fix: remove panic in calculateGPUPairScore

pkg/device/nvidia/calculate_score.go

@@ -22,6 +22,7 @@ import (
 
 	"github.com/NVIDIA/go-nvlib/pkg/nvlib/device"
 	"github.com/NVIDIA/go-nvml/pkg/nvml"
+	"k8s.io/klog/v2"
 )
 
 // Device represents a GPU device as reported by NVML, including all of its
@@ -201,6 +202,8 @@ func calculateGPUScore(devices []*Device) ListDeviceScore {
 	return ds
 }
 
+// calculateGPUPairScore calculates a heuristic score representing the P2P connection
+// strength between two GPUs. A higher score implies better/faster communication links.
 func calculateGPUPairScore(gpu0 *Device, gpu1 *Device) int {
 	if gpu0 == nil || gpu1 == nil {
 		return 0
@@ -212,7 +215,8 @@ func calculateGPUPairScore(gpu0 *Device, gpu1 *Device) int {
 
 	if len(gpu0.Links[gpu1.Index]) != len(gpu1.Links[gpu0.Index]) {
 		err := fmt.Errorf("internal error in bestEffort GPU allocator: all P2PLinks between 2 GPUs should be bidirectional")
-		panic(err)
+		klog.ErrorS(err, "P2PLinks mismatch", "gpu0", gpu0.UUID, "gpu1", gpu1.UUID)
+		return 0
 	}
 
 	score := 0

pkg/device/nvidia/calculate_score_test.go

@@ -106,17 +106,49 @@
 					UUID:  "gpu3",
 					Score: map[string]int{"gpu0": 200, "gpu1": 100, "gpu2": 200},
 				},
+		},
+		{
+			name: "asymmetric links",
+			args: []*Device{
+				{
+					Index: 0,
+					nvlibDevice: nvlibDevice{
+						UUID: "gpu0",
+					},
+					Links: map[int][]P2PLink{
+						1: {{Type: SingleNVLINKLink}},
+					},
+				},
+				{
+					Index: 1,
+					nvlibDevice: nvlibDevice{
+						UUID: "gpu1",
+					},
+					Links: map[int][]P2PLink{
+						0: {{Type: SingleNVLINKLink}, {Type: SingleNVLINKLink}}, // Asymmetric: 2 links here vs 1 above
+					},
+				},
+			},
+			want: ListDeviceScore{
+				{
+					UUID:  "gpu0",
+					Score: map[string]int{"gpu1": 0},
+				},
+				{
+					UUID:  "gpu1",
+					Score: map[string]int{"gpu0": 0},
+				},
 			},
 		},
 	}
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
 			scoreList := make(ListDeviceScore, len(test.args))
 			for i, gpuI := range test.args {
 				score := make(map[string]int)
 				for j, gpuJ := range test.args {
 					if i == j {
 						continue
 					}
 					score[gpuJ.UUID] = calculateGPUPairScore(gpuI, gpuJ)
 				}