoint behavior, configure it on the `DynamoCheckpoint` object with `spec.gpuMemoryService.enabled`.
+
 Useful inspection commands:
 
 ```bash

lib/gpu_memory_service/cli/__init__.py

@@ -4,7 +4,13 @@
 """CLI for GPU Memory Service."""
 
 from gpu_memory_service.cli.args import Config, parse_args
-from gpu_memory_service.cli.runner import main
+
+
+def main():
+    from gpu_memory_service.cli.runner import main as runner_main
+
+    return runner_main()
+
 
 __all__ = [
     "Config",

lib/gpu_memory_service/cli/server.py

+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""GMS server entry point.
+
+Launches two GMS server processes per GPU (one for weights, one for kv_cache).
+Writes a ready file once all expected UDS sockets are present. Monitors an
+optional checkpoint stop file and shuts down cleanly when it appears.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import signal
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+from gpu_memory_service.common.utils import get_socket_path
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
+)
+logger = logging.getLogger(__name__)
+
+_TAGS = ("weights", "kv_cache")
+_READY_FILE = "gms-ready"
+
+
+def _ready_file_path() -> Path:
+    return Path(os.environ.get("GMS_SOCKET_DIR", "/tmp")) / _READY_FILE
+
+
+def _list_devices() -> list[int]:
+    import pynvml
+
+    pynvml.nvmlInit()
+    try:
+        count = pynvml.nvmlDeviceGetCount()
+    finally:
+        pynvml.nvmlShutdown()
+
+    if count == 0:
+        raise SystemExit("no nvidia devices found")
+    return list(range(count))
+
+
+def _optional_checkpoint_stop_file() -> Path | None:
+    control_dir = os.environ.get("GMS_CONTROL_DIR")
+    if not control_dir:
+        return None
+    return Path(control_dir) / "checkpoint-done"
+
+
+def main() -> None:
+    ready_file = _ready_file_path()
+    ready_file.unlink(missing_ok=True)
+
+    devices = _list_devices()
+    processes = []
+    for device in devices:
+        for tag in _TAGS:
+            proc = subprocess.Popen(
+                [
+                    sys.executable,
+                    "-m",
+                    "gpu_memory_service",
+                    "--device",
+                    str(device),
+                    "--tag",
+                    tag,
+                ]
+            )
+            logger.info("Started GMS device=%d tag=%s pid=%d", device, tag, proc.pid)
+            processes.append(proc)
+
+    def shutdown() -> None:
+        for process in processes:
+            if process.poll() is None:
+                process.terminate()
+
+    def terminate(*_args) -> None:
+        shutdown()
+        raise SystemExit(0)
+
+    signal.signal(signal.SIGTERM, terminate)
+    signal.signal(signal.SIGINT, terminate)
+
+    stop_file = _optional_checkpoint_stop_file()
+    ready_written = False
+    while True:
+        stop_requested = stop_file is not None and stop_file.exists()
+        if stop_requested:
+            logger.info("checkpoint stop requested; shutting down GMS servers")
+            shutdown()
+
+        if not ready_written:
+            sockets_ready = all(
+                os.path.exists(get_socket_path(device, tag))
+                for device in devices
+                for tag in _TAGS
+            )
+            if sockets_ready:
+                ready_file.write_text("ready", encoding="utf-8")
+                ready_written = True
+
+        running = False
+        for process in processes:
+            exit_code = process.poll()
+            if exit_code is None:
+                running = True
+                continue
+            if stop_requested:
+                continue
+            shutdown()
+            raise SystemExit(exit_code)
+
+        if not running:
+            return
+        time.sleep(1)
+
+
+if __name__ == "__main__":
+    main()

lib/gpu_memory_service/cli/snapshot/__init__.py

+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0

lib/gpu_memory_service/cli/snapshot/loader.py

+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""GMS checkpoint loader entry point.
+
+Waits for the GMS server UDS socket on each device, then loads saved GMS
+state from a checkpoint directory into the running GMS servers. Devices
+are loaded in parallel to saturate PVC bandwidth.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+from gpu_memory_service.common.utils import get_socket_path
+from gpu_memory_service.snapshot.storage_client import GMSStorageClient
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
+)
+logger = logging.getLogger(__name__)
+
+_WEIGHTS_TAG = "weights"
+_DEFAULT_MAX_WORKERS = 8
+
+
+def _list_devices() -> list[int]:
+    import pynvml
+
+    pynvml.nvmlInit()
+    try:
+        count = pynvml.nvmlDeviceGetCount()
+    finally:
+        pynvml.nvmlShutdown()
+
+    if count == 0:
+        raise SystemExit("no nvidia devices found")
+    return list(range(count))
+
+
+def _wait_for_weights_socket(device: int) -> None:
+    socket_path = get_socket_path(device, _WEIGHTS_TAG)
+    while not os.path.exists(socket_path):
+        time.sleep(1)
+
+
+def _load_device(checkpoint_dir: str, device: int, max_workers: int) -> None:
+    _wait_for_weights_socket(device)
+    input_dir = os.path.join(checkpoint_dir, f"device-{device}")
+    logger.info("Loading GMS checkpoint: device=%d input_dir=%s", device, input_dir)
+    t0 = time.monotonic()
+    client = GMSStorageClient(
+        socket_path=get_socket_path(device),
+        device=device,
+    )
+    client.load_to_gms(
+        input_dir,
+        max_workers=max_workers,
+        clear_existing=True,
+    )
+    elapsed = time.monotonic() - t0
+    logger.info("GMS checkpoint loaded: device=%d elapsed=%.2fs", device, elapsed)
+
+
+def main() -> None:
+    checkpoint_dir = os.environ["GMS_CHECKPOINT_DIR"]
+    max_workers = int(os.environ.get("GMS_LOAD_WORKERS", str(_DEFAULT_MAX_WORKERS)))
+    devices = _list_devices()
+
+    t0 = time.monotonic()
+    with ThreadPoolExecutor(max_workers=len(devices)) as pool:
+        futures = {
+            pool.submit(_load_device, checkpoint_dir, dev, max_workers): dev
+            for dev in devices
+        }
+        for future in as_completed(futures):
+            dev = futures[future]
+            future.result()
+            logger.info("Device %d load complete", dev)
+    elapsed = time.monotonic() - t0
+    logger.info("All %d devices loaded in %.2fs", len(devices), elapsed)
+
+    while True:
+        time.sleep(3600)
+
+
+if __name__ == "__main__":
+    main()

lib/gpu_memory_service/cli/snapshot/saver.py

+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""GMS checkpoint saver entry point.
+
+Waits for the checkpoint pod to reach Ready=True, then saves GMS state from
+each device in parallel. Writes a stop file to signal the GMS server to shut
+down after save completes.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import ssl
+import time
+import urllib.error
+import urllib.request
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+from typing import Any
+
+from gpu_memory_service.common.utils import get_socket_path
+from gpu_memory_service.snapshot.storage_client import GMSStorageClient
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
+)
+logger = logging.getLogger(__name__)
+
+_WEIGHTS_TAG = "weights"
+_SERVICE_ACCOUNT_TOKEN = Path("/var/run/secrets/kubernetes.io/serviceaccount/token")
+_SERVICE_ACCOUNT_CA = "/var/run/secrets/kubernetes.io/serviceaccount/ca.crt"
+
+
+def _list_devices() -> list[int]:
+    import pynvml
+
+    pynvml.nvmlInit()
+    try:
+        count = pynvml.nvmlDeviceGetCount()
+    finally:
+        pynvml.nvmlShutdown()
+
+    if count == 0:
+        raise SystemExit("no nvidia devices found")
+    return list(range(count))
+
+
+def _wait_for_weights_socket(device: int) -> None:
+    socket_path = get_socket_path(device, _WEIGHTS_TAG)
+    while not os.path.exists(socket_path):
+        time.sleep(1)
+
+
+def _checkpoint_pod_ready(pod: dict[str, Any]) -> bool:
+    status = pod.get("status") or {}
+    if str(status.get("phase", "")).strip() != "Running":
+        return False
+    for condition in status.get("conditions") or []:
+        if (
+            condition.get("type") == "Ready"
+            and str(condition.get("status", "")).strip() == "True"
+        ):
+            return True
+    return False
+
+
+def _main_terminated(pod: dict[str, Any]) -> bool:
+    status = pod.get("status") or {}
+    for container in status.get("containerStatuses") or []:
+        if container.get("name") != "main":
+            continue
+        return bool((container.get("state") or {}).get("terminated"))
+    return False
+
+
+def main() -> None:
+    service_token = _SERVICE_ACCOUNT_TOKEN.read_text(encoding="utf-8").strip()
+    ssl_context = ssl.create_default_context(cafile=_SERVICE_ACCOUNT_CA)
+    pod_api_url = (
+        "https://"
+        + os.environ["KUBERNETES_SERVICE_HOST"]
+        + ":"
+        + os.environ.get("KUBERNETES_SERVICE_PORT_HTTPS", "443")
+        + f"/api/v1/namespaces/{os.environ['POD_NAMESPACE']}/pods/{os.environ['POD_NAME']}"
+    )
+    checkpoint_dir = os.environ["GMS_CHECKPOINT_DIR"]
+
+    def checkpoint_pod() -> dict[str, Any]:
+        request = urllib.request.Request(
+            pod_api_url,
+            headers={"Authorization": f"Bearer {service_token}"},
+        )
+        with urllib.request.urlopen(
+            request,
+            context=ssl_context,
+            timeout=5,
+        ) as response:
+            return json.load(response)
+
+    logger.info("Waiting for checkpoint pod Ready=True before GMS save")
+    while True:
+        try:
+            pod = checkpoint_pod()
+        except (urllib.error.URLError, TimeoutError, ssl.SSLError, OSError):
+            time.sleep(1)
+            continue
+
+        if _checkpoint_pod_ready(pod):
+            break
+        if _main_terminated(pod):
+            raise SystemExit("main container terminated before GMS save could start")
+        time.sleep(1)
+
+    def _save_device(device: int, max_workers: int) -> None:
+        _wait_for_weights_socket(device)
+        output_dir = os.path.join(checkpoint_dir, f"device-{device}")
+        logger.info(
+            "Saving GMS checkpoint: device=%d output_dir=%s",
+            device,
+            output_dir,
+        )
+        t0 = time.monotonic()
+        client = GMSStorageClient(
+            output_dir,
+            socket_path=get_socket_path(device),
+            device=device,
+        )
+        client.save(max_workers=max_workers)
+        elapsed = time.monotonic() - t0
+        logger.info("GMS checkpoint saved: device=%d elapsed=%.2fs", device, elapsed)
+
+    max_workers = int(os.environ.get("GMS_SAVE_WORKERS", "8"))
+    logger.info("Checkpoint pod is Ready; starting GMS save")
+    try:
+        devices = _list_devices()
+        t0 = time.monotonic()
+        with ThreadPoolExecutor(max_workers=len(devices)) as pool:
+            futures = {
+                pool.submit(_save_device, dev, max_workers): dev for dev in devices
+            }
+            for future in as_completed(futures):
+                future.result()
+        elapsed = time.monotonic() - t0
+        logger.info("All %d devices saved in %.2fs", len(devices), elapsed)
+    finally:
+        (Path(os.environ["GMS_CONTROL_DIR"]) / "checkpoint-done").write_text(
+            "done",
+            encoding="utf-8",
+        )
+
+
+if __name__ == "__main__":
+    main()

lib/gpu_memory_service/cli/storage_runner.py

+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""GMS Storage Client CLI.
+
+Provides two subcommands for saving and loading GPU Memory Service state:
+
+* ``save`` – connect to a running GMS server in RO mode and write every
+              allocation plus all metadata to a sharded binary directory.
+* ``load`` – connect to a running GMS server in RW mode, read tensor data
+              from a saved directory, and commit the state so readers can
+              acquire the RO lock.
+
+Usage examples::
+
+    # Save GMS state to disk
+    gms-storage-client save --output-dir /mnt/nvme/save --device 0
+
+    # Load a previous save back into a fresh GMS server
+    gms-storage-client load --input-dir /mnt/nvme/save --device 0
+"""
+
+import argparse
+import logging
+import sys
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
+)
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Shared helpers
+# ---------------------------------------------------------------------------
+
+
+def _configure_logging(verbose: bool) -> None:
+    if verbose:
+        logging.getLogger().setLevel(logging.DEBUG)
+        logging.getLogger("gpu_memory_service").setLevel(logging.DEBUG)
+
+
+def _resolve_socket(device: int, socket_path) -> str:
+    if socket_path is not None:
+        return socket_path
+    from gpu_memory_service.common.utils import get_socket_path
+
+    return get_socket_path(device)
+
+
+# ---------------------------------------------------------------------------
+# Subcommand implementations
+# ---------------------------------------------------------------------------
+
+
+def _run_save(args) -> None:
+    """Execute the save subcommand."""
+    from gpu_memory_service.snapshot.storage_client import GMSStorageClient
+
+    _configure_logging(args.verbose)
+    socket_path = _resolve_socket(args.device, args.socket_path)
+
+    logger.info(
+        "Saving GMS state: device=%s, socket=%s, output_dir=%s, save_workers=%s",
+        args.device,
+        socket_path,
+        args.output_dir,
+        args.save_workers,
+    )
+
+    client = GMSStorageClient(
+        args.output_dir,
+        socket_path=socket_path,
+        device=args.device,
+        timeout_ms=args.timeout_ms,
+        shard_size_bytes=args.shard_size_bytes,
+    )
+
+    manifest = client.save(max_workers=args.save_workers)
+    shard_count = len({a.tensor_file for a in manifest.allocations})
+
+    logger.info(
+        "Save complete: %s allocations written to %s (%s shards)",
+        len(manifest.allocations),
+        args.output_dir,
+        shard_count,
+    )
+    logger.info("Layout hash: %s", manifest.layout_hash)
+
+
+def _run_load(args) -> None:
+    """Execute the load subcommand."""
+    from gpu_memory_service.snapshot.storage_client import GMSStorageClient
+
+    _configure_logging(args.verbose)
+    socket_path = _resolve_socket(args.device, args.socket_path)
+
+    logger.info(
+        "Loading GMS state: device=%s, socket=%s, input_dir=%s, clear_existing=%s",
+        args.device,
+        socket_path,
+        args.input_dir,
+        not args.no_clear,
+    )
+
+    client = GMSStorageClient(
+        socket_path=socket_path,
+        device=args.device,
+        timeout_ms=args.timeout_ms,
+    )
+
+    id_map = client.load_to_gms(
+        args.input_dir,
+        max_workers=args.workers,
+        clear_existing=not args.no_clear,
+    )
+
+    logger.info("Load complete: %s allocations committed to GMS", len(id_map))
+    for old_id, new_id in id_map.items():
+        logger.info("  %s → %s", old_id, new_id)
+
+
+# ---------------------------------------------------------------------------
+# Argument parsing
+# ---------------------------------------------------------------------------
+
+_SHARD_SIZE_DEFAULT = 4 * 1024**3  # 4 GiB
+
+
+def _build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        prog="gms-storage-client",
+        description="Save and load GPU Memory Service state to/from disk.",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    subparsers = parser.add_subparsers(dest="subcommand")
+
+    # -- save ---------------------------------------------------------------
+    save_p = subparsers.add_parser(
+        "save",
+        help="Save GMS state to a sharded binary directory.",
+        description=(
+            "Connect to a running GMS server in RO mode and export every "
+            "allocation plus all metadata to a compact sharded binary format."
+        ),
+    )
+    save_p.add_argument(
+        "--output-dir",
+        required=True,
+        help="Directory to write into (created if absent).",
+    )
+    save_p.add_argument(
+        "--device",
+        type=int,
+        default=0,
+        help="CUDA device index (default: 0).",
+    )
+    save_p.add_argument(
+        "--socket-path",
+        type=str,
+        default=None,
+        help="GMS Unix socket path. Default uses GPU UUID-based path.",
+    )
+    save_p.add_argument(
+        "--timeout-ms",
+        type=int,
+        default=None,
+        help="Timeout in milliseconds for acquiring the RO lock.",
+    )
+    save_p.add_argument(
+        "--shard-size-bytes",
+        type=int,
+        default=_SHARD_SIZE_DEFAULT,
+        help=(
+            f"Soft upper bound per shard file in bytes "
+            f"(default: {_SHARD_SIZE_DEFAULT // 1024**3} GiB).  "
+            "Decrease to increase parallelism on save/load; increase to "
+            "reduce file count."
+        ),
+    )
+    save_p.add_argument(
+        "--save-workers",
+        type=int,
+        default=8,
+        help="Thread pool size for parallel shard writes (default: 8).",
+    )
+    save_p.add_argument(
+        "--verbose",
+        "-v",
+        action="store_true",
+        help="Enable verbose logging.",
+    )
+
+    # -- load ---------------------------------------------------------------
+    load_p = subparsers.add_parser(
+        "load",
+        help="Load a saved GMS state back into a running GMS server.",
+        description=(
+            "Connect to a running GMS server in RW mode, read tensor data "
+            "from a saved directory (reading each shard file sequentially), "
+            "and commit the state so readers can acquire the RO lock."
+        ),
+    )
+    load_p.add_argument(
+        "--input-dir",
+        required=True,
+        help="Directory previously created by the save subcommand.",
+    )
+    load_p.add_argument(
+        "--device",
+        type=int,
+        default=0,
+        help="CUDA device index (default: 0).",
+    )
+    load_p.add_argument(
+        "--socket-path",
+        type=str,
+        default=None,
+        help="GMS Unix socket path. Default uses GPU UUID-based path.",
+    )
+    load_p.add_argument(
+        "--timeout-ms",
+        type=int,
+        default=None,
+        help="Timeout in milliseconds for acquiring the RW lock.",
+    )
+    load_p.add_argument(
+        "--workers",
+        type=int,
+        default=8,
+        help="Thread pool size for parallel shard reads (default: 8).",
+    )
+    load_p.add_argument(
+        "--no-clear",
+        action="store_true",
+        default=False,
+        help=(
+            "Do not clear existing GMS allocations before loading. "
+            "Default behaviour clears the server to produce an exact replica."
+        ),
+    )
+    load_p.add_argument(
+        "--verbose",
+        "-v",
+        action="store_true",
+        help="Enable verbose logging.",
+    )
+
+    return parser
+
+
+# ---------------------------------------------------------------------------
+# Entry point
+# ---------------------------------------------------------------------------
+
+
+def main() -> None:
+    """Entry point for the GMS Storage Client CLI."""
+    parser = _build_parser()
+    args = parser.parse_args()
+
+    if args.subcommand is None:
+        parser.print_help()
+        sys.exit(1)
+
+    if args.subcommand == "save":
+        _run_save(args)
+    elif args.subcommand == "load":
+        _run_load(args)
+    else:
+        parser.print_help()
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

lib/gpu_memory_service/client/memory_manager.py

@@ -37,7 +37,7 @@ from typing import Dict, List, Optional
 from gpu_memory_service.client.session import _GMSClientSession
 from gpu_memory_service.common.cuda_utils import (
     align_to_granularity,
-    cuda_set_current_device,
+    cuda_ensure_initialized,
     cuda_synchronize,
     cuda_validate_pointer,
     cumem_address_free,
@@ -134,22 +134,25 @@ class GMSClientMemoryManager:
         socket_path: str,
         *,
         device: int = 0,
+        tag: Optional[str] = None,
     ) -> None:
         self.socket_path = socket_path
         self.device = device
+        self.tag = tag
 
         self._client: Optional[_GMSClientSession] = None
         self._mappings: Dict[int, LocalMapping] = {}  # va -> mapping
         self._inverse_mapping: Dict[str, int] = {}
 
         self._unmapped = False
+        self._aborted = False
         self._granted_lock_type: Optional[GrantedLockType] = None
 
         # VA-stable unmap/remap state
         self._va_preserved = False
         self._last_memory_layout_hash: str = ""
 
-        cuda_set_current_device(self.device)
+        cuda_ensure_initialized()
         self.granularity = cumem_get_allocation_granularity(device)
 
     # ==================== Properties ====================
@@ -183,9 +186,34 @@ class GMSClientMemoryManager:
 
         Updates self._granted_lock_type based on granted lock type. Saves memory layout hash
         for stale detection if server is in committed state.
+
+        On reconnect after abort (e.g. after CRIU restore on a different GPU),
+        refreshes the socket path from the current GPU UUID so we connect to
+        the correct GMS server.
         """
         if self._client is not None:
             raise RuntimeError("Memory manager is already connected")
+
+        # After abort + CRIU restore the process may be on a different GPU.
+        # Re-derive socket path from current UUID so we talk to the right server.
+        if self._aborted and self.tag is not None:
+            from gpu_memory_service.common.utils import (
+                get_socket_path,
+                invalidate_uuid_cache,
+            )
+
+            invalidate_uuid_cache()
+            new_path = get_socket_path(self.device, self.tag)
+            if new_path != self.socket_path:
+                logger.info(
+                    "Refreshed socket path for tag=%s: %s -> %s",
+                    self.tag,
+                    self.socket_path,
+                    new_path,
+                )
+                self.socket_path = new_path
+            self._aborted = False
+
         self._client = _GMSClientSession(
             self.socket_path,
             lock_type=lock_type,
@@ -211,6 +239,7 @@ class GMSClientMemoryManager:
         Clean callers should unmap first. This also supports abrupt session
         drop with live mappings still present.
         """
+        self._aborted = True
         if self._client is not None:
             try:
                 self._client.close()
@@ -464,8 +493,6 @@ class GMSClientMemoryManager:
         Checks layout hash for staleness. Validates each allocation still
         exists and size matches before remapping.
         """
-        cuda_set_current_device(self.device)
-
         # Stale layout check
         current_hash = self.get_memory_layout_hash()
         if (
@@ -580,18 +607,30 @@ class GMSClientMemoryManager:
 
     # ==================== Lifecycle ====================
 
-    def close(self) -> None:
-        """Strict cleanup.
+    def close(self, *, best_effort: bool = False) -> None:
+        """Cleanup mappings and abort.
 
         synchronize + unmap all + free all VAs + abort.
-        """
-        cuda_synchronize()
-
-        for va in list(self._mappings.keys()):
-            self.unmap_va(va)
-            self.free_va(va)
 
-        self.abort()
+        Args:
+            best_effort: If True, skip cuda_synchronize and swallow
+                errors during cleanup. Used after checkpoint where
+                cuda-checkpoint may have torn down the device context
+                (cuda_synchronize calls os._exit via fail()).
+        """
+        if best_effort:
+            try:
+                self.abort()
+            except Exception:
+                pass
+            self._mappings.clear()
+            self._inverse_mapping.clear()
+        else:
+            cuda_synchronize()
+            for va in list(self._mappings.keys()):
+                self.unmap_va(va)
+                self.free_va(va)
+            self.abort()
         self._unmapped = False
         self._va_preserved = False
         from gpu_memory_service.client.torch.allocator import (

lib/gpu_memory_service/client/torch/allocator.py

@@ -127,7 +127,7 @@ def get_or_create_gms_client_memory_manager(
             )
         return state.manager
 
-    manager = GMSClientMemoryManager(socket_path, device=device)
+    manager = GMSClientMemoryManager(socket_path, device=device, tag=tag)
     manager.connect(mode, timeout_ms=timeout_ms)
 
     mem_pool = None

lib/gpu_memory_service/common/cuda_utils.py

@@ -5,7 +5,6 @@
 
 from __future__ import annotations
 
-import atexit
 import os
 
 from gpu_memory_service.common.locks import GrantedLockType
@@ -24,9 +23,6 @@ except ImportError:
 
     cuda = _MissingCuda()
 
-_primary_contexts: dict[int, object] = {}
-_primary_context_release_registered = False
-
 
 def cuda_check_result(result: cuda.CUresult, name: str) -> None:
     if result != cuda.CUresult.CUDA_SUCCESS:
@@ -169,28 +165,3 @@ def cuda_validate_pointer(va: int) -> None:
 def cuda_synchronize() -> None:
     (result,) = cuda.cuCtxSynchronize()
     cuda_check_result(result, "cuCtxSynchronize")
-
-
-def cuda_set_current_device(device: int) -> None:
-    global _primary_context_release_registered
-
-    ctx = _primary_contexts.get(device)
-    if ctx is None:
-        result, ctx = cuda.cuDevicePrimaryCtxRetain(device)
-        cuda_check_result(result, "cuDevicePrimaryCtxRetain")
-        _primary_contexts[device] = ctx
-        if not _primary_context_release_registered:
-            _primary_context_release_registered = True
-            atexit.register(_release_primary_contexts)
-    (result,) = cuda.cuCtxSetCurrent(ctx)
-    cuda_check_result(result, "cuCtxSetCurrent")
-
-
-def _release_primary_contexts() -> None:
-    for device in list(_primary_contexts):
-        try:
-            (result,) = cuda.cuDevicePrimaryCtxRelease(device)
-        except Exception:
-            continue
-        if result == cuda.CUresult.CUDA_SUCCESS:
-            _primary_contexts.pop(device, None)

lib/gpu_memory_service/common/utils.py

@@ -17,11 +17,19 @@ def fail(message: str, *args, exc_info=None) -> NoReturn:
     os._exit(1)
 
 
+_uuid_cache: dict[int, str] = {}
+
+
+def invalidate_uuid_cache() -> None:
+    """Clear cached GPU UUIDs. Call after CRIU restore when GPU assignment may change."""
+    _uuid_cache.clear()
+
+
 def get_socket_path(device: int, tag: str = "weights") -> str:
     """Get GMS socket path for the given CUDA device and tag.
 
     The socket path is based on GPU UUID, making it stable across different
-    CUDA_VISIBLE_DEVICES configurations.
+    CUDA_VISIBLE_DEVICES configurations. UUIDs are cached per device index.
 
     Args:
         device: CUDA device index.
@@ -30,12 +38,16 @@ def get_socket_path(device: int, tag: str = "weights") -> str:
         Socket path
         (e.g., "<tempdir>/gms_GPU-12345678-1234-1234-1234-123456789abc_weights.sock").
     """
-    import pynvml
-
-    pynvml.nvmlInit()
-    try:
-        handle = pynvml.nvmlDeviceGetHandleByIndex(device)
-        uuid = pynvml.nvmlDeviceGetUUID(handle)
-    finally:
-        pynvml.nvmlShutdown()
-    return os.path.join(tempfile.gettempdir(), f"gms_{uuid}_{tag}.sock")
+    uuid = _uuid_cache.get(device)
+    if uuid is None:
+        import pynvml  # deferred: not available in all environments
+
+        pynvml.nvmlInit()
+        try:
+            handle = pynvml.nvmlDeviceGetHandleByIndex(device)
+            uuid = pynvml.nvmlDeviceGetUUID(handle)
+        finally:
+            pynvml.nvmlShutdown()
+        _uuid_cache[device] = uuid
+    socket_dir = os.environ.get("GMS_SOCKET_DIR") or tempfile.gettempdir()
+    return os.path.join(socket_dir, f"gms_{uuid}_{tag}.sock")