diff --git a/api/openapi-spec/swagger.json b/api/openapi-spec/swagger.json
index 5731c0204f..25c7036bd8 100644
--- a/api/openapi-spec/swagger.json
+++ b/api/openapi-spec/swagger.json
@@ -517,7 +517,7 @@
         },
         "numProcPerNode": {
           "description": "Number of processes per node. This value is inserted into the `--nproc-per-node` argument of the `torchrun` CLI. Supported values: `auto`, `cpu`, `gpu`, or int value. Defaults to `auto`.",
-          "type": "string"
+          "$ref": "#/definitions/k8s.io.apimachinery.pkg.util.intstr.IntOrString"
         }
       }
     },
@@ -716,7 +716,7 @@
         },
         "numProcPerNode": {
           "description": "Number of processes/workers/slots on every training node. For the Torch runtime: `auto`, `cpu`, `gpu`, or int value can be set. For the MPI runtime only int value can be set.",
-          "type": "string"
+          "$ref": "#/definitions/k8s.io.apimachinery.pkg.util.intstr.IntOrString"
         },
         "resourcesPerNode": {
           "description": "Compute resources for each training node.",
diff --git a/manifests/base/crds/trainer.kubeflow.org_clustertrainingruntimes.yaml b/manifests/base/crds/trainer.kubeflow.org_clustertrainingruntimes.yaml
index 2765763a36..a57a33e00f 100644
--- a/manifests/base/crds/trainer.kubeflow.org_clustertrainingruntimes.yaml
+++ b/manifests/base/crds/trainer.kubeflow.org_clustertrainingruntimes.yaml
@@ -587,17 +587,20 @@ spec:
                             type: integer
                         type: object
                       numProcPerNode:
+                        anyOf:
+                        - type: integer
+                        - type: string
                         default: auto
                         description: |-
                           Number of processes per node.
                           This value is inserted into the `--nproc-per-node` argument of the `torchrun` CLI.
                           Supported values: `auto`, `cpu`, `gpu`, or int value.
                           Defaults to `auto`.
-                        type: string
+                        x-kubernetes-int-or-string: true
                         x-kubernetes-validations:
                         - message: NumProcPerNode must be equal to auto, cpu, gpu,
                             or int value
-                          rule: self in ['auto', 'cpu', 'gpu'] || type(self) == int
+                          rule: self > 0 || self in ['auto', 'cpu', 'gpu']
                     type: object
                 type: object
                 x-kubernetes-validations:
diff --git a/manifests/base/crds/trainer.kubeflow.org_trainingruntimes.yaml b/manifests/base/crds/trainer.kubeflow.org_trainingruntimes.yaml
index 0800575e0a..c7725e57a8 100644
--- a/manifests/base/crds/trainer.kubeflow.org_trainingruntimes.yaml
+++ b/manifests/base/crds/trainer.kubeflow.org_trainingruntimes.yaml
@@ -587,17 +587,20 @@ spec:
                             type: integer
                         type: object
                       numProcPerNode:
+                        anyOf:
+                        - type: integer
+                        - type: string
                         default: auto
                         description: |-
                           Number of processes per node.
                           This value is inserted into the `--nproc-per-node` argument of the `torchrun` CLI.
                           Supported values: `auto`, `cpu`, `gpu`, or int value.
                           Defaults to `auto`.
-                        type: string
+                        x-kubernetes-int-or-string: true
                         x-kubernetes-validations:
                         - message: NumProcPerNode must be equal to auto, cpu, gpu,
                             or int value
-                          rule: self in ['auto', 'cpu', 'gpu'] || type(self) == int
+                          rule: self > 0 || self in ['auto', 'cpu', 'gpu']
                     type: object
                 type: object
                 x-kubernetes-validations:
diff --git a/manifests/base/crds/trainer.kubeflow.org_trainjobs.yaml b/manifests/base/crds/trainer.kubeflow.org_trainjobs.yaml
index 5c4c7cb7f4..16a2294cb9 100644
--- a/manifests/base/crds/trainer.kubeflow.org_trainjobs.yaml
+++ b/manifests/base/crds/trainer.kubeflow.org_trainjobs.yaml
@@ -3138,11 +3138,14 @@ spec:
                     format: int32
                     type: integer
                   numProcPerNode:
+                    anyOf:
+                    - type: integer
+                    - type: string
                     description: |-
                       Number of processes/workers/slots on every training node.
                       For the Torch runtime: `auto`, `cpu`, `gpu`, or int value can be set.
                       For the MPI runtime only int value can be set.
-                    type: string
+                    x-kubernetes-int-or-string: true
                   resourcesPerNode:
                     description: Compute resources for each training node.
                     properties:
diff --git a/pkg/apis/trainer/v1alpha1/trainingruntime_types.go b/pkg/apis/trainer/v1alpha1/trainingruntime_types.go
index 6d55280348..6aec8be8d2 100644
--- a/pkg/apis/trainer/v1alpha1/trainingruntime_types.go
+++ b/pkg/apis/trainer/v1alpha1/trainingruntime_types.go
@@ -19,6 +19,7 @@ package v1alpha1
 import (
 	autoscalingv2 "k8s.io/api/autoscaling/v2"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/util/intstr"
 	jobsetv1alpha2 "sigs.k8s.io/jobset/api/jobset/v1alpha2"
 )
 
@@ -174,11 +175,10 @@ type TorchMLPolicySource struct {
 	// Number of processes per node.
 	// This value is inserted into the `--nproc-per-node` argument of the `torchrun` CLI.
 	// Supported values: `auto`, `cpu`, `gpu`, or int value.
-	// TODO (andreyvelich): Add kubebuilder validation.
 	// Defaults to `auto`.
 	// +kubebuilder:default="auto"
-	// +kubebuilder:validation:XValidation:rule="self in ['auto', 'cpu', 'gpu'] || type(self) == int", message="NumProcPerNode must be equal to auto, cpu, gpu, or int value"
-	NumProcPerNode *string `json:"numProcPerNode,omitempty"`
+	// +kubebuilder:validation:XValidation:rule="self > 0 || self in ['auto', 'cpu', 'gpu']", message="NumProcPerNode must be equal to auto, cpu, gpu, or int value"
+	NumProcPerNode *intstr.IntOrString `json:"numProcPerNode,omitempty"`
 
 	// Elastic policy for the PyTorch training.
 	ElasticPolicy *TorchElasticPolicy `json:"elasticPolicy,omitempty"`
diff --git a/pkg/apis/trainer/v1alpha1/trainjob_types.go b/pkg/apis/trainer/v1alpha1/trainjob_types.go
index 03ef0816ad..0ea6ddbbf4 100644
--- a/pkg/apis/trainer/v1alpha1/trainjob_types.go
+++ b/pkg/apis/trainer/v1alpha1/trainjob_types.go
@@ -19,6 +19,7 @@ package v1alpha1
 import (
 	corev1 "k8s.io/api/core/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/util/intstr"
 )
 
 const (
@@ -194,7 +195,7 @@ type Trainer struct {
 	// Number of processes/workers/slots on every training node.
 	// For the Torch runtime: `auto`, `cpu`, `gpu`, or int value can be set.
 	// For the MPI runtime only int value can be set.
-	NumProcPerNode *string `json:"numProcPerNode,omitempty"`
+	NumProcPerNode *intstr.IntOrString `json:"numProcPerNode,omitempty"`
 }
 
 // DatasetConfig represents the desired dataset configuration.
diff --git a/pkg/apis/trainer/v1alpha1/zz_generated.deepcopy.go b/pkg/apis/trainer/v1alpha1/zz_generated.deepcopy.go
index 3548d01a18..40a9d91af6 100644
--- a/pkg/apis/trainer/v1alpha1/zz_generated.deepcopy.go
+++ b/pkg/apis/trainer/v1alpha1/zz_generated.deepcopy.go
@@ -24,6 +24,7 @@ import (
 	v1 "k8s.io/api/core/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	runtime "k8s.io/apimachinery/pkg/runtime"
+	intstr "k8s.io/apimachinery/pkg/util/intstr"
 )
 
 // DeepCopyInto is an autogenerated deepcopy function, copying the receiver, writing into out. in must be non-nil.
@@ -566,7 +567,7 @@ func (in *TorchMLPolicySource) DeepCopyInto(out *TorchMLPolicySource) {
 	*out = *in
 	if in.NumProcPerNode != nil {
 		in, out := &in.NumProcPerNode, &out.NumProcPerNode
-		*out = new(string)
+		*out = new(intstr.IntOrString)
 		**out = **in
 	}
 	if in.ElasticPolicy != nil {
@@ -776,7 +777,7 @@ func (in *Trainer) DeepCopyInto(out *Trainer) {
 	}
 	if in.NumProcPerNode != nil {
 		in, out := &in.NumProcPerNode, &out.NumProcPerNode
-		*out = new(string)
+		*out = new(intstr.IntOrString)
 		**out = **in
 	}
 	return
diff --git a/pkg/apis/trainer/v1alpha1/zz_generated.openapi.go b/pkg/apis/trainer/v1alpha1/zz_generated.openapi.go
index b8d53431b8..c0ee8376c7 100644
--- a/pkg/apis/trainer/v1alpha1/zz_generated.openapi.go
+++ b/pkg/apis/trainer/v1alpha1/zz_generated.openapi.go
@@ -974,8 +974,7 @@ func schema_pkg_apis_trainer_v1alpha1_TorchMLPolicySource(ref common.ReferenceCa
 					"numProcPerNode": {
 						SchemaProps: spec.SchemaProps{
 							Description: "Number of processes per node. This value is inserted into the `--nproc-per-node` argument of the `torchrun` CLI. Supported values: `auto`, `cpu`, `gpu`, or int value. Defaults to `auto`.",
-							Type:        []string{"string"},
-							Format:      "",
+							Ref:         ref("k8s.io/apimachinery/pkg/util/intstr.IntOrString"),
 						},
 					},
 					"elasticPolicy": {
@@ -988,7 +987,7 @@ func schema_pkg_apis_trainer_v1alpha1_TorchMLPolicySource(ref common.ReferenceCa
 			},
 		},
 		Dependencies: []string{
-			"github.com/kubeflow/trainer/pkg/apis/trainer/v1alpha1.TorchElasticPolicy"},
+			"github.com/kubeflow/trainer/pkg/apis/trainer/v1alpha1.TorchElasticPolicy", "k8s.io/apimachinery/pkg/util/intstr.IntOrString"},
 	}
 }
 
@@ -1352,15 +1351,14 @@ func schema_pkg_apis_trainer_v1alpha1_Trainer(ref common.ReferenceCallback) comm
 					"numProcPerNode": {
 						SchemaProps: spec.SchemaProps{
 							Description: "Number of processes/workers/slots on every training node. For the Torch runtime: `auto`, `cpu`, `gpu`, or int value can be set. For the MPI runtime only int value can be set.",
-							Type:        []string{"string"},
-							Format:      "",
+							Ref:         ref("k8s.io/apimachinery/pkg/util/intstr.IntOrString"),
 						},
 					},
 				},
 			},
 		},
 		Dependencies: []string{
-			"k8s.io/api/core/v1.EnvVar", "k8s.io/api/core/v1.ResourceRequirements"},
+			"k8s.io/api/core/v1.EnvVar", "k8s.io/api/core/v1.ResourceRequirements", "k8s.io/apimachinery/pkg/util/intstr.IntOrString"},
 	}
 }
 
diff --git a/pkg/client/applyconfiguration/trainer/v1alpha1/torchmlpolicysource.go b/pkg/client/applyconfiguration/trainer/v1alpha1/torchmlpolicysource.go
index c9d14b1ec1..6c0bae78dc 100644
--- a/pkg/client/applyconfiguration/trainer/v1alpha1/torchmlpolicysource.go
+++ b/pkg/client/applyconfiguration/trainer/v1alpha1/torchmlpolicysource.go
@@ -16,10 +16,14 @@
 
 package v1alpha1
 
+import (
+	intstr "k8s.io/apimachinery/pkg/util/intstr"
+)
+
 // TorchMLPolicySourceApplyConfiguration represents a declarative configuration of the TorchMLPolicySource type for use
 // with apply.
 type TorchMLPolicySourceApplyConfiguration struct {
-	NumProcPerNode *string                               `json:"numProcPerNode,omitempty"`
+	NumProcPerNode *intstr.IntOrString                   `json:"numProcPerNode,omitempty"`
 	ElasticPolicy  *TorchElasticPolicyApplyConfiguration `json:"elasticPolicy,omitempty"`
 }
 
@@ -32,7 +36,7 @@ func TorchMLPolicySource() *TorchMLPolicySourceApplyConfiguration {
 // WithNumProcPerNode sets the NumProcPerNode field in the declarative configuration to the given value
 // and returns the receiver, so that objects can be built by chaining "With" function invocations.
 // If called multiple times, the NumProcPerNode field is set to the value of the last call.
-func (b *TorchMLPolicySourceApplyConfiguration) WithNumProcPerNode(value string) *TorchMLPolicySourceApplyConfiguration {
+func (b *TorchMLPolicySourceApplyConfiguration) WithNumProcPerNode(value intstr.IntOrString) *TorchMLPolicySourceApplyConfiguration {
 	b.NumProcPerNode = &value
 	return b
 }
diff --git a/pkg/client/applyconfiguration/trainer/v1alpha1/trainer.go b/pkg/client/applyconfiguration/trainer/v1alpha1/trainer.go
index d9991106ab..66e1aed89d 100644
--- a/pkg/client/applyconfiguration/trainer/v1alpha1/trainer.go
+++ b/pkg/client/applyconfiguration/trainer/v1alpha1/trainer.go
@@ -17,6 +17,7 @@
 package v1alpha1
 
 import (
+	intstr "k8s.io/apimachinery/pkg/util/intstr"
 	v1 "k8s.io/client-go/applyconfigurations/core/v1"
 )
 
@@ -29,7 +30,7 @@ type TrainerApplyConfiguration struct {
 	Env              []v1.EnvVarApplyConfiguration              `json:"env,omitempty"`
 	NumNodes         *int32                                     `json:"numNodes,omitempty"`
 	ResourcesPerNode *v1.ResourceRequirementsApplyConfiguration `json:"resourcesPerNode,omitempty"`
-	NumProcPerNode   *string                                    `json:"numProcPerNode,omitempty"`
+	NumProcPerNode   *intstr.IntOrString                        `json:"numProcPerNode,omitempty"`
 }
 
 // TrainerApplyConfiguration constructs a declarative configuration of the Trainer type for use with
@@ -98,7 +99,7 @@ func (b *TrainerApplyConfiguration) WithResourcesPerNode(value *v1.ResourceRequi
 // WithNumProcPerNode sets the NumProcPerNode field in the declarative configuration to the given value
 // and returns the receiver, so that objects can be built by chaining "With" function invocations.
 // If called multiple times, the NumProcPerNode field is set to the value of the last call.
-func (b *TrainerApplyConfiguration) WithNumProcPerNode(value string) *TrainerApplyConfiguration {
+func (b *TrainerApplyConfiguration) WithNumProcPerNode(value intstr.IntOrString) *TrainerApplyConfiguration {
 	b.NumProcPerNode = &value
 	return b
 }
diff --git a/pkg/runtime/core/trainingruntime_test.go b/pkg/runtime/core/trainingruntime_test.go
index 4274136a53..0c04ff8b64 100644
--- a/pkg/runtime/core/trainingruntime_test.go
+++ b/pkg/runtime/core/trainingruntime_test.go
@@ -19,7 +19,6 @@ package core
 import (
 	"context"
 	"fmt"
-	"k8s.io/utils/ptr"
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
@@ -27,6 +26,7 @@ import (
 	corev1 "k8s.io/api/core/v1"
 	"k8s.io/apimachinery/pkg/api/resource"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/util/intstr"
 	"sigs.k8s.io/controller-runtime/pkg/client"
 	schedulerpluginsv1alpha1 "sigs.k8s.io/scheduler-plugins/apis/scheduling/v1alpha1"
 
@@ -264,7 +264,7 @@ func TestTrainingRuntimeNewObjects(t *testing.T) {
 		"succeeded to build JobSet with Torch values from the TrainJob": {
 			trainingRuntime: testingutil.MakeTrainingRuntimeWrapper(metav1.NamespaceDefault, "test-runtime").RuntimeSpec(
 				testingutil.MakeTrainingRuntimeSpecWrapper(testingutil.MakeTrainingRuntimeWrapper(metav1.NamespaceDefault, "test-runtime").Spec).
-					TorchPolicy(100, ptr.To("auto")).
+					TorchPolicy(100, intstr.FromString("auto")).
 					ContainerTrainer("test:runtime", []string{"runtime"}, []string{"runtime"}, resRequests).
 					Obj(),
 			).Obj(),
@@ -274,7 +274,7 @@ func TestTrainingRuntimeNewObjects(t *testing.T) {
 				Trainer(
 					testingutil.MakeTrainJobTrainerWrapper().
 						NumNodes(30).
-						NumProcPerNode(ptr.To("3")).
+						NumProcPerNode(intstr.FromInt32(3)).
 						Obj(),
 				).
 				Obj(),
@@ -318,7 +318,7 @@ func TestTrainingRuntimeNewObjects(t *testing.T) {
 		"succeeded to build JobSet with Torch values from the Runtime and envs.": {
 			trainingRuntime: testingutil.MakeTrainingRuntimeWrapper(metav1.NamespaceDefault, "test-runtime").RuntimeSpec(
 				testingutil.MakeTrainingRuntimeSpecWrapper(testingutil.MakeTrainingRuntimeWrapper(metav1.NamespaceDefault, "test-runtime").Spec).
-					TorchPolicy(100, ptr.To("auto")).
+					TorchPolicy(100, intstr.FromString("auto")).
 					ContainerTrainer("test:runtime", []string{"runtime"}, []string{"runtime"}, resRequests).
 					ContainerTrainerEnv(
 						[]corev1.EnvVar{
diff --git a/pkg/runtime/framework/plugins/mpi/mpi.go b/pkg/runtime/framework/plugins/mpi/mpi.go
index 5d02679d05..d26c58c7c4 100644
--- a/pkg/runtime/framework/plugins/mpi/mpi.go
+++ b/pkg/runtime/framework/plugins/mpi/mpi.go
@@ -94,7 +94,7 @@ func (m *MPI) EnforceMLPolicy(info *runtime.Info, trainJob *trainer.TrainJob) er
 
 	numProcPerNode := strconv.Itoa(int(*info.RuntimePolicy.MLPolicy.MPI.NumProcPerNode))
 	if trainJob.Spec.Trainer != nil && trainJob.Spec.Trainer.NumProcPerNode != nil {
-		numProcPerNode = *trainJob.Spec.Trainer.NumProcPerNode
+		numProcPerNode = (*trainJob.Spec.Trainer.NumProcPerNode).String()
 	}
 	info.Trainer.NumProcPerNode = numProcPerNode
 
diff --git a/pkg/runtime/framework/plugins/torch/torch.go b/pkg/runtime/framework/plugins/torch/torch.go
index 02e86073c4..af9a04c456 100644
--- a/pkg/runtime/framework/plugins/torch/torch.go
+++ b/pkg/runtime/framework/plugins/torch/torch.go
@@ -21,6 +21,7 @@ import (
 	"fmt"
 
 	corev1 "k8s.io/api/core/v1"
+	"k8s.io/apimachinery/pkg/util/intstr"
 	"k8s.io/apimachinery/pkg/util/sets"
 	"k8s.io/apimachinery/pkg/util/validation/field"
 	"k8s.io/utils/ptr"
@@ -66,9 +67,9 @@ func (t *Torch) EnforceMLPolicy(info *runtime.Info, trainJob *trainer.TrainJob)
 	}
 	info.Trainer.NumNodes = numNodes
 
-	numProcPerNode := info.RuntimePolicy.MLPolicy.Torch.NumProcPerNode
+	numProcPerNode := ptr.Deref(info.RuntimePolicy.MLPolicy.Torch.NumProcPerNode, intstr.FromString("auto"))
 	if trainJob.Spec.Trainer != nil && trainJob.Spec.Trainer.NumProcPerNode != nil {
-		numProcPerNode = trainJob.Spec.Trainer.NumProcPerNode
+		numProcPerNode = ptr.Deref(trainJob.Spec.Trainer.NumProcPerNode, intstr.FromString("auto"))
 	}
 
 	// Update envs for Info object.
@@ -84,7 +85,7 @@ func (t *Torch) EnforceMLPolicy(info *runtime.Info, trainJob *trainer.TrainJob)
 		},
 		{
 			Name:  constants.TorchEnvNumProcPerNode,
-			Value: ptr.Deref(numProcPerNode, "auto"),
+			Value: numProcPerNode.String(),
 		},
 		{
 			Name: constants.TorchEnvNodeRank,
diff --git a/pkg/util/testing/wrapper.go b/pkg/util/testing/wrapper.go
index 23b38545cd..f2d735b12f 100644
--- a/pkg/util/testing/wrapper.go
+++ b/pkg/util/testing/wrapper.go
@@ -22,6 +22,7 @@ import (
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/runtime/schema"
 	"k8s.io/apimachinery/pkg/types"
+	"k8s.io/apimachinery/pkg/util/intstr"
 	"k8s.io/utils/ptr"
 	jobsetv1alpha2 "sigs.k8s.io/jobset/api/jobset/v1alpha2"
 	schedulerpluginsv1alpha1 "sigs.k8s.io/scheduler-plugins/apis/scheduling/v1alpha1"
@@ -392,8 +393,8 @@ func (t *TrainJobTrainerWrapper) NumNodes(numNodes int32) *TrainJobTrainerWrappe
 	return t
 }
 
-func (t *TrainJobTrainerWrapper) NumProcPerNode(numProcPerNode *string) *TrainJobTrainerWrapper {
-	t.Trainer.NumProcPerNode = numProcPerNode
+func (t *TrainJobTrainerWrapper) NumProcPerNode(numProcPerNode intstr.IntOrString) *TrainJobTrainerWrapper {
+	t.Trainer.NumProcPerNode = &numProcPerNode
 	return t
 }
 
@@ -689,12 +690,12 @@ func (s *TrainingRuntimeSpecWrapper) NumNodes(numNodes int32) *TrainingRuntimeSp
 	return s
 }
 
-func (s *TrainingRuntimeSpecWrapper) TorchPolicy(numNodes int32, numProcPerNode *string) *TrainingRuntimeSpecWrapper {
+func (s *TrainingRuntimeSpecWrapper) TorchPolicy(numNodes int32, numProcPerNode intstr.IntOrString) *TrainingRuntimeSpecWrapper {
 	s.MLPolicy = &trainer.MLPolicy{
 		NumNodes: &numNodes,
 		MLPolicySource: trainer.MLPolicySource{
 			Torch: &trainer.TorchMLPolicySource{
-				NumProcPerNode: numProcPerNode,
+				NumProcPerNode: &numProcPerNode,
 			},
 		},
 	}
diff --git a/sdk/docs/TrainerV1alpha1TorchMLPolicySource.md b/sdk/docs/TrainerV1alpha1TorchMLPolicySource.md
index 6dcb288e30..e1cf9b351f 100644
--- a/sdk/docs/TrainerV1alpha1TorchMLPolicySource.md
+++ b/sdk/docs/TrainerV1alpha1TorchMLPolicySource.md
@@ -5,7 +5,7 @@ TorchMLPolicySource represents a PyTorch runtime configuration.
 Name | Type | Description | Notes
 ------------ | ------------- | ------------- | -------------
 **elastic_policy** | [**TrainerV1alpha1TorchElasticPolicy**](TrainerV1alpha1TorchElasticPolicy.md) |  | [optional] 
-**num_proc_per_node** | **str** | Number of processes per node. This value is inserted into the `--nproc-per-node` argument of the `torchrun` CLI. Supported values: `auto`, `cpu`, `gpu`, or int value. Defaults to `auto`. | [optional] 
+**num_proc_per_node** | [**K8sIoApimachineryPkgUtilIntstrIntOrString**](K8sIoApimachineryPkgUtilIntstrIntOrString.md) |  | [optional] 
 
 [[Back to Model list]](../README.md#documentation-for-models) [[Back to API list]](../README.md#documentation-for-api-endpoints) [[Back to README]](../README.md)
 
diff --git a/sdk/docs/TrainerV1alpha1Trainer.md b/sdk/docs/TrainerV1alpha1Trainer.md
index ac3623f559..6af31458ad 100644
--- a/sdk/docs/TrainerV1alpha1Trainer.md
+++ b/sdk/docs/TrainerV1alpha1Trainer.md
@@ -9,7 +9,7 @@ Name | Type | Description | Notes
 **env** | [**list[V1EnvVar]**](V1EnvVar.md) | List of environment variables to set in the training container. These values will be merged with the TrainingRuntime's trainer environments. | [optional] 
 **image** | **str** | Docker image for the training container. | [optional] 
 **num_nodes** | **int** | Number of training nodes. | [optional] 
-**num_proc_per_node** | **str** | Number of processes/workers/slots on every training node. For the Torch runtime: `auto`, `cpu`, `gpu`, or int value can be set. For the MPI runtime only int value can be set. | [optional] 
+**num_proc_per_node** | [**K8sIoApimachineryPkgUtilIntstrIntOrString**](K8sIoApimachineryPkgUtilIntstrIntOrString.md) |  | [optional] 
 **resources_per_node** | [**V1ResourceRequirements**](V1ResourceRequirements.md) |  | [optional] 
 
 [[Back to Model list]](../README.md#documentation-for-models) [[Back to API list]](../README.md#documentation-for-api-endpoints) [[Back to README]](../README.md)
diff --git a/sdk/kubeflow/trainer/models/trainer_v1alpha1_torch_ml_policy_source.py b/sdk/kubeflow/trainer/models/trainer_v1alpha1_torch_ml_policy_source.py
index be0bcca026..6a4c6c6022 100644
--- a/sdk/kubeflow/trainer/models/trainer_v1alpha1_torch_ml_policy_source.py
+++ b/sdk/kubeflow/trainer/models/trainer_v1alpha1_torch_ml_policy_source.py
@@ -34,7 +34,7 @@ class TrainerV1alpha1TorchMLPolicySource(object):
     """
     openapi_types = {
         'elastic_policy': 'TrainerV1alpha1TorchElasticPolicy',
-        'num_proc_per_node': 'str'
+        'num_proc_per_node': 'K8sIoApimachineryPkgUtilIntstrIntOrString'
     }
 
     attribute_map = {
@@ -82,10 +82,9 @@ def elastic_policy(self, elastic_policy):
     def num_proc_per_node(self):
         """Gets the num_proc_per_node of this TrainerV1alpha1TorchMLPolicySource.  # noqa: E501
 
-        Number of processes per node. This value is inserted into the `--nproc-per-node` argument of the `torchrun` CLI. Supported values: `auto`, `cpu`, `gpu`, or int value. Defaults to `auto`.  # noqa: E501
 
         :return: The num_proc_per_node of this TrainerV1alpha1TorchMLPolicySource.  # noqa: E501
-        :rtype: str
+        :rtype: K8sIoApimachineryPkgUtilIntstrIntOrString
         """
         return self._num_proc_per_node
 
@@ -93,10 +92,9 @@ def num_proc_per_node(self):
     def num_proc_per_node(self, num_proc_per_node):
         """Sets the num_proc_per_node of this TrainerV1alpha1TorchMLPolicySource.
 
-        Number of processes per node. This value is inserted into the `--nproc-per-node` argument of the `torchrun` CLI. Supported values: `auto`, `cpu`, `gpu`, or int value. Defaults to `auto`.  # noqa: E501
 
         :param num_proc_per_node: The num_proc_per_node of this TrainerV1alpha1TorchMLPolicySource.  # noqa: E501
-        :type: str
+        :type: K8sIoApimachineryPkgUtilIntstrIntOrString
         """
 
         self._num_proc_per_node = num_proc_per_node
diff --git a/sdk/kubeflow/trainer/models/trainer_v1alpha1_trainer.py b/sdk/kubeflow/trainer/models/trainer_v1alpha1_trainer.py
index 18ac69b23a..57e857b971 100644
--- a/sdk/kubeflow/trainer/models/trainer_v1alpha1_trainer.py
+++ b/sdk/kubeflow/trainer/models/trainer_v1alpha1_trainer.py
@@ -38,7 +38,7 @@ class TrainerV1alpha1Trainer(object):
         'env': 'list[V1EnvVar]',
         'image': 'str',
         'num_nodes': 'int',
-        'num_proc_per_node': 'str',
+        'num_proc_per_node': 'K8sIoApimachineryPkgUtilIntstrIntOrString',
         'resources_per_node': 'V1ResourceRequirements'
     }
 
@@ -201,10 +201,9 @@ def num_nodes(self, num_nodes):
     def num_proc_per_node(self):
         """Gets the num_proc_per_node of this TrainerV1alpha1Trainer.  # noqa: E501
 
-        Number of processes/workers/slots on every training node. For the Torch runtime: `auto`, `cpu`, `gpu`, or int value can be set. For the MPI runtime only int value can be set.  # noqa: E501
 
         :return: The num_proc_per_node of this TrainerV1alpha1Trainer.  # noqa: E501
-        :rtype: str
+        :rtype: K8sIoApimachineryPkgUtilIntstrIntOrString
         """
         return self._num_proc_per_node
 
@@ -212,10 +211,9 @@ def num_proc_per_node(self):
     def num_proc_per_node(self, num_proc_per_node):
         """Sets the num_proc_per_node of this TrainerV1alpha1Trainer.
 
-        Number of processes/workers/slots on every training node. For the Torch runtime: `auto`, `cpu`, `gpu`, or int value can be set. For the MPI runtime only int value can be set.  # noqa: E501
 
         :param num_proc_per_node: The num_proc_per_node of this TrainerV1alpha1Trainer.  # noqa: E501
-        :type: str
+        :type: K8sIoApimachineryPkgUtilIntstrIntOrString
         """
 
         self._num_proc_per_node = num_proc_per_node
diff --git a/test/integration/controller/trainjob_controller_test.go b/test/integration/controller/trainjob_controller_test.go
index 88b1fb6688..9378554788 100644
--- a/test/integration/controller/trainjob_controller_test.go
+++ b/test/integration/controller/trainjob_controller_test.go
@@ -26,6 +26,7 @@ import (
 	"k8s.io/apimachinery/pkg/api/resource"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/runtime/schema"
+	"k8s.io/apimachinery/pkg/util/intstr"
 	"k8s.io/utils/ptr"
 	"sigs.k8s.io/controller-runtime/pkg/client"
 	jobsetv1alpha2 "sigs.k8s.io/jobset/api/jobset/v1alpha2"
@@ -278,7 +279,7 @@ var _ = ginkgo.Describe("TrainJob controller", ginkgo.Ordered, func() {
 			trainingRuntime = testingutil.MakeTrainingRuntimeWrapper(ns.Name, "alpha").
 				RuntimeSpec(
 					testingutil.MakeTrainingRuntimeSpecWrapper(testingutil.MakeTrainingRuntimeWrapper(metav1.NamespaceDefault, "alpha").Spec).
-						TorchPolicy(100, ptr.To("auto")).
+						TorchPolicy(100, intstr.FromString("auto")).
 						ContainerTrainer("test:runtime", []string{"runtime"}, []string{"runtime"}, resRequests).
 						Obj()).
 				Obj()
diff --git a/test/integration/webhooks/trainingruntime_webhook_test.go b/test/integration/webhooks/trainingruntime_webhook_test.go
index bffb37507d..839293a3d9 100644
--- a/test/integration/webhooks/trainingruntime_webhook_test.go
+++ b/test/integration/webhooks/trainingruntime_webhook_test.go
@@ -21,6 +21,7 @@ import (
 	"github.com/onsi/gomega"
 	corev1 "k8s.io/api/core/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/util/intstr"
 	"k8s.io/utils/ptr"
 	"sigs.k8s.io/controller-runtime/pkg/client"
 
@@ -164,7 +165,7 @@ var _ = ginkgo.Describe("TrainingRuntime marker validations and defaulting", gin
 					runtime.Spec.MLPolicy = &trainer.MLPolicy{
 						MLPolicySource: trainer.MLPolicySource{
 							Torch: &trainer.TorchMLPolicySource{
-								NumProcPerNode: ptr.To("auto"),
+								NumProcPerNode: ptr.To(intstr.FromString("auto")),
 							},
 						},
 					}