fix(nnctl): allow CPU cloud deploys

Allow Verda cloud deploy to accept CPU-only spot instance types.

Keep the multi-GPU guard in place so large GPU shapes remain blocked.

This lets e2e workers use CPU capacity when GPUs are not required.

Signed-off-by: Ville Vesilehto <ville@vesilehto.fi>

Author: thevilledev

nnctl/internal/cli/cloud.go

@@ -309,7 +309,7 @@ func (a *App) printCloudDeployResult(result *verdacloud.DeployResult, jsonOutput
 		if location == "" {
 			location = result.Policy.LocationSelection
 		}
-		fmt.Fprintf(a.stdout(), "dry run: would deploy %s as a spot single-GPU instance in %s\n", result.Request.InstanceType, location)
+		fmt.Fprintf(a.stdout(), "dry run: would deploy %s as a spot instance in %s\n", result.Request.InstanceType, location)
 		fmt.Fprintf(a.stdout(), "hostname: %s\n", result.Request.Hostname)
 		fmt.Fprintf(a.stdout(), "source os volume: %s\n", result.SourceOSVolumeID)
 		if result.Policy.SourceOSVolumeLocked {

nnctl/internal/cli/cloud_test.go

@@ -28,7 +28,8 @@ func TestCloudDeployDryRunJSON(t *testing.T) {
 			LocationSelection    string `json:"location_selection"`
 			SourceOSVolumeLocked bool   `json:"source_os_volume_locked"`
 			SpotOnly             bool   `json:"spot_only"`
-			SingleGPU            bool   `json:"single_gpu"`
+			AllowsCPU            bool   `json:"allows_cpu"`
+			MaxGPUCount          int    `json:"max_gpu_count"`
 		} `json:"policy"`
 		SourceOSVolumeID string `json:"source_os_volume_id"`
 		Request          struct {
@@ -51,7 +52,7 @@ func TestCloudDeployDryRunJSON(t *testing.T) {
 	if result.SourceOSVolumeID != "vol-golden" {
 		t.Fatalf("SourceOSVolumeID = %q", result.SourceOSVolumeID)
 	}
-	if result.Policy.LocationCode != "" || result.Policy.LocationSelection != "source_os_volume_location" || !result.Policy.SourceOSVolumeLocked || !result.Policy.SpotOnly || !result.Policy.SingleGPU {
+	if result.Policy.LocationCode != "" || result.Policy.LocationSelection != "source_os_volume_location" || !result.Policy.SourceOSVolumeLocked || !result.Policy.SpotOnly || !result.Policy.AllowsCPU || result.Policy.MaxGPUCount != 1 {
 		t.Fatalf("policy was not encoded: %#v", result.Policy)
 	}
 }

nnctl/internal/cli/cobra.go

@@ -234,8 +234,8 @@ changes, so remote benchmark runs can be tied back to a specific git ref.`,
 func (a *App) newCloudCommand() *cobra.Command {
 	cmd := &cobra.Command{
 		Use:   "cloud <command>",
-		Short: "Deploy cloud GPU benchmark workers",
-		Long:  "Deploys cloud GPU benchmark workers for nnctl agents. Verda deployments are restricted to spot, single-GPU instances.",
+		Short: "Deploy cloud benchmark workers",
+		Long:  "Deploys cloud benchmark workers for nnctl agents. Verda deployments are restricted to spot CPU-only or single-GPU instances.",
 		RunE: func(cmd *cobra.Command, args []string) error {
 			return cmd.Help()
 		},
@@ -257,10 +257,10 @@ func (a *App) newCloudDeployCommand() *cobra.Command {
 	}
 	cmd := &cobra.Command{
 		Use:   "deploy",
-		Short: "Deploy a Verda spot GPU worker",
-		Long: `Deploys one Verda GPU instance for nnctl benchmark work.
-The deployment policy is intentionally narrow: spot only and single-GPU instance
-types only. By default nnctl chooses the cheapest currently available spot
+		Short: "Deploy a Verda spot benchmark worker",
+		Long: `Deploys one Verda spot instance for nnctl benchmark work.
+The deployment policy is intentionally narrow: spot only, with CPU-only and
+single-GPU instance types accepted. By default nnctl uses the source OS volume
 location for the requested instance type. Userdata defaults to the embedded
 script from nnctl/internal/cloud/verda/bootstrap.sh.`,
 		Example: `  nnctl cloud deploy --instance-type 1V100.6V --source-os-volume-id volume_id --ssh-key-id ssh_key_id
@@ -279,7 +279,7 @@ script from nnctl/internal/cloud/verda/bootstrap.sh.`,
 	cmd.Flags().StringVar(&opts.Hostname, "hostname", opts.Hostname, "instance hostname")
 	cmd.Flags().StringVar(&opts.Description, "description", opts.Description, "instance description")
 	cmd.Flags().StringArrayVar(&opts.SSHKeyIDs, "ssh-key-id", opts.SSHKeyIDs, "Verda SSH key ID to attach (repeatable)")
-	cmd.Flags().StringVar(&opts.LocationCode, "location-code", opts.LocationCode, "Verda location code; defaults to cheapest currently available spot location")
+	cmd.Flags().StringVar(&opts.LocationCode, "location-code", opts.LocationCode, "Verda location code; defaults to source OS volume location")
 	cmd.Flags().StringVar(&opts.StartupScriptName, "startup-script-name", opts.StartupScriptName, "Verda startup script name")
 	cmd.Flags().StringVar(&opts.userDataFile, "user-data-file", opts.userDataFile, "read userdata from a file instead of the embedded script")
 	cmd.Flags().StringVar(&opts.BaseURL, "base-url", opts.BaseURL, "Verda API base URL")

nnctl/internal/cloud/verda/deploy.go

@@ -17,7 +17,8 @@ const (
 	ExplicitLocation       = "explicit"
 	SourceOSVolumeLocation = "source_os_volume_location"
 	SpotContract           = "SPOT"
-	DefaultDescription     = "nnctl GPU benchmark worker"
+	DefaultDescription     = "nnctl benchmark worker"
+	MaxDeployGPUCount      = 1
 	VolumeReadyTimeout     = 20 * time.Minute
 	VolumeReadyPoll        = 5 * time.Second
 )
@@ -131,6 +132,8 @@ type DeployPolicy struct {
 	SourceOSVolumeLocked           bool   `json:"source_os_volume_locked"`
 	SpotOnly                       bool   `json:"spot_only"`
 	SingleGPU                      bool   `json:"single_gpu"`
+	AllowsCPU                      bool   `json:"allows_cpu"`
+	MaxGPUCount                    int    `json:"max_gpu_count"`
 	CleanupClonedOSVolumeOnFailure bool   `json:"cleanup_cloned_os_volume_on_failure"`
 }
 
@@ -182,6 +185,8 @@ func Deploy(ctx context.Context, client Client, opts DeployOptions) (*DeployResu
 			SourceOSVolumeLocked:           true,
 			SpotOnly:                       true,
 			SingleGPU:                      true,
+			AllowsCPU:                      true,
+			MaxGPUCount:                    MaxDeployGPUCount,
 			CleanupClonedOSVolumeOnFailure: !normalized.KeepClonedOSVolumeOnFailure,
 		},
 		SourceOSVolumeID: normalized.SourceOSVolumeID,
@@ -219,8 +224,8 @@ func Deploy(ctx context.Context, client Client, opts DeployOptions) (*DeployResu
 	if err != nil {
 		return nil, fmt.Errorf("get Verda instance type %q: %w", normalized.InstanceType, err)
 	}
-	if instanceType.GPUCount != 1 {
-		return nil, fmt.Errorf("instance type %q has %d GPUs; nnctl cloud deploy only accepts single-GPU instance types", normalized.InstanceType, instanceType.GPUCount)
+	if instanceType.GPUCount > MaxDeployGPUCount {
+		return nil, fmt.Errorf("instance type %q has %d GPUs; nnctl cloud deploy accepts CPU-only and single-GPU instance types", normalized.InstanceType, instanceType.GPUCount)
 	}
 	result.InstanceType = &instanceType
 

nnctl/internal/cloud/verda/deploy_test.go

@@ -54,7 +54,7 @@ func TestDeployDryRunBuildsSpotAutoPlacementRequestWithoutClient(t *testing.T) {
 	if !result.DryRun {
 		t.Fatal("DryRun = false")
 	}
-	if result.Policy.LocationCode != "" || result.Policy.LocationSelection != SourceOSVolumeLocation || !result.Policy.SourceOSVolumeLocked || !result.Policy.SpotOnly || !result.Policy.SingleGPU || !result.Policy.CleanupClonedOSVolumeOnFailure {
+	if result.Policy.LocationCode != "" || result.Policy.LocationSelection != SourceOSVolumeLocation || !result.Policy.SourceOSVolumeLocked || !result.Policy.SpotOnly || !result.Policy.AllowsCPU || result.Policy.MaxGPUCount != MaxDeployGPUCount || !result.Policy.CleanupClonedOSVolumeOnFailure {
 		t.Fatalf("unexpected policy: %#v", result.Policy)
 	}
 	if !result.Request.IsSpot || result.Request.Contract != SpotContract || result.Request.LocationCode != "" {
@@ -80,14 +80,46 @@ func TestDeployRejectsMultiGPUInstanceTypes(t *testing.T) {
 	if err == nil {
 		t.Fatal("expected multi-GPU instance type error")
 	}
-	if !strings.Contains(err.Error(), "single-GPU") {
-		t.Fatalf("error did not mention single-GPU policy: %v", err)
+	if !strings.Contains(err.Error(), "CPU-only and single-GPU") {
+		t.Fatalf("error did not mention CPU/single-GPU policy: %v", err)
 	}
 	if client.createdScript {
 		t.Fatal("startup script should not be created for rejected instance type")
 	}
 }
 
+func TestDeployCreatesSpotCPUInstanceInSourceOSVolumeLocation(t *testing.T) {
+	client := &fakeClient{
+		sourceVolume: Volume{ID: "vol-golden", Name: "golden", Status: "detached", Location: FinlandLocationCode, IsOSVolume: true},
+		instanceType: InstanceType{InstanceType: "4C.16M", GPUCount: 0},
+		placements: []SpotPlacement{
+			{LocationCode: FinlandLocationCode, SpotPrice: 1, PriceKnown: true, Currency: "eur"},
+		},
+		script:       StartupScript{ID: "script-1", Name: "userdata"},
+		clonedVolume: Volume{ID: "vol-clone-1", Name: "worker-os", Status: "cloning", Location: FinlandLocationCode},
+		readyVolume:  Volume{ID: "vol-clone-1", Name: "worker-os", Status: "detached", Location: FinlandLocationCode},
+		instance:     Instance{ID: "inst-1", Hostname: "worker", Status: "new", InstanceType: "4C.16M", Location: FinlandLocationCode, IsSpot: true},
+	}
+	opts := DefaultDeployOptions("4C.16M")
+	opts.SourceOSVolumeID = "vol-golden"
+	opts.Hostname = "worker"
+
+	result, err := Deploy(context.Background(), client, opts)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if result.InstanceType == nil || result.InstanceType.GPUCount != 0 {
+		t.Fatalf("unexpected instance type result: %#v", result.InstanceType)
+	}
+	if result.Policy.MaxGPUCount != MaxDeployGPUCount || !result.Policy.AllowsCPU {
+		t.Fatalf("unexpected policy: %#v", result.Policy)
+	}
+	if client.createRequest.InstanceType != "4C.16M" || client.createRequest.LocationCode != FinlandLocationCode {
+		t.Fatalf("unexpected create request: %#v", client.createRequest)
+	}
+}
+
 func TestDeployCreatesSpotInstanceInSourceOSVolumeLocation(t *testing.T) {
 	client := &fakeClient{
 		sourceVolume: Volume{ID: "vol-golden", Name: "golden", Status: "detached", Location: FinlandLocationCode, IsOSVolume: true},