feat(gpu): add NVIDIA GPU power monitoring via NVML

  Add experimental GPU power monitoring support:
  - NVIDIA backend using go-nvml for device discovery and power readings
  - GPU sharing mode detection (time-slicing, exclusive, partitioned)
  - Per-process power attribution based on compute utilization
  - experimental.gpu.enabled config flag (disabled by default)
  - added tests

  Sharing mode detection:
  - Partitioned (MIG): GPU partitioned into isolated instances
  - Exclusive: Single process has exclusive GPU access (nvidia-smi -c EXCLUSIVE_PROCESS)
  - Time-slicing: Multiple processes share GPU (default mode)

  Detection logic:
  1. Check if MIG enabled → Partitioned
  2. Check NVML compute mode:
     - EXCLUSIVE_PROCESS/THREAD → Exclusive
     - DEFAULT → Time-slicing

  Includes:
  - nvidia/collector.go: GPUPowerMeter implementation
  - nvidia/detector.go: sharing mode detection
  - nvidia/nvml.go: NVML wrapper with MIG support
  - nvidia/types.go: NVIDIA-specific ComputeMode type
  - Dockerfile: enable CGO for NVML
  - Manifests: mount nvidia driver libs, set LD_LIBRARY_PATH

Signed-off-by: Vimal Kumar <vimal78@gmail.com>

Author: vimalk78

Dockerfile

@@ -11,6 +11,7 @@ WORKDIR /workspace
 COPY . .
 
 RUN make build \
+  CGO_ENABLED=1 \
   PRODUCTION=1 \
   VERSION=${VERSION} \
   GIT_COMMIT=${GIT_COMMIT} \

cmd/kepler/main.go

@@ -14,6 +14,7 @@ import (
 
 	"github.com/sustainable-computing-io/kepler/config"
 	"github.com/sustainable-computing-io/kepler/internal/device"
+	"github.com/sustainable-computing-io/kepler/internal/device/gpu"
 	"github.com/sustainable-computing-io/kepler/internal/exporter/prometheus"
 	"github.com/sustainable-computing-io/kepler/internal/exporter/stdout"
 	"github.com/sustainable-computing-io/kepler/internal/k8s/pod"
@@ -24,6 +25,9 @@ import (
 	"github.com/sustainable-computing-io/kepler/internal/server"
 	"github.com/sustainable-computing-io/kepler/internal/service"
 	"github.com/sustainable-computing-io/kepler/internal/version"
+
+	// Register GPU backends via init()
+	_ "github.com/sustainable-computing-io/kepler/internal/device/gpu/nvidia"
 )
 
 func main() {
@@ -187,6 +191,22 @@ func createServices(logger *slog.Logger, cfg *config.Config) ([]service.Service,
 		}
 	}
 
+	// Discover GPU devices if enabled (experimental feature)
+	if cfg.IsFeatureEnabled(config.ExperimentalGPUFeature) {
+		gpuMeters := gpu.DiscoverAll(logger)
+		if len(gpuMeters) > 0 {
+			logger.Info("GPU power monitoring enabled",
+				"vendors", len(gpuMeters),
+				"registered_vendors", gpu.RegisteredVendors())
+			// GPU meters are initialized by DiscoverAll, store for future use
+			// TODO: integrate with monitor for per-process power attribution
+			_ = gpuMeters
+		} else {
+			logger.Warn("GPU power monitoring enabled but no GPUs discovered",
+				"registered_vendors", gpu.RegisteredVendors())
+		}
+	}
+
 	// Add Prometheus exporter if enabled
 	if cfg.IsFeatureEnabled(config.PrometheusFeature) {
 		promExporter, err := createPrometheusExporter(logger, cfg, apiServer, pm, redfishService)

codecov.yml

@@ -0,0 +1,14 @@
+# SPDX-FileCopyrightText: 2025 The Kepler Authors
+# SPDX-License-Identifier: Apache-2.0
+
+coverage:
+  status:
+    project:
+      default:
+        threshold: 1%
+    patch:
+      default:
+        target: 70%
+
+ignore:
+  - internal/device/gpu/nvidia/nvml_lib.go

compose/default/kepler/etc/kepler/config.yaml

@@ -83,3 +83,5 @@ experimental:
   hwmon:
     enabled: false # Enable experimental hwmon power monitoring
     zones: [] # List of zones to enable (default enable all)
+  gpu:
+    enabled: false # Enable experimental GPU power monitoring

compose/dev/kepler-dev/etc/kepler/config.yaml

@@ -83,3 +83,5 @@ experimental:
   hwmon:
     enabled: false # Enable experimental hwmon power monitoring
     zones: [] # List of zones to enable (default enable all)
+  gpu:
+    enabled: false # Enable experimental GPU power monitoring

config/config.go

@@ -255,6 +255,9 @@ const (
 	ExperimentalHwmonEnabledFlag = "experimental.hwmon.enabled"
 	ExperimentalHwmonZonesFlag   = "experimental.hwmon.zones"
 
+	// Experimental GPU flags
+	ExperimentalGPUEnabledFlag = "experimental.gpu.enabled"
+
 // WARN:  dev settings shouldn't be exposed as flags as flags are intended for end users
 )
 
@@ -409,6 +412,9 @@ func RegisterFlags(app *kingpin.Application) ConfigUpdaterFn {
 	hwmonEnabled := app.Flag(ExperimentalHwmonEnabledFlag, "Enable experimental hwmon power monitoring").Default("false").Bool()
 	hwmonZones := app.Flag(ExperimentalHwmonZonesFlag, "Hwmon zone filter (power labels to monitor)").Strings()
 
+	// experimental GPU
+	gpuEnabled := app.Flag(ExperimentalGPUEnabledFlag, "Enable experimental GPU power monitoring").Default("false").Bool()
+
 	return func(cfg *Config) error {
 		// Logging settings
 		if flagsSet[LogLevelFlag] {
@@ -481,6 +487,9 @@ func RegisterFlags(app *kingpin.Application) ConfigUpdaterFn {
 			return err
 		}
 
+		// Apply experimental GPU settings
+		applyGPUConfig(cfg, flagsSet, gpuEnabled)
+
 		cfg.sanitize()
 		return cfg.Validate()
 	}
@@ -604,6 +613,28 @@ func applyHwmonFlags(hwmon *Hwmon, flagsSet map[string]bool, enabled *bool, zone
 	}
 }
 
+// applyGPUConfig applies GPU configuration from flags
+func applyGPUConfig(cfg *Config, flagsSet map[string]bool, enabled *bool) {
+	// Early exit if no GPU flags are set and config file does not have experimental section
+	if !hasGPUFlags(flagsSet) && cfg.Experimental == nil {
+		return
+	}
+
+	// Initialize experimental section if needed
+	if cfg.Experimental == nil {
+		cfg.Experimental = &Experimental{}
+	}
+
+	if flagsSet[ExperimentalGPUEnabledFlag] {
+		cfg.Experimental.GPU.Enabled = enabled
+	}
+}
+
+// hasGPUFlags returns true if any GPU experimental flags are set
+func hasGPUFlags(flagsSet map[string]bool) bool {
+	return flagsSet[ExperimentalGPUEnabledFlag]
+}
+
 // resolveNodeName resolves the node name using the following precedence:
 // 1. CLI flag / config.yaml (--experimental.platform.redfish.node-name)
 // 2. Kubernetes node name

docs/user/configuration.md

@@ -38,6 +38,7 @@ You can configure Kepler by passing flags when starting the service. The followi
 | `--experimental.platform.redfish.config-file` | Path to experimental Redfish BMC configuration file                     | `""`                            | Any valid file path                                                |
 | `--experimental.hwmon.enabled`                | Enable experimental hwmon power monitoring                              | `false`                         | `true`, `false`                                                    |
 | `--experimental.hwmon.zones`                  | hwmon zones to be enabled (can be specified multiple times)             | All available zones             | Any valid hwmon zone name                                          |
+| `--experimental.gpu.enabled`                  | Enable experimental GPU power monitoring                                | `false`                         | `true`, `false`                                                    |
 
 ### 💡 Examples
 
@@ -73,6 +74,9 @@ kepler --experimental.hwmon.enabled=true \
        --experimental.hwmon.zones=power1 \
        --experimental.hwmon.zones=power2
 
+# Enable experimental GPU power monitoring
+kepler --experimental.gpu.enabled=true
+
 # Export only node and container level metrics
 kepler --metrics=node --metrics=container
 
@@ -153,6 +157,8 @@ experimental:   # experimental features (no stability guarantees)
   hwmon:        # hwmon power monitoring
     enabled: false                    # Enable hwmon power monitoring (default: false)
     zones: []                         # hwmon zones to be enabled, empty enables all available zones
+  gpu:          # GPU power monitoring
+    enabled: false                    # Enable GPU power monitoring (default: false)
 
 # WARN: DO NOT ENABLE THIS IN PRODUCTION - for development/testing only
 dev:
@@ -328,6 +334,8 @@ experimental:
   hwmon:
     enabled: false
     zones: []
+  gpu:
+    enabled: false
 ```
 
 ⚠️ **WARNING**: This section contains experimental features with no stability guarantees.
@@ -394,6 +402,21 @@ experimental:
     zones: ["power1", "power2"]
 ```
 
+#### GPU Power Monitoring
+
+- **enabled**: Enable experimental GPU power monitoring (default: false)
+  - When enabled, Kepler will collect power metrics from NVIDIA GPUs using NVML
+  - Requires NVIDIA drivers and NVML library to be available
+  - Supports per-process power attribution based on GPU compute utilization
+
+**Example:**
+
+```yaml
+experimental:
+  gpu:
+    enabled: true
+```
+
 ### 🧑‍🔬 Development Configuration
 
 ```yaml

go.mod

@@ -6,6 +6,7 @@ toolchain go1.24.9
 
 require (
 	dario.cat/mergo v1.0.2
+	github.com/NVIDIA/go-nvml v0.13.0-1
 	github.com/alecthomas/kingpin/v2 v2.4.0
 	github.com/go-logr/logr v1.4.2
 	github.com/oklog/run v1.1.0

go.sum

@@ -1,5 +1,7 @@
 dario.cat/mergo v1.0.2 h1:85+piFYR1tMbRrLcDwR18y4UKJ3aH1Tbzi24VRW1TK8=
 dario.cat/mergo v1.0.2/go.mod h1:E/hbnu0NxMFBjpMIE34DRGLWqDy0g5FuKDhCb31ngxA=
+github.com/NVIDIA/go-nvml v0.13.0-1 h1:OLX8Jq3dONuPOQPC7rndB6+iDmDakw0XTYgzMxObkEw=
+github.com/NVIDIA/go-nvml v0.13.0-1/go.mod h1:+KNA7c7gIBH7SKSJ1ntlwkfN80zdx8ovl4hrK3LmPt4=
 github.com/alecthomas/kingpin/v2 v2.4.0 h1:f48lwail6p8zpO1bC4TxtqACaGqHYA22qkHjHpqDjYY=
 github.com/alecthomas/kingpin/v2 v2.4.0/go.mod h1:0gyi0zQnjuFk8xrkNKamJoyUo382HRL7ATRpFZCw6tE=
 github.com/alecthomas/units v0.0.0-20211218093645-b94a6e3cc137 h1:s6gZFSlWYmbqAuRjVTiNNhvNRfY2Wxp9nhfyel4rklc=

hack/config.yaml

@@ -83,3 +83,5 @@ experimental:
   hwmon:
     enabled: false # Enable experimental hwmon power monitoring
     zones: [] # List of zones to enable (default enable all)
+  gpu:
+    enabled: false # Enable experimental GPU power monitoring