fix(monitor): avoid uint64 underflow when deriving the legacy device-memory offset label

cmd/vGPUmonitor/feedback_test.go

@@ -38,6 +38,7 @@ type stubInfo struct {
 	ctxSize    []uint64
 	modSize    []uint64
 	bufSize    []uint64
+	offset     []uint64
 	smUtil     []uint64
 	lastKernel int64
 }
@@ -60,7 +61,7 @@ func (s *stubInfo) DeviceUUID(i int) string {
 func (s *stubInfo) DeviceMemoryContextSize(i int) uint64 { return slot(s.ctxSize, i) }
 func (s *stubInfo) DeviceMemoryModuleSize(i int) uint64  { return slot(s.modSize, i) }
 func (s *stubInfo) DeviceMemoryBufferSize(i int) uint64  { return slot(s.bufSize, i) }
-func (s *stubInfo) DeviceMemoryOffset(int) uint64        { return 0 }
+func (s *stubInfo) DeviceMemoryOffset(i int) uint64      { return slot(s.offset, i) }
 func (s *stubInfo) DeviceMemoryTotal(i int) uint64       { return slot(s.total, i) }
 func (s *stubInfo) DeviceSmUtil(i int) uint64            { return slot(s.smUtil, i) }
 func (s *stubInfo) SetDeviceSmLimit(uint64)              {}

cmd/vGPUmonitor/metrics.go

@@ -463,7 +463,8 @@ func (cc ClusterManagerCollector) collectContainerMetrics(ch chan<- prometheus.M
 			klog.Errorf("Failed to send device memory desc: %v", err)
 			return err
 		}
-		memoryOffset := memoryTotal - memoryContextSize - memoryModuleSize - memoryBufferSize
+		// Send legacy metric with additional memory details
+		memoryOffset := c.Info.DeviceMemoryOffset(i) // Get the memory offset for the device
 		memoryLabels := append(labels, fmt.Sprint(memoryContextSize), fmt.Sprint(memoryModuleSize), fmt.Sprint(memoryBufferSize), fmt.Sprint(memoryOffset))
 		sendLegacyMetric(ch, legacyCtrDeviceMemorydesc, prometheus.GaugeValue, float64(memoryTotal), memoryLabels...)
 

cmd/vGPUmonitor/metrics_container_test.go

@@ -154,3 +154,47 @@ func TestCollectContainerMetricsSkipsInvalidUTF8UUID(t *testing.T) {
 		t.Errorf("got %d metrics for an invalid UUID, want 0", len(metrics))
 	}
 }
+
+// TestCollectContainerMetricsLegacyOffsetNoUnderflow verifies that the legacy
+// device-memory offset label uses the recorded value without uint64 underflow.
+func TestCollectContainerMetricsLegacyOffsetNoUnderflow(t *testing.T) {
+	legacyCtrDeviceMemorydesc = prometheus.NewDesc(
+		"Device_memory_desc_of_container",
+		"Container device memory description",
+		[]string{"podnamespace", "podname", "ctrname", "vdeviceid", "deviceuuid", "context", "module", "data", "offset"}, nil,
+	)
+	t.Cleanup(func() { legacyCtrDeviceMemorydesc = prev })
+
+	cu := &nvidia.ContainerUsage{Info: &stubInfo{
+		uuids:   []string{testUUID},
+		total:   []uint64{100},
+		ctxSize: []uint64{1000},
+		modSize: []uint64{500},
+		bufSize: []uint64{300},
+		offset:  []uint64{7},
+	}}
+
+	pod := &corev1.Pod{ObjectMeta: metav1.ObjectMeta{Namespace: "team-a", Name: "trainer-0"}}
+	ctr := corev1.Container{Name: "worker"}
+	cc := ClusterManagerCollector{ClusterManager: &ClusterManager{LegacyMetrics: true}}
+	ch := make(chan prometheus.Metric, 64)
+	if err := cc.collectContainerMetrics(ch, pod, ctr, cu, 0); err != nil {
+		t.Fatal(err)
+	}
+	close(ch)
+
+	found := false
+	for m := range ch {
+		if m.Desc() != legacyCtrDeviceMemorydesc {
+			continue
+		}
+		found = true
+		_, labels := gaugeValue(t, m)
+		if labels["offset"] != "7" {
+			t.Errorf("offset label = %q, want %q (got a wrapped/underflowed value instead of the recorded offset)", labels["offset"], "7")
+		}
+	}
+	if !found {
+		t.Fatal("legacy device memory metric was not emitted")
+	}
+}