Replace custom GPU filter/scorer with generic endpoint-attribute plugins

The DCGM extractor now emits ScalarMetricValue (instead of a custom
GPUUtilization type), making it compatible with the generic
endpoint-attribute-filter and endpoint-attribute-scorer already in main.

Signed-off-by: noalimoy <nlimoy@redhat.com>

Author: noalimoy

cmd/epp/runner/runner.go

@@ -108,7 +108,6 @@ import (
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/requesthandling/parsers/vllmhttp"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/filter/bylabel"
 	endpointattributefilter "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/filter/endpointattribute"
-	filtergpuutil "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/filter/gpuutilization"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/filter/prefixcacheaffinity"
 	sessionaffinityfilter "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/filter/sessionaffinity"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/filter/sloheadroomtier"
@@ -121,7 +120,6 @@ import (
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/scorer/activerequest"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/scorer/contextlengthaware"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/scorer/endpointattribute"
-	scorergpuutil "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/scorer/gpuutilization"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/scorer/kvcacheutilization"
 	latencyscorer "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/scorer/latency"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/scheduling/scorer/loadaware"
@@ -577,10 +575,6 @@ func (r *Runner) registerInTreePlugins() {
 	fwkplugin.Register(srcdcgm.DCGMDataSourceType, srcdcgm.DCGMDataSourceFactory)
 	fwkplugin.Register(attrgpu.DCGMExtractorType, extdcgm.DCGMExtractorFactory)
 
-	// GPU utilization filter/scorer
-	fwkplugin.Register(filtergpuutil.PluginType, filtergpuutil.Factory)
-	fwkplugin.Register(scorergpuutil.PluginType, scorergpuutil.Factory)
-
 	fwkplugin.Register(prefix.PrefixCacheScorerPluginType, prefix.PrefixCachePluginFactory)
 	fwkplugin.Register(maxscore.MaxScorePickerType, maxscore.MaxScorePickerFactory)
 	fwkplugin.Register(random.RandomPickerType, random.RandomPickerFactory)

deploy/config/sim-epp-gpu-config.yaml

@@ -9,10 +9,27 @@ plugins:
     port: 9400
 - type: dcgm-extractor
   name: dcgm-extractor
-- type: gpu-utilization-filter
+- type: endpoint-attribute-filter
+  name: gpu-utilization-filter
   parameters:
-    threshold: 0.90
-- type: gpu-utilization-scorer
+    attribute: "GPUUtilization/dcgm-extractor"
+    onMissing: "Pass"
+    fallbackOnEmpty: true
+    algorithm:
+      type: "threshold"
+      threshold:
+        operator: "LessThanOrEqual"
+        value: 0.90
+- type: endpoint-attribute-scorer
+  name: gpu-utilization-scorer
+  parameters:
+    attributeKey: "GPUUtilization/dcgm-extractor"
+    algorithm:
+      type: "linear_lower_is_better"
+      normalization:
+        fixedRange:
+          min: 0.0
+          max: 1.0
 - type: kv-cache-utilization-scorer
 - type: max-score-picker
 - type: single-profile-handler

pkg/epp/framework/plugins/datalayer/attribute/gpu/README.md

@@ -1,15 +1,11 @@
 # GPU Attributes
 
-This package defines the data structures for GPU hardware metrics collected from NVIDIA DCGM Exporter.
+This package defines the data key for GPU hardware metrics collected from NVIDIA DCGM Exporter.
 
-## `GPUUtilization`
+## `GPUUtilizationDataKey`
 
-A normalized GPU compute utilization value in [0.0, 1.0], derived from `DCGM_FI_DEV_GPU_UTIL` (which reports 0-100). For multi-GPU pods the extractor aggregates across visible devices using `max`.
+- **Key string**: `GPUUtilization/dcgm-extractor`
+- **Value type**: `ScalarMetricValue` (from `attribute/metrics`)
 
-- **Key**: `GPUUtilizationDataKey` (`GPUUtilization/dcgm-extractor`)
-
-## Producers
-
-The following plugins produce this attribute:
-
-- **`dcgm-extractor`** (Data Layer): Extracts GPU utilization from the DCGM Exporter Prometheus endpoint.
+For usage instructions, cluster prerequisites, and EPP config examples see
+[`extractor/dcgm/README.md`](../../../extractor/dcgm/README.md).

pkg/epp/framework/plugins/datalayer/attribute/gpu/data_types.go

@@ -17,7 +17,6 @@ limitations under the License.
 package gpu
 
 import (
-	fwkdl "github.com/llm-d/llm-d-router/pkg/epp/framework/interface/datalayer"
 	"github.com/llm-d/llm-d-router/pkg/epp/framework/interface/plugin"
 )
 
@@ -27,19 +26,6 @@ const (
 )
 
 // GPUUtilizationDataKey identifies per-endpoint GPU utilization published
-// by the DCGM extractor and consumed by GPU-aware filters and scorers.
+// by the DCGM extractor and consumed by the generic endpoint-attribute
+// filter and scorer.
 var GPUUtilizationDataKey = plugin.NewDataKey("GPUUtilization", DCGMExtractorType)
-
-// GPUUtilization is a normalized GPU compute utilization value in [0.0, 1.0],
-// derived from DCGM_FI_DEV_GPU_UTIL (which reports 0-100).
-// For multi-GPU pods the extractor aggregates across visible devices.
-type GPUUtilization float64
-
-func (v GPUUtilization) Clone() fwkdl.Cloneable {
-	return v
-}
-
-// ReadGPUUtilization retrieves GPU utilization from an endpoint's AttributeMap.
-func ReadGPUUtilization(attrs fwkdl.AttributeMap, key string) (GPUUtilization, bool) {
-	return fwkdl.ReadAttribute[GPUUtilization](attrs, key)
-}

pkg/epp/framework/plugins/datalayer/extractor/dcgm/README.md

@@ -4,7 +4,8 @@
 
 The DCGM Extractor converts the Prometheus metrics response from a
 `dcgm-data-source` into a per-endpoint GPU utilization attribute consumed by
-GPU-aware filters and scorers.
+the generic `endpoint-attribute-filter` and `endpoint-attribute-scorer`
+(configurable for any scalar attribute, not hardwired to GPU).
 
 ## What it does
 
@@ -16,18 +17,77 @@ GPU-aware filters and scorers.
    - If the sample has no `pod` label, it is kept (sidecar / unlabeled payloads).
 4. Aggregates matching samples using `max` (highest-utilized GPU for the pod).
 5. Normalizes the value from 0-100 to [0.0, 1.0].
-6. Stores the result as a `GPUUtilization` attribute on the corresponding endpoint.
+6. Stores the result as `ScalarMetricValue` on the corresponding endpoint.
 
 ## Attributes produced
 
-- `GPUUtilization` stored at attribute key `GPUUtilizationDataKey`
-  (`GPUUtilization/dcgm-extractor`) on each endpoint.
-
-```go
-key := attrgpu.GPUUtilizationDataKey.String()
-util, ok := attrgpu.ReadGPUUtilization(endpoint.GetAttributes(), key)
-```
+- **Type:** `ScalarMetricValue` (from `attribute/metrics`)
+- **Key string:** `GPUUtilization/dcgm-extractor`
 
 ## Configuration
 
 No configuration parameters.
+
+## Cluster prerequisites
+
+- **DCGM Exporter** must be running on the cluster and accessible from the
+  EPP. See [`source/dcgm/README.md`](../../source/dcgm/README.md) for
+  data source configuration (port, DaemonSet vs sidecar, TLS).
+
+## EPP config example
+
+Use the generic `endpoint-attribute-filter` and `endpoint-attribute-scorer`
+to consume the produced attribute. The `attribute`/`attributeKey` parameter
+must match the key string above.
+
+```yaml
+apiVersion: llm-d.ai/v1alpha1
+kind: EndpointPickerConfig
+plugins:
+- type: dcgm-data-source
+  name: dcgm-source
+  parameters:
+    port: 9400
+- type: dcgm-extractor
+  name: dcgm-extractor
+- type: endpoint-attribute-filter
+  name: gpu-utilization-filter
+  parameters:
+    attribute: "GPUUtilization/dcgm-extractor"
+    onMissing: "Pass"
+    fallbackOnEmpty: true
+    algorithm:
+      type: "threshold"
+      threshold:
+        operator: "LessThanOrEqual"
+        value: 0.90
+- type: endpoint-attribute-scorer
+  name: gpu-utilization-scorer
+  parameters:
+    attributeKey: "GPUUtilization/dcgm-extractor"
+    algorithm:
+      type: "linear_lower_is_better"
+      normalization:
+        fixedRange:
+          min: 0.0
+          max: 1.0
+- type: kv-cache-utilization-scorer
+- type: max-score-picker
+- type: single-profile-handler
+- type: decode-filter
+dataLayer:
+  sources:
+  - pluginRef: dcgm-source
+    extractors:
+    - pluginRef: dcgm-extractor
+schedulingProfiles:
+- name: default
+  plugins:
+  - pluginRef: decode-filter
+  - pluginRef: gpu-utilization-filter
+  - pluginRef: gpu-utilization-scorer
+    weight: 2.0
+  - pluginRef: kv-cache-utilization-scorer
+    weight: 1.0
+  - pluginRef: max-score-picker
+```

pkg/epp/framework/plugins/datalayer/extractor/dcgm/extractor.go

@@ -26,6 +26,7 @@ import (
 	fwkdl "github.com/llm-d/llm-d-router/pkg/epp/framework/interface/datalayer"
 	fwkplugin "github.com/llm-d/llm-d-router/pkg/epp/framework/interface/plugin"
 	attrgpu "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/datalayer/attribute/gpu"
+	attrmetrics "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/datalayer/attribute/metrics"
 	sourcemetrics "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/datalayer/source/metrics"
 )
 
@@ -101,14 +102,14 @@ func (e *Extractor) Extract(_ context.Context, in fwkdl.PollInput[sourcemetrics.
 		return fmt.Errorf("metric %q present but has no samples for pod %q", gpuUtilMetricName, podName)
 	}
 
-	normalized := attrgpu.GPUUtilization(maxUtil / 100.0)
+	normalized := attrmetrics.ScalarMetricValue(maxUtil / 100.0)
 	in.Endpoint.GetAttributes().Put(e.dk.String(), normalized)
 	return nil
 }
 
 // Produces declares the data key this extractor publishes.
 func (e *Extractor) Produces() map[fwkplugin.DataKey]any {
-	return map[fwkplugin.DataKey]any{e.dk: attrgpu.GPUUtilization(0)}
+	return map[fwkplugin.DataKey]any{e.dk: attrmetrics.ScalarMetricValue(0)}
 }
 
 // sampleBelongsToPod reports whether m should contribute to this endpoint's

pkg/epp/framework/plugins/datalayer/extractor/dcgm/extractor_test.go

@@ -28,6 +28,7 @@ import (
 
 	fwkdl "github.com/llm-d/llm-d-router/pkg/epp/framework/interface/datalayer"
 	attrgpu "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/datalayer/attribute/gpu"
+	attrmetrics "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/datalayer/attribute/metrics"
 	sourcemetrics "github.com/llm-d/llm-d-router/pkg/epp/framework/plugins/datalayer/source/metrics"
 )
 
@@ -215,7 +216,7 @@ func TestExtractorExtract(t *testing.T) {
 				if diff := cmp.Diff(before, after); diff == "" {
 					t.Error("expected attribute to be updated, but no change detected")
 				}
-				val, ok := attrgpu.ReadGPUUtilization(attr, key)
+				val, ok := attrmetrics.ReadScalarMetricValue(attr, key)
 				if !ok {
 					t.Fatal("GPU utilization not found in attributes after extract")
 				}
@@ -247,7 +248,23 @@ func TestFactory_SetsName(t *testing.T) {
 func TestProduces_DeclaresGPUUtilization(t *testing.T) {
 	ext := NewDCGMExtractor()
 	produced := ext.Produces()
-	if _, ok := produced[attrgpu.GPUUtilizationDataKey]; !ok {
+	val, ok := produced[attrgpu.GPUUtilizationDataKey]
+	if !ok {
 		t.Error("Produces must declare GPUUtilizationDataKey")
 	}
+	if _, isScalar := val.(attrmetrics.ScalarMetricValue); !isScalar {
+		t.Errorf("Produces value type = %T, want ScalarMetricValue", val)
+	}
+}
+
+// TestAttributeKeyContract verifies that the attribute key string stored by
+// the extractor matches the value that generic endpoint-attribute-filter and
+// endpoint-attribute-scorer expect in their YAML config. A mismatch means
+// the filter/scorer silently skip all endpoints (attribute not found).
+func TestAttributeKeyContract(t *testing.T) {
+	const expectedKey = "GPUUtilization/dcgm-extractor"
+	got := attrgpu.GPUUtilizationDataKey.String()
+	if got != expectedKey {
+		t.Errorf("GPUUtilizationDataKey.String() = %q, want %q (must match YAML attribute config)", got, expectedKey)
+	}
 }