feat: allow MPS shared GPU requests greater than one

Honor FailRequestsGreaterThanOne in the MPS sharing strategy the same
way it is honored for time-slicing. Previously this flag was forcibly
set to true and then ignored, making it impossible to request more than
one shared GPU unit under MPS. Now the flag defaults to false, allowing
multi-unit MPS requests unless explicitly restricted.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: radu-malliu

api/config/v1/config.go

@@ -68,12 +68,11 @@ func NewConfig(c *cli.Context, flags []cli.Flag) (*Config, error) {
 		config.Flags.NvidiaDevRoot = config.Flags.NvidiaDriverRoot
 	}
 
-	// We explicitly set sharing.mps.failRequestsGreaterThanOne = true
-	// This can be relaxed in certain cases -- such as a single GPU -- but
-	// requires additional logic around when it's OK to combine requests and
-	// makes the semantics of a request unclear.
-	if config.Sharing.MPS != nil {
-		config.Sharing.MPS.FailRequestsGreaterThanOne = true
+	// Default sharing.mps.failRequestsGreaterThanOne to true if not explicitly set in config.
+	// Set it to false in the device plugin config to allow pods to request more than one MPS-shared GPU unit.
+	if config.Sharing.MPS != nil && config.Sharing.MPS.FailRequestsGreaterThanOne == nil {
+		t := true
+		config.Sharing.MPS.FailRequestsGreaterThanOne = &t
 	}
 
 	return config, nil

api/config/v1/replicas.go

@@ -28,7 +28,7 @@ import (
 // ReplicatedResources defines generic options for replicating devices.
 type ReplicatedResources struct {
 	RenameByDefault            bool                 `json:"renameByDefault,omitempty"            yaml:"renameByDefault,omitempty"`
-	FailRequestsGreaterThanOne bool                 `json:"failRequestsGreaterThanOne,omitempty" yaml:"failRequestsGreaterThanOne,omitempty"`
+	FailRequestsGreaterThanOne *bool                `json:"failRequestsGreaterThanOne,omitempty" yaml:"failRequestsGreaterThanOne,omitempty"`
 	Resources                  []ReplicatedResource `json:"resources,omitempty"                  yaml:"resources,omitempty"`
 }
 
@@ -179,14 +179,11 @@ func (s *ReplicatedResources) UnmarshalJSON(b []byte) error {
 		return err
 	}
 
-	failRequestsGreaterThanOne, exists := ts["failRequestsGreaterThanOne"]
-	if !exists {
-		failRequestsGreaterThanOne = []byte(`false`)
-	}
-
-	err = json.Unmarshal(failRequestsGreaterThanOne, &s.FailRequestsGreaterThanOne)
-	if err != nil {
-		return err
+	if failRequestsGreaterThanOne, exists := ts["failRequestsGreaterThanOne"]; exists {
+		err = json.Unmarshal(failRequestsGreaterThanOne, &s.FailRequestsGreaterThanOne)
+		if err != nil {
+			return err
+		}
 	}
 
 	resources, exists := ts["resources"]

internal/rm/rm.go

@@ -74,20 +74,14 @@ func (r *resourceManager) ValidateRequest(ids AnnotatedIDs) error {
 	// error out if more than one resource is being allocated.
 	includesReplicas := ids.AnyHasAnnotations()
 	numRequestedDevices := len(ids)
+	failRequestsGTOne := r.config.Sharing.ReplicatedResources().FailRequestsGreaterThanOne
 	switch r.config.Sharing.SharingStrategy() {
 	case spec.SharingStrategyTimeSlicing:
-		if includesReplicas && numRequestedDevices > 1 && r.config.Sharing.ReplicatedResources().FailRequestsGreaterThanOne {
+		if includesReplicas && numRequestedDevices > 1 && failRequestsGTOne != nil && *failRequestsGTOne {
 			return fmt.Errorf("%w: maximum request size for shared resources is 1; found %d", errInvalidRequest, numRequestedDevices)
 		}
 	case spec.SharingStrategyMPS:
-		// For MPS sharing, we explicitly ignore the FailRequestsGreaterThanOne
-		// value in the sharing settings.
-		// This setting was added to timeslicing after the initial release and
-		// is set to `false` to maintain backward compatibility with existing
-		// deployments. If we do extend MPS to allow multiple devices to be
-		// requested, the MPS API will be extended separately from the
-		// time-slicing API.
-		if includesReplicas && numRequestedDevices > 1 {
+		if includesReplicas && numRequestedDevices > 1 && failRequestsGTOne != nil && *failRequestsGTOne {
 			return fmt.Errorf("%w: maximum request size for shared resources is 1; found %d", errInvalidRequest, numRequestedDevices)
 		}
 	}

internal/rm/rm_test.go

@@ -20,6 +20,7 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/require"
+	"k8s.io/utils/ptr"
 
 	spec "github.com/NVIDIA/k8s-device-plugin/api/config/v1"
 )
@@ -94,7 +95,7 @@ func TestValidateRequest(t *testing.T) {
 			description: "timeslicing with two devices -- failRequestsGreaterThanOne",
 			sharing: spec.Sharing{
 				TimeSlicing: spec.ReplicatedResources{
-					FailRequestsGreaterThanOne: true,
+					FailRequestsGreaterThanOne: ptr.To(true),
 					Resources: []spec.ReplicatedResource{
 						{
 							Name:     "nvidia.com/gpu",
@@ -151,13 +152,12 @@ func TestValidateRequest(t *testing.T) {
 				"device1::1": nil,
 			},
 			requestDevicesIDs: []string{"device0::1", "device1::0"},
-			expectedError:     errInvalidRequest,
 		},
 		{
 			description: "MPS with two devices -- failRequestsGreaterThanOne",
 			sharing: spec.Sharing{
 				MPS: &spec.ReplicatedResources{
-					FailRequestsGreaterThanOne: true,
+					FailRequestsGreaterThanOne: ptr.To(true),
 					Resources: []spec.ReplicatedResource{
 						{
 							Name:     "nvidia.com/gpu",