Support for slim-api, ctsm model & fixes

api/v1/aiplatform_types.go

@@ -177,8 +177,8 @@ type VectorDBStorageSpec struct {
 
 // FeatureSpec defines the features to enable in the AIPlatform
 type FeatureSpec struct {
-	// Name of the feature, e.g. "saia" or "seca"
-	// +kubebuilder:validation:Enum=saia;seca
+	// Name of the feature, e.g. "saia", "seca" or "slim"
+	// +kubebuilder:validation:Enum=saia;seca;slim
 	Name string `json:"name,omitempty"`
 	// ServiceAccountName is the name of the service account to use for the feature
 	ServiceAccountName string `json:"serviceAccountName,omitempty"`

config/configs/applications.yaml

@@ -34,6 +34,7 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- if .Replicas.Gemma431bIt }}
   - args:
       application_name: Gemma431bIt
       deployment_configs:
@@ -170,6 +171,8 @@ applications:
         # Airgap k0s has no Redis; cloud sets this to "False" and wires
         # RESPONSES_REDIS_ADDRESS to its in-namespace Redis StatefulSet.
         DISABLE_RESPONSES_API_REDIS: "True"
+{{- end }}
+{{- if .Replicas.GptOss20b }}
   - args:
       application_name: GptOss20b
       deployment_configs:
@@ -274,6 +277,8 @@ applications:
         # See Gemma431bIt above for rationale. Must be "True" in airgap (no
         # Redis) so vLLM uses NoOpOpenAIServingResponses.
         DISABLE_RESPONSES_API_REDIS: "True"
+{{- end }}
+{{- if .Replicas.UaeLarge }}
   - args:
       application_name: UaeLarge
       deployment_configs:
@@ -352,6 +357,8 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- end }}
+{{- if .Replicas.AllMinilmL6V2 }}
   - args:
       application_name: AllMinilmL6V2
       deployment_configs:
@@ -424,6 +431,8 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- end }}
+{{- if .Replicas.BiEncoder }}
   - args:
       application_name: BiEncoder
       deployment_configs:
@@ -493,6 +502,8 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- end }}
+{{- if .Replicas.MbartTranslator }}
   - args:
       application_name: MbartTranslator
       custom_deployment_import_path: mbart_translator:MbartTranslatorDeployment
@@ -553,6 +564,8 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- end }}
+{{- if .Replicas.XlmRobertaLanguageClassifier }}
   - args:
       application_name: XlmRobertaLanguageClassifier
       deployment_configs:
@@ -631,6 +644,7 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- end }}
   - args:
       application_name: PromptInjectionTfidf
       custom_deployment_import_path: prompt_injection_tfidf:PromptInjectionTfidfDeployment
@@ -670,6 +684,84 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- if .Replicas.FmTimeseries }}
+  # FmTimeseries is a slim-feature model (see config/configs/features/slim.yaml).
+  # .Replicas.FmTimeseries is only populated when the "slim" feature is enabled,
+  # so this whole application is rendered only then — with saia alone it would
+  # otherwise emit "<no value>" for the replica counts and break the serve config.
+  - args:
+      application_name: FmTimeseries
+      # Cisco Time Series Model (CTSM) — a 250M-param decoder-only time-series
+      # forecaster served as a custom PyTorch deployment (NOT vLLM), so it has
+      # no engine_args / tensor_parallel_size / gpu_memory_utilization. It shares
+      # a GPU (num_gpus: 0.1). Weights are staged from HuggingFace
+      # (cisco-ai/cisco-time-series-model-1.0) to model_artifacts/fm_timeseries;
+      # CHECKPOINT_FILE names the checkpoint inside that prefix.
+      custom_deployment_import_path: fm_timeseries.main:FMTimeseriesDeployment
+      deployment_configs:
+        FMTimeseriesDeployment:
+          options:
+            autoscaling_config:
+              max_replicas: {{.Replicas.FmTimeseries}}
+              min_replicas: {{.Replicas.FmTimeseries}}
+            ray_actor_options:
+              num_gpus: 0.1
+      deployment_type: custom_deployment
+      model_definition:
+        model_config:
+          num_layers: 25
+        model_id: fm_timeseries
+        model_loader:
+          blob_storage:
+            # Only the checkpoint is needed — CiscoTsmMR reads its model config
+            # from inside torch_model.pt (loaded_checkpoint["config"]), not from
+            # config.json. artifacts_list restricts the download to exactly these
+            # files under blob_prefix, so this must name a file that exists in
+            # the bucket (verified: model_artifacts/fm_timeseries/torch_model.pt).
+            artifacts_list:
+              - torch_model.pt
+            blob_prefix: model_artifacts/fm_timeseries
+    name: FmTimeseries
+    import_path: main:create_serve_app
+    route_prefix: /fm_timeseries
+    runtime_env:
+      working_dir: "file:///home/ray/ray/applications/generic_application.zip"
+      env_vars:
+        API_VERSION: "v1"
+        APPLICATION_NAME: fm_timeseries
+        # Checkpoint filename inside blob_prefix. The public HF repo ships
+        # torch_model.pt (the deployment builds the path as
+        # <local_dir>/CHECKPOINT_FILE), so this must match a staged artifact.
+        CHECKPOINT_FILE: torch_model.pt
+        ARTIFACTS_S3_BUCKET: "{{.ArtifactBucketName}}"
+        S3_BUCKET: "{{.ArtifactBucketName}}"
+        ARTIFACTS_PROVIDER: "{{.ArtifactsProvider}}"
+        CLOUD_PROVIDER: "{{.CloudProvider}}"
+        S3COMPAT_OBJECT_STORE_ENDPOINT_URL: "{{.S3CompatObjectStoreEndpointUrl}}"
+        S3COMPAT_OBJECT_STORE_ACCESS_KEY: "{{.S3CompatObjectStoreAccessKey}}"
+        S3COMPAT_OBJECT_STORE_SECRET_KEY: "{{.S3CompatObjectStoreSecretKey}}"
+        # AWS / boto3 standard credential names — populated whenever the
+        # operator can load credentials from spec.objectStorage.secretRef. For
+        # CLOUD_PROVIDER=aws these are the values boto3 reads (the S3COMPAT_*
+        # names above are only consumed by the s3compat shim). Both code paths
+        # share the same source-of-truth Secret keys (s3_access_key /
+        # s3_secret_key) so emitting both pairs is safe — each provider only
+        # reads its own. AWS_REGION lets boto3 resolve the default regional S3
+        # endpoint when no AWS_ENDPOINT_URL is set; required for any AWS S3
+        # bucket outside us-east-1 to avoid PermanentRedirect on the first call.
+        AWS_ACCESS_KEY_ID: "{{.S3CompatObjectStoreAccessKey}}"
+        AWS_SECRET_ACCESS_KEY: "{{.S3CompatObjectStoreSecretKey}}"
+        AWS_REGION: "{{.Region}}"
+        AWS_DEFAULT_REGION: "{{.Region}}"
+        ENABLE_AUTHN: "false"
+        ENABLE_AUTHZ: "false"
+        SERVICE_EXTERNAL_NAME: "ai-platform-models"
+        SERVICE_INTERNAL_NAME: "ai_platform_models"
+        SERVICE_NAME: "ai_platform_models"
+        SKIP_VERIFICATION: "true"
+        USE_SYSTEM_PERMISSIONS: "true"
+{{- end }}
+{{- if .Replicas.CrossEncoder }}
   - args:
       application_name: CrossEncoder
       deployment_configs:
@@ -743,6 +835,8 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- end }}
+{{- if .Replicas.E5LanguageClassifier }}
   - args:
       application_name: E5LanguageClassifier
       deployment_configs:
@@ -821,6 +915,8 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- end }}
+{{- if .Replicas.PromptInjectionCrossEncoder }}
   - args:
       application_name: PromptInjectionCrossEncoder
       deployment_configs:
@@ -886,6 +982,7 @@ applications:
         SERVICE_NAME: "ai_platform_models"
         SKIP_VERIFICATION: "true"
         USE_SYSTEM_PERMISSIONS: "true"
+{{- end }}
   - args:
       application_name: PromptInjectionClassifier
       deployment_type: classification_model_deployment

config/configs/features/slim.yaml

@@ -0,0 +1,11 @@
+applicationScale:
+  FmTimeseries: 1
+instanceScale:
+  L40S:
+    l40s-1-gpu: 1
+  H100:
+    h100-1-gpu: 1
+  H100_NVL:
+    h100-nvl-1-gpu: 1
+  RTX_PRO_6000_BLACKWELL:
+    rtx-pro-6000-blackwell-1-gpu: 1

config/crd/bases/ai.splunk.com_aiplatforms.yaml

@@ -1054,10 +1054,11 @@ spec:
                   description: FeatureSpec defines the features to enable in the AIPlatform
                   properties:
                     name:
-                      description: Name of the feature, e.g. "saia" or "seca"
+                      description: Name of the feature, e.g. "saia", "seca" or "slim"
                       enum:
                       - saia
                       - seca
+                      - slim
                       type: string
                     scaleFactor:
                       description: ScaleFactor is the desired fixed number of replicas

config/crd/bases/ai.splunk.com_aiservices.yaml

@@ -1045,10 +1045,11 @@ spec:
                 description: Feature defines the features to be enabled for the AIService
                 properties:
                   name:
-                    description: Name of the feature, e.g. "saia" or "seca"
+                    description: Name of the feature, e.g. "saia", "seca" or "slim"
                     enum:
                     - saia
                     - seca
+                    - slim
                     type: string
                   scaleFactor:
                     description: ScaleFactor is the desired fixed number of replicas

config/manager/kustomization.yaml

@@ -13,6 +13,7 @@ patches:
     \ - name: RELATED_IMAGE_RAY_WORKER\n    value: \"667741767953.dkr.ecr.us-west-2.amazonaws.com/ml-platform/ray/ray-worker-gpu:build-17\"\n
     \ - name: RELATED_IMAGE_WEAVIATE\n    value: \"semitechnologies/weaviate:stable-v1.28-007846a\"\n
     \ - name: RELATED_IMAGE_SAIA_API\n    value: \"667741767953.dkr.ecr.us-west-2.amazonaws.com/ml-platform/saia/saia-api:build-1\"\n
+    \ - name: RELATED_IMAGE_SLIM_API\n    value: \"667741767953.dkr.ecr.us-west-2.amazonaws.com/ml-platform/slim/slim-api:build-1\"\n
     \ - name: RELATED_IMAGE_POST_INSTALL_HOOK\n    value: \"667741767953.dkr.ecr.us-west-2.amazonaws.com/ml-platform/saia/saia-data-loader:build-1\"\n
     \ - name: RELATED_IMAGE_FLUENT_BIT\n    value: \"fluent/fluent-bit:1.9.6\"\n  -
     name: RELATED_IMAGE_OTEL_COLLECTOR\n    value: \"otel/opentelemetry-collector-contrib:0.122.1\"\n

helm-chart/splunk-ai-operator/crds/ai.splunk.com_aiplatforms.yaml

@@ -1054,10 +1054,11 @@ spec:
                   description: FeatureSpec defines the features to enable in the AIPlatform
                   properties:
                     name:
-                      description: Name of the feature, e.g. "saia" or "seca"
+                      description: Name of the feature, e.g. "saia", "seca" or "slim"
                       enum:
                       - saia
                       - seca
+                      - slim
                       type: string
                     scaleFactor:
                       description: ScaleFactor is the desired fixed number of replicas

helm-chart/splunk-ai-operator/crds/ai.splunk.com_aiservices.yaml

@@ -1019,6 +1019,14 @@ spec:
                     type: string
                 type: object
                 x-kubernetes-map-type: atomic
+              aiPlatformScheme:
+                default: http
+                description: AIPlatformScheme specifies the URL scheme for the AI
+                  Platform service ("http" or "https")
+                enum:
+                - http
+                - https
+                type: string
               aiPlatformUrl:
                 description: AIPlatformUrl specifies the URL for the AI Platform (deprecated,
                   use AIPlatformRef)
@@ -1037,10 +1045,11 @@ spec:
                 description: Feature defines the features to be enabled for the AIService
                 properties:
                   name:
-                    description: Name of the feature, e.g. "saia" or "seca"
+                    description: Name of the feature, e.g. "saia", "seca" or "slim"
                     enum:
                     - saia
                     - seca
+                    - slim
                     type: string
                   scaleFactor:
                     description: ScaleFactor is the desired fixed number of replicas

helm-chart/splunk-ai-operator/templates/deployment.yaml

@@ -78,6 +78,8 @@ spec:
             value: {{ .Values.saiaApiImage }}
           - name: RELATED_IMAGE_SAIA_API_V2
             value: {{ .Values.saiaApiV2Image }}
+          - name: RELATED_IMAGE_SLIM_API
+            value: {{ .Values.slimApiImage }}
           - name: RELATED_IMAGE_POST_INSTALL_HOOK
             value: {{ .Values.saiaSchemaImage }}
           - name: RELATED_IMAGE_OTEL_COLLECTOR

helm-chart/splunk-ai-operator/values.yaml

@@ -110,6 +110,9 @@ saiaApiImage: "docker.io/splunk/saia-api:1.1.0"
 saiaApiV2Image: "docker.io/splunk/saia-api-v2:1.1.0"
 saiaSchemaImage: "docker.io/splunk/saia-data-loader:1.1.0"
 
+# SLIM API image
+slimApiImage: "docker.io/splunk/slim-api:1.0.0"
+
 # OpenTelemetry Collector sidecar image
 otelCollectorImage: "docker.io/otel/opentelemetry-collector-contrib:0.122.1"
 

helm-chart/splunk-ai-platform/values.yaml

@@ -56,7 +56,7 @@ serviceAccountName: ""
 gpuInstanceType: "g6.24xlarge"
 
 # List of defined features to enable in the AI platform
-# Options include: saia, seca
+# Options include: saia, seca, slim
 features: []
 # Example:
 # - name: "saia"

pkg/ai/features/registry.go

@@ -4,9 +4,11 @@ import (
 	"github.com/splunk/splunk-ai-operator/pkg/ai/features/common"
 	"github.com/splunk/splunk-ai-operator/pkg/ai/features/saia"
 	"github.com/splunk/splunk-ai-operator/pkg/ai/features/seca"
+	"github.com/splunk/splunk-ai-operator/pkg/ai/features/slim"
 )
 
 var FeatureFactories = map[string]common.FeatureFactory{
 	"saia": &saia.SaiaFactory{},
 	"seca": &seca.SecaFactory{},
+	"slim": &slim.SlimFactory{},
 }

pkg/ai/features/slim/factory.go

@@ -0,0 +1,21 @@
+package slim
+
+import (
+	"context"
+
+	aiv1 "github.com/splunk/splunk-ai-operator/api/v1"
+	"github.com/splunk/splunk-ai-operator/pkg/ai/features/common"
+	"k8s.io/apimachinery/pkg/runtime"
+	"k8s.io/client-go/tools/record"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+)
+
+type SlimFactory struct{}
+
+func (f *SlimFactory) New(ctx context.Context, c client.Client, scheme *runtime.Scheme, ai *aiv1.AIService, recorder record.EventRecorder) (common.FeatureHandler, error) {
+	return &SlimReconciler{
+		Client:   c,
+		Scheme:   scheme,
+		Recorder: recorder,
+	}, nil
+}

pkg/ai/features/slim/factory_test.go

@@ -0,0 +1,44 @@
+package slim
+
+import (
+	"context"
+	"testing"
+
+	aiv1 "github.com/splunk/splunk-ai-operator/api/v1"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/runtime"
+	"k8s.io/client-go/kubernetes/scheme"
+	"k8s.io/client-go/tools/record"
+	"sigs.k8s.io/controller-runtime/pkg/client/fake"
+)
+
+func TestSlimFactory_New(t *testing.T) {
+	s := runtime.NewScheme()
+	_ = scheme.AddToScheme(s)
+	_ = aiv1.AddToScheme(s)
+
+	factory := &SlimFactory{}
+	fakeClient := fake.NewClientBuilder().WithScheme(s).Build()
+	recorder := record.NewFakeRecorder(10)
+
+	aiService := &aiv1.AIService{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      "test-service",
+			Namespace: "default",
+		},
+	}
+
+	handler, err := factory.New(context.Background(), fakeClient, s, aiService, recorder)
+
+	require.NoError(t, err)
+	require.NotNil(t, handler)
+
+	// Verify it returns a SlimReconciler
+	reconciler, ok := handler.(*SlimReconciler)
+	assert.True(t, ok, "Expected handler to be *SlimReconciler")
+	assert.NotNil(t, reconciler.Client)
+	assert.NotNil(t, reconciler.Scheme)
+	assert.NotNil(t, reconciler.Recorder)
+}