fix: Add gpu monitoring support for some machines (#10133)

Refs #9504 #8397

agent/utils/ai_tools/gpu/gpu.go

@@ -59,11 +59,5 @@ func (n NvidiaSMI) LoadGpuInfo() (*common.GpuInfo, error) {
 		break
 	}
 
-	if version == "v12" || version == "v11" {
-		return schema.Parse(data, version)
-	} else {
-		global.LOG.Errorf("don't support such schema version %s", version)
-	}
-
-	return &common.GpuInfo{}, nil
+	return schema.Parse(data, version)
 }

agent/utils/ai_tools/gpu/schema/parser.go

@@ -33,9 +33,12 @@ func Parse(buf []byte, version string) (*common.GpuInfo, error) {
 
 		gpuItem.Temperature = s.Gpu[i].Temperature.GpuTemp
 		gpuItem.PerformanceState = s.Gpu[i].PerformanceState
-		if version == "v12" {
+		if version == "v12" || version == "v13" {
 			gpuItem.PowerDraw = s.Gpu[i].GpuPowerReadings.PowerDraw
-			gpuItem.MaxPowerLimit = s.Gpu[i].GpuPowerReadings.MaxPowerLimit
+			if len(gpuItem.PowerDraw) == 0 {
+				gpuItem.PowerDraw = s.Gpu[i].GpuPowerReadings.InstantPowerDraw
+			}
+			gpuItem.MaxPowerLimit = s.Gpu[i].GpuPowerReadings.CurrentPowerLimit
 		} else {
 			gpuItem.PowerDraw = s.Gpu[i].PowerReadings.PowerDraw
 			gpuItem.MaxPowerLimit = s.Gpu[i].PowerReadings.MaxPowerLimit

agent/utils/ai_tools/gpu/schema/types.go

@@ -108,6 +108,7 @@ type smi struct {
 		} `xml:"gpu_operation_mode"`
 		GpuPartNumber    string `xml:"gpu_part_number"`
 		GpuPowerReadings struct {
+			InstantPowerDraw    string `xml:"instant_power_draw"`
 			CurrentPowerLimit   string `xml:"current_power_limit"`
 			DefaultPowerLimit   string `xml:"default_power_limit"`
 			MaxPowerLimit       string `xml:"max_power_limit"`