(helm) Add override files to match Compose

helm/overrides/values-a100-40gb.yaml

@@ -0,0 +1,85 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES.
+# All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+# GPU-specific overrides for A100 40GB (loaded by harness when --deployment-type helm --sku a100-40gb).
+# Sets NIM_TRITON_MAX_BATCH_SIZE=1 per NIM to match docker-compose.a100-40gb.yaml.
+
+nimOperator:
+  page_elements:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+      - name: NIM_TRITON_CPU_THREADS_PRE_PROCESSOR
+        value: "2"
+      - name: OMP_NUM_THREADS
+        value: "2"
+      - name: NIM_TRITON_CPU_THREADS_POST_PROCESSOR
+        value: "1"
+      - name: NIM_ENABLE_OTEL
+        value: "true"
+      - name: NIM_OTEL_SERVICE_NAME
+        value: "page-elements"
+      - name: NIM_OTEL_TRACES_EXPORTER
+        value: "otlp"
+      - name: NIM_OTEL_METRICS_EXPORTER
+        value: "console"
+      - name: NIM_OTEL_EXPORTER_OTLP_ENDPOINT
+        value: "http://otel-collector:4318"
+      - name: TRITON_OTEL_URL
+        value: "http://otel-collector:4318/v1/traces"
+      - name: TRITON_OTEL_RATE
+        value: "1"
+
+  graphic_elements:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_RATE_LIMIT
+        value: "3"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+      - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
+        value: "2048"
+      - name: OMP_NUM_THREADS
+        value: "1"
+
+  table_structure:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_RATE_LIMIT
+        value: "3"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+      - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
+        value: "2048"
+      - name: OMP_NUM_THREADS
+        value: "1"
+
+  nemoretriever_ocr_v1:
+    env:
+      - name: OMP_NUM_THREADS
+        value: "8"
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+
+  llama_3_2_nv_rerankqa_1b_v2:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"

helm/overrides/values-a10g.yaml

@@ -0,0 +1,85 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES.
+# All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+# GPU-specific overrides for A10G (loaded by harness when --deployment-type helm --sku a10g).
+# Sets NIM_TRITON_MAX_BATCH_SIZE=1 per NIM to match docker-compose.a10g.yaml.
+
+nimOperator:
+  page_elements:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+      - name: NIM_TRITON_CPU_THREADS_PRE_PROCESSOR
+        value: "2"
+      - name: OMP_NUM_THREADS
+        value: "2"
+      - name: NIM_TRITON_CPU_THREADS_POST_PROCESSOR
+        value: "1"
+      - name: NIM_ENABLE_OTEL
+        value: "true"
+      - name: NIM_OTEL_SERVICE_NAME
+        value: "page-elements"
+      - name: NIM_OTEL_TRACES_EXPORTER
+        value: "otlp"
+      - name: NIM_OTEL_METRICS_EXPORTER
+        value: "console"
+      - name: NIM_OTEL_EXPORTER_OTLP_ENDPOINT
+        value: "http://otel-collector:4318"
+      - name: TRITON_OTEL_URL
+        value: "http://otel-collector:4318/v1/traces"
+      - name: TRITON_OTEL_RATE
+        value: "1"
+
+  graphic_elements:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_RATE_LIMIT
+        value: "3"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+      - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
+        value: "2048"
+      - name: OMP_NUM_THREADS
+        value: "1"
+
+  table_structure:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_RATE_LIMIT
+        value: "3"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+      - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
+        value: "2048"
+      - name: OMP_NUM_THREADS
+        value: "1"
+
+  nemoretriever_ocr_v1:
+    env:
+      - name: OMP_NUM_THREADS
+        value: "8"
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+
+  llama_3_2_nv_rerankqa_1b_v2:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"

helm/overrides/values-l40s.yaml

@@ -0,0 +1,85 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES.
+# All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+# GPU-specific overrides for L40S (loaded by harness when --deployment-type helm --sku l40s).
+# Sets NIM_TRITON_MAX_BATCH_SIZE=1 per NIM to match docker-compose.l40s.yaml.
+
+nimOperator:
+  page_elements:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+      - name: NIM_TRITON_CPU_THREADS_PRE_PROCESSOR
+        value: "2"
+      - name: OMP_NUM_THREADS
+        value: "2"
+      - name: NIM_TRITON_CPU_THREADS_POST_PROCESSOR
+        value: "1"
+      - name: NIM_ENABLE_OTEL
+        value: "true"
+      - name: NIM_OTEL_SERVICE_NAME
+        value: "page-elements"
+      - name: NIM_OTEL_TRACES_EXPORTER
+        value: "otlp"
+      - name: NIM_OTEL_METRICS_EXPORTER
+        value: "console"
+      - name: NIM_OTEL_EXPORTER_OTLP_ENDPOINT
+        value: "http://otel-collector:4318"
+      - name: TRITON_OTEL_URL
+        value: "http://otel-collector:4318/v1/traces"
+      - name: TRITON_OTEL_RATE
+        value: "1"
+
+  graphic_elements:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_RATE_LIMIT
+        value: "3"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+      - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
+        value: "2048"
+      - name: OMP_NUM_THREADS
+        value: "1"
+
+  table_structure:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_RATE_LIMIT
+        value: "3"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+      - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
+        value: "2048"
+      - name: OMP_NUM_THREADS
+        value: "1"
+
+  nemoretriever_ocr_v1:
+    env:
+      - name: OMP_NUM_THREADS
+        value: "8"
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"
+
+  llama_3_2_nv_rerankqa_1b_v2:
+    env:
+      - name: NIM_HTTP_API_PORT
+        value: "8000"
+      - name: NIM_TRITON_LOG_VERBOSE
+        value: "1"
+      - name: NIM_TRITON_MAX_BATCH_SIZE
+        value: "1"

tools/harness/README.md

@@ -782,7 +782,7 @@ uv run nv-ingest-harness-run --case=e2e --dataset=bo767 --managed --no-dump-logs
 
 ### GPU-Specific Configuration (SKU Override)
 
-The harness supports GPU-specific configuration overrides for Docker Compose deployments via the `--sku` option:
+The harness supports GPU-specific configuration overrides via the `--sku` option for both Docker Compose and Helm deployments:
 
 ```bash
 # A10G GPU settings
@@ -796,15 +796,10 @@ uv run nv-ingest-harness-run --case=e2e --dataset=bo767 --managed --sku=a100-40g
 ```
 
 **How it works:**
-- Loads GPU-specific override file: `docker-compose.<sku>.yaml`
-- Merges with base `docker-compose.yaml` configuration
-- Override settings take precedence (typically batch sizes, memory limits, etc.)
-- Only applies to Docker Compose deployments (ignored for Helm)
-
-**Available SKUs:**
-- `a10g` - NVIDIA A10G GPU settings
-- `l40s` - NVIDIA L40S GPU settings
-- `a100-40gb` - NVIDIA A100 40GB GPU settings
+- **Compose:** Loads `docker-compose.<sku>.yaml` and merges with base `docker-compose.yaml` (override takes precedence).
+- **Helm:** Loads `helm/overrides/values-<sku>.yaml` via `helm upgrade -f`; merged with chart defaults (and any `helm_values_file` / `helm_values` still override).
+
+**Available SKUs:** `a10g` (NVIDIA A10G), `l40s` (NVIDIA L40S), `a100-40gb` (NVIDIA A100 40GB).
 
 ## Nightly Benchmarks
 

tools/harness/plans/SERVICE_MANAGER.md

@@ -45,6 +45,7 @@ tools/harness/src/nv_ingest_harness/service_manager/
    - Installs/upgrades Helm releases with version support
    - Supports both remote charts (from Helm repos) and local `./helm` chart
    - Supports custom values files and inline values
+   - **SKU override**: Optional `helm/overrides/values-<sku>.yaml` via `-f` when `--sku` is set
    - **Port-forwarding**: Automatically sets up resilient `kubectl port-forward` for services
    - **Wildcard matching**: Supports dynamic service name patterns (e.g., `*embed*`)
    - **Auto-restart**: Port-forwards automatically restart on pod restarts/failures
@@ -58,12 +59,11 @@ tools/harness/src/nv_ingest_harness/service_manager/
      - Includes timestamps and handles multi-container pods
 
 4. **create_service_manager()**: Factory function that creates the appropriate manager based on `deployment_type` config
-   - Accepts `sku` parameter for Docker Compose GPU-specific overrides
-   - Passes `sku` to `DockerComposeManager` constructor
+   - Accepts `sku` parameter; passes it to `DockerComposeManager` or `HelmManager` for GPU-specific overrides
 
 ## SKU Override Support
 
-The service manager supports GPU-specific configuration overrides for Docker Compose deployments via the `--sku` CLI option.
+The service manager supports GPU-specific configuration overrides via the `--sku` CLI option for both Docker Compose and Helm deployments.
 
 ### Implementation Details
 
@@ -74,8 +74,7 @@ The service manager supports GPU-specific configuration overrides for Docker Com
 
 #### Service Manager Factory (`service_manager/__init__.py`)
 - `create_service_manager()` accepts `sku` parameter
-- Passes `sku` to `DockerComposeManager` constructor
-- Ignored for Helm deployments (only applies to Docker Compose)
+- Passes `sku` to `DockerComposeManager` (Compose) or `HelmManager` (Helm)
 
 #### Docker Compose Manager (`service_manager/docker_compose.py`)
 - Accepts `sku` parameter in `__init__()`
@@ -84,8 +83,14 @@ The service manager supports GPU-specific configuration overrides for Docker Com
 - Uses `_build_compose_cmd()` helper to construct commands
 - All docker compose commands include override file via multiple `-f` flags
 
+#### Helm Manager (`service_manager/helm.py`)
+- Accepts `sku` parameter in `__init__()`
+- In `start()`, if `sku` is set and `helm/overrides/values-<sku>.yaml` exists, adds `-f <path>` to the `helm upgrade --install` command
+- Prints warning if override file is specified but not found
+
 ### SKU Parameter Flow
 
+**Compose:**
 ```
 CLI (run.py/nightly.py)
   └─> --sku=a10g
@@ -96,10 +101,18 @@ CLI (run.py/nightly.py)
                            └─> Used by: start(), stop(), dump_logs()
 ```
 
+**Helm:**
+```
+CLI (run.py/nightly.py)
+  └─> --sku=a10g
+       └─> create_service_manager(config, repo_root, sku)
+            └─> HelmManager(config, repo_root, sku)
+                 └─> start(): if helm/overrides/values-a10g.yaml exists, add -f <path> to helm upgrade --install
+```
+
 ### Available SKU Override Files
-- `docker-compose.a10g.yaml` - NVIDIA A10G GPU settings
-- `docker-compose.a100-40gb.yaml` - NVIDIA A100 40GB GPU settings
-- `docker-compose.l40s.yaml` - NVIDIA L40S GPU settings
+- **Compose:** `docker-compose.a10g.yaml`, `docker-compose.l40s.yaml`, `docker-compose.a100-40gb.yaml` (repo root)
+- **Helm:** `helm/overrides/values-a10g.yaml`, `helm/overrides/values-l40s.yaml`, `helm/overrides/values-a100-40gb.yaml`
 
 ### Usage Examples
 
@@ -117,6 +130,12 @@ python -m nv_ingest_harness.cli.nightly \
   --deployment-type=compose \
   --managed \
   --sku=l40s
+
+# Helm with A10G GPU override (loads helm/overrides/values-a10g.yaml)
+python -m nv_ingest_harness.cli.run \
+  --dataset=bo767 --case=e2e --managed \
+  --deployment-type=helm \
+  --sku=a10g
 ```
 
 ### Override File Structure
@@ -330,7 +349,7 @@ The service manager is controlled via CLI flags:
 - `--deployment-type=<type>`: Set deployment type (`compose` or `helm`)
   - Overrides `deployment_type` in YAML config
   - Defaults to `compose` if not specified in either place
-- `--sku=<sku>`: GPU-specific override file for Docker Compose (e.g., `a10g`, `l40s`, `a100-40gb`)
+- `--sku=<sku>`: GPU-specific override (Compose: `docker-compose.<sku>.yaml`; Helm: `helm/overrides/values-<sku>.yaml`). e.g. `a10g`, `l40s`, `a100-40gb`
 - `--no-build`: Skip building Docker images (Docker Compose only)
 - `--keep-up`: Keep services running after test completes (does not apply to port-forwards)
 - `--doc-analysis`: Show per-document element breakdown in results

tools/harness/src/nv_ingest_harness/cli/run.py

@@ -354,8 +354,8 @@ def close(self):
     "--sku",
     type=str,
     default=None,
-    help="GPU SKU for Docker Compose override file (e.g., a10g, a100-40gb, l40s). Only applies to managed Compose "
-    "services.",
+    help="GPU SKU for override file (Compose: docker-compose.<sku>.yaml; Helm: helm/overrides/values-<sku>.yaml). "
+    "Applies to managed Compose and Helm deployments (e.g., a10g, a100-40gb, l40s).",
 )
 @click.option(
     "--dump-logs/--no-dump-logs",