Support shared memory size

.github/ISSUE_TEMPLATE/BUG_REPORT.md

@@ -21,7 +21,6 @@ assignees: ''
 **Environment**:
 
 - Kubernetes version (use `kubectl version`):
-- LWS version:
 - llmaz version (use `git describe --tags --dirty --always`):
 - Cloud provider or hardware configuration:
 - OS (e.g: `cat /etc/os-release`):

.github/ISSUE_TEMPLATE/ENHANCEMENT.md

@@ -2,7 +2,7 @@
 name: Enhancement Request
 about: Suggest an enhancement to the llmaz project
 title: ''
-labels: kind/feature
+labels: feature
 assignees: ''
 
 ---

.golangci.yaml

@@ -1,11 +1,6 @@
 run:
-  deadline: 5m
+  timeout: 5m
   allow-parallel-runners: true
-  dirs:
-    - api
-    - cmd
-    - pkg
-    - test
 
 issues:
   # don't skip warning about doc comments

api/core/v1alpha1/model_types.go

@@ -18,6 +18,7 @@ package v1alpha1
 
 import (
 	v1 "k8s.io/api/core/v1"
+	"k8s.io/apimachinery/pkg/api/resource"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 )
 
@@ -131,6 +132,9 @@ type InferenceConfig struct {
 	// +kubebuilder:validation:MaxItems=8
 	// +optional
 	Flavors []Flavor `json:"flavors,omitempty"`
+	// SharedMemorySize represents the size of /dev/shm required in the runtime of
+	// inference workload.
+	SharedMemorySize *resource.Quantity `json:"sharedMemorySize,omitempty"`
 }
 
 type ModelName string

config/crd/bases/llmaz.io_openmodels.yaml

@@ -105,6 +105,15 @@ spec:
                       type: object
                     maxItems: 8
                     type: array
+                  sharedMemorySize:
+                    anyOf:
+                    - type: integer
+                    - type: string
+                    description: |-
+                      SharedMemorySize represents the size of /dev/shm required in the runtime of
+                      inference workload.
+                    pattern: ^(\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))))?$
+                    x-kubernetes-int-or-string: true
                 type: object
               source:
                 description: |-

pkg/controller/inference/playground_controller.go

@@ -265,7 +265,7 @@ func buildWorkloadTemplate(models []*coreapi.OpenModel, playground *inferenceapi
 
 	if multiHost {
 		workload.LeaderWorkerTemplate.LeaderTemplate = &template
-		workload.LeaderWorkerTemplate.WorkerTemplate = buildWorkerTemplate(playground, backendRuntime)
+		workload.LeaderWorkerTemplate.WorkerTemplate = buildWorkerTemplate(models, playground, backendRuntime)
 	} else {
 		workload.LeaderWorkerTemplate.WorkerTemplate = template
 	}
@@ -366,12 +366,30 @@ func buildTemplate(models []*coreapi.OpenModel, playground *inferenceapi.Playgro
 		},
 	}
 
+	// construct /dev/shm size
+	if models[0].Spec.InferenceConfig != nil && models[0].Spec.InferenceConfig.SharedMemorySize != nil {
+		template.Spec.Volumes = append(template.Spec.Volumes, corev1.Volume{
+			Name: "dshm",
+			VolumeSource: corev1.VolumeSource{
+				EmptyDir: &corev1.EmptyDirVolumeSource{
+					Medium:    corev1.StorageMediumMemory,
+					SizeLimit: models[0].Spec.InferenceConfig.SharedMemorySize,
+				},
+			},
+		})
+
+		template.Spec.Containers[0].VolumeMounts = append(template.Spec.Containers[0].VolumeMounts, corev1.VolumeMount{
+			Name:      "dshm",
+			MountPath: "/dev/shm",
+		})
+	}
+
 	return template, nil
 }
 
 // This is a copy of buildTemplate with some refactors, only used in multi-nodes cases.
 // Worker template has no args, no contain port.
-func buildWorkerTemplate(playground *inferenceapi.Playground, backendRuntime *inferenceapi.BackendRuntime) corev1.PodTemplateSpec {
+func buildWorkerTemplate(models []*coreapi.OpenModel, playground *inferenceapi.Playground, backendRuntime *inferenceapi.BackendRuntime) corev1.PodTemplateSpec {
 	parser := helper.NewBackendRuntimeParser(backendRuntime)
 
 	envs := parser.Envs()
@@ -423,6 +441,24 @@ func buildWorkerTemplate(playground *inferenceapi.Playground, backendRuntime *in
 		},
 	}
 
+	// construct /dev/shm size
+	if models[0].Spec.InferenceConfig != nil && models[0].Spec.InferenceConfig.SharedMemorySize != nil {
+		template.Spec.Volumes = append(template.Spec.Volumes, corev1.Volume{
+			Name: "dshm",
+			VolumeSource: corev1.VolumeSource{
+				EmptyDir: &corev1.EmptyDirVolumeSource{
+					Medium:    corev1.StorageMediumMemory,
+					SizeLimit: models[0].Spec.InferenceConfig.SharedMemorySize,
+				},
+			},
+		})
+
+		template.Spec.Containers[0].VolumeMounts = append(template.Spec.Containers[0].VolumeMounts, corev1.VolumeMount{
+			Name:      "dshm",
+			MountPath: "/dev/shm",
+		})
+	}
+
 	return template
 }
 

test/util/mock.go

@@ -37,6 +37,7 @@ func MockASampleModel() *coreapi.OpenModel {
 		InferenceFlavors(
 			*wrapper.MakeFlavor("a100").SetRequest("nvidia.com/gpu", "1").Obj(),
 			*wrapper.MakeFlavor("a10").SetRequest("nvidia.com/gpu", "2").Obj()).
+		SharedMemorySize("1Gi").
 		Obj()
 }
 

test/util/validation/validate_playground.go

@@ -260,6 +260,18 @@ func ValidatePlayground(ctx context.Context, k8sClient client.Client, playground
 				return errors.New("command not right")
 			}
 		}
+
+		if models[0].Spec.InferenceConfig != nil && models[0].Spec.InferenceConfig.SharedMemorySize != nil {
+			if multiHost {
+				if *models[0].Spec.InferenceConfig.SharedMemorySize != *service.Spec.WorkloadTemplate.LeaderWorkerTemplate.LeaderTemplate.Spec.Volumes[0].EmptyDir.SizeLimit {
+					return fmt.Errorf("expected SharedMemorySize %s, got %s", models[0].Spec.InferenceConfig.SharedMemorySize.String(), service.Spec.WorkloadTemplate.LeaderWorkerTemplate.LeaderTemplate.Spec.Volumes[0].EmptyDir.SizeLimit.String())
+				}
+			}
+			if *models[0].Spec.InferenceConfig.SharedMemorySize != *service.Spec.WorkloadTemplate.LeaderWorkerTemplate.WorkerTemplate.Spec.Volumes[0].EmptyDir.SizeLimit {
+				return fmt.Errorf("expected SharedMemorySize %s, got %s", models[0].Spec.InferenceConfig.SharedMemorySize.String(), service.Spec.WorkloadTemplate.LeaderWorkerTemplate.WorkerTemplate.Spec.Volumes[0].EmptyDir.SizeLimit.String())
+			}
+		}
+
 		return nil
 
 	}, util.IntegrationTimeout, util.Interval).Should(gomega.Succeed())

test/util/wrapper/model.go

@@ -107,6 +107,15 @@ func (w *ModelWrapper) Label(k, v string) *ModelWrapper {
 	return w
 }
 
+func (w *ModelWrapper) SharedMemorySize(v string) *ModelWrapper {
+	if w.Spec.InferenceConfig == nil {
+		w.Spec.InferenceConfig = &coreapi.InferenceConfig{}
+	}
+	value := resource.MustParse(v)
+	w.Spec.InferenceConfig.SharedMemorySize = &value
+	return w
+}
+
 func MakeFlavor(name string) *FlavorWrapper {
 	return &FlavorWrapper{
 		coreapi.Flavor{