Add managed Helm service mode to retriever harness

nemo_retriever/harness/examples/managed-helm-nrl-26.05.yaml

@@ -0,0 +1,78 @@
+# Managed NRL Helm service example for the 26.05 perf sweep.
+#
+# Run from the repository root:
+#   retriever harness run --config nemo_retriever/harness/examples/managed-helm-nrl-26.05.yaml
+#
+# This expects the target namespace to already contain these non-chart-managed secrets:
+#   - ngc-secret: docker registry pull secret for nvcr.io
+#   - ngc-api: opaque secret with the NGC API key entries required by the NIM operator
+# Keep credentials out of this file and out of command history.
+
+# RC6 is the current known-good chart for core NIM coverage. When a new RC is
+# validated, bump the chart version below. If the service image RC differs from
+# the chart package RC, change only x-nrl-service-image-tag.
+x-nrl-chart-version: &nrl_chart_version "26.05-RC6"
+x-nrl-service-image-tag: &nrl_service_image_tag "26.05-RC6"
+
+# This runner class uses A100 MIG resources rather than nvidia.com/gpu.
+# These maps are passed through helm_set as --set-json and replace each NIM's
+# default full-GPU limit. Adjust or remove them for non-MIG clusters.
+x-a100-mig-1g-resource: &a100_mig_1g_resource
+  limits:
+    nvidia.com/mig-1g.10gb: 1
+x-a100-mig-2g-resource: &a100_mig_2g_resource
+  limits:
+    nvidia.com/mig-2g.20gb: 1
+
+active:
+  dataset: bo767
+  preset: dgx_8gpu
+  run_mode: service
+  manage_service: true
+  keep_up: false
+
+  input_type: pdf
+  recall_required: false
+  evaluation_mode: none
+  service_max_concurrency: 8
+
+  helm_bin: microk8s helm
+  kubectl_bin: microk8s kubectl
+  helm_sudo: true
+  kubectl_sudo: true
+
+  helm_chart: nim-nvstaging/nemo-retriever
+  helm_chart_version: *nrl_chart_version
+  helm_release: nv-ingest
+  helm_namespace: nv-ingest
+  helm_values_file: ../helm-profiles/core.yaml
+  helm_timeout: 1800
+  readiness_timeout: 1800
+  helm_service_local_port: 17670
+
+  helm_set:
+    # RC chart defaults still point at localhost:32000/nemo-retriever-service:latest.
+    # Keep these service.image overrides for remote RC chart runs.
+    service.image.repository: nvcr.io/nvstaging/nim/nrl-service
+    service.image.tag: *nrl_service_image_tag
+    service.image.pullPolicy: Always
+    ngcImagePullSecret.create: false
+    ngcImagePullSecret.name: ngc-secret
+    ngcApiSecret.create: false
+    ngcApiSecret.name: ngc-api
+    # Replace the chart's default nvidia.com/gpu NIM requests for A100 MIG nodes.
+    nimOperator.page_elements.resources: *a100_mig_1g_resource
+    nimOperator.table_structure.resources: *a100_mig_1g_resource
+    nimOperator.ocr.resources: *a100_mig_2g_resource
+    nimOperator.vlm_embed.resources: *a100_mig_2g_resource
+
+presets:
+  # Service-mode runs use service_max_concurrency; chart sizing lives in Helm values.
+  dgx_8gpu: {}
+
+datasets:
+  bo767:
+    path: /datasets/nv-ingest/bo767
+    input_type: pdf
+    recall_required: false
+    evaluation_mode: none

nemo_retriever/harness/helm-profiles/all-optional.yaml

@@ -0,0 +1,30 @@
+# Full optional-service coverage profile.
+topology:
+  mode: standalone
+  otel:
+    enabled: true
+
+serviceConfig:
+  vectordb:
+    enabled: true
+
+nims:
+  enabled: true
+
+nimOperator:
+  page_elements:
+    enabled: true
+  table_structure:
+    enabled: true
+  ocr:
+    enabled: true
+  vlm_embed:
+    enabled: true
+  rerankqa:
+    enabled: true
+  nemotron_parse:
+    enabled: true
+  nemotron_3_nano_omni_30b_a3b_reasoning:
+    enabled: true
+  audio:
+    enabled: true

nemo_retriever/harness/helm-profiles/audio-video.yaml

@@ -0,0 +1,35 @@
+# Audio/video managed service profile. The audio NIM is gRPC-only.
+service:
+  installFfmpeg: true
+
+topology:
+  mode: standalone
+  otel:
+    enabled: false
+
+serviceConfig:
+  nimEndpoints:
+    audioGrpcEndpoint: "audio:50051"
+  vectordb:
+    enabled: true
+
+nims:
+  enabled: true
+
+nimOperator:
+  page_elements:
+    enabled: true
+  table_structure:
+    enabled: true
+  ocr:
+    enabled: true
+  vlm_embed:
+    enabled: true
+  rerankqa:
+    enabled: false
+  nemotron_parse:
+    enabled: false
+  nemotron_3_nano_omni_30b_a3b_reasoning:
+    enabled: false
+  audio:
+    enabled: true

nemo_retriever/harness/helm-profiles/core.yaml

@@ -0,0 +1,30 @@
+# Core managed NRL service profile: service, VectorDB, and core NIMs.
+topology:
+  mode: standalone
+  otel:
+    enabled: false
+
+serviceConfig:
+  vectordb:
+    enabled: true
+
+nims:
+  enabled: true
+
+nimOperator:
+  page_elements:
+    enabled: true
+  table_structure:
+    enabled: true
+  ocr:
+    enabled: true
+  vlm_embed:
+    enabled: true
+  rerankqa:
+    enabled: false
+  nemotron_parse:
+    enabled: false
+  nemotron_3_nano_omni_30b_a3b_reasoning:
+    enabled: false
+  audio:
+    enabled: false

nemo_retriever/harness/helm-profiles/no-nims-external.yaml

@@ -0,0 +1,36 @@
+# Service-only profile for externally managed NIM endpoints.
+topology:
+  mode: standalone
+  otel:
+    enabled: false
+
+serviceConfig:
+  nimEndpoints:
+    pageElementsInvokeUrl: "http://page-elements-nim.example.invalid:8000/v1/infer"
+    tableStructureInvokeUrl: "http://table-structure-nim.example.invalid:8000/v1/infer"
+    ocrInvokeUrl: "http://ocr-nim.example.invalid:8000/v1/infer"
+    embedInvokeUrl: "http://embed-nim.example.invalid:8000/v1/embeddings"
+    audioGrpcEndpoint: "audio-nim.example.invalid:50051"
+  vectordb:
+    enabled: true
+
+nims:
+  enabled: false
+
+nimOperator:
+  page_elements:
+    enabled: false
+  table_structure:
+    enabled: false
+  ocr:
+    enabled: false
+  vlm_embed:
+    enabled: false
+  rerankqa:
+    enabled: false
+  nemotron_parse:
+    enabled: false
+  nemotron_3_nano_omni_30b_a3b_reasoning:
+    enabled: false
+  audio:
+    enabled: false

nemo_retriever/harness/helm-profiles/split.yaml

@@ -0,0 +1,52 @@
+# Split-topology profile for gateway/realtime/batch readiness coverage.
+topology:
+  mode: split
+  otel:
+    enabled: true
+  realtime:
+    hpa:
+      metrics:
+        cpu:
+          enabled: true
+        queueDepthRatio:
+          enabled: false
+        processingLatencyP95:
+          enabled: false
+  batch:
+    hpa:
+      metrics:
+        cpu:
+          enabled: true
+        queueDepthRatio:
+          enabled: false
+        processingLatencyP95:
+          enabled: false
+
+autoscaling:
+  queueDepth:
+    backend: cpu
+
+serviceConfig:
+  vectordb:
+    enabled: true
+
+nims:
+  enabled: true
+
+nimOperator:
+  page_elements:
+    enabled: true
+  table_structure:
+    enabled: true
+  ocr:
+    enabled: true
+  vlm_embed:
+    enabled: true
+  rerankqa:
+    enabled: false
+  nemotron_parse:
+    enabled: false
+  nemotron_3_nano_omni_30b_a3b_reasoning:
+    enabled: false
+  audio:
+    enabled: false