fix(amd): round core requests up to whole WGPs

pkg/device/amd/device.go

@@ -336,6 +336,11 @@ func (amddevice *AMDDevices) Fit(devices []*device.DeviceUsage, request device.C
 			}
 			coreReq = dev.Totalcore * k.Coresreq / 100
 			coreReq = max(coreReq, 1)
+			// RDNA applies the CU mask per WGP, so the device plugin hands out
+			// whole WGPs; account for the same count.
+			if unit := cuPerWGP(dev.CustomInfo); unit > 1 {
+				coreReq = int32(min((int64(coreReq)+int64(unit)-1)/int64(unit)*int64(unit), int64(dev.Totalcore)))
+			}
 			coreReq = min(coreReq, dev.Totalcore)
 		} else if dev.Totalmem > 0 && memReq >= dev.Totalmem {
 			// Memreq omitted or zero means whole-card memory; treat core request as whole-card as well.
@@ -381,3 +386,12 @@ func (amddevice *AMDDevices) Fit(devices []*device.DeviceUsage, request device.C
 	}
 	return false, tmpDevs, common.GenReason(reason, len(devices))
 }
+
+// cuPerWGP returns how many CUs the device plugin allocates together, as
+// published in the registration custominfo; 1 when absent.
+func cuPerWGP(info map[string]any) int32 {
+	if v, ok := info["cuPerWGP"].(float64); ok && v > 1 && v <= math.MaxInt32 && v == math.Trunc(v) {
+		return int32(v)
+	}
+	return 1
+}

pkg/device/amd/device_test.go

@@ -415,6 +415,35 @@ func TestDevices_Fit(t *testing.T) {
 		assert.Equal(t, "", reason)
 	})
 
+	t.Run("rounds cores up to whole WGPs on RDNA", func(t *testing.T) {
+		for _, tc := range []struct {
+			name                  string
+			total, used, coresReq int32
+			info                  map[string]any
+			wantCores             int32
+			wantOK                bool
+		}{
+			{"rdna odd cores", 64, 0, 5, map[string]any{"cuPerWGP": float64(2)}, 4, true},
+			{"cdna keeps single cus", 64, 0, 5, map[string]any{}, 3, true},
+			{"one-wgp apu", 2, 0, 50, map[string]any{"cuPerWGP": float64(2)}, 2, true},
+			{"huge wgp size does not overflow", 4, 0, 75, map[string]any{"cuPerWGP": float64(math.MaxInt32)}, 4, true},
+			{"fractional wgp size is ignored", 64, 0, 5, map[string]any{"cuPerWGP": 2.5}, 3, true},
+			{"apu wgp already taken", 2, 2, 50, map[string]any{"cuPerWGP": float64(2)}, 0, false},
+		} {
+			devices := []*device.DeviceUsage{{
+				ID: "dev-0", Count: 10, Totalmem: 1000, Totalcore: tc.total, Usedcores: tc.used,
+				Type: AMDDevice, Health: true, CustomInfo: tc.info,
+			}}
+			req := device.ContainerDeviceRequest{Nums: 1, Type: AMDDevice, Memreq: 100, Coresreq: tc.coresReq}
+			pod := &corev1.Pod{ObjectMeta: metav1.ObjectMeta{Annotations: map[string]string{}}}
+			ok, got, _ := dev.Fit(devices, req, pod, &device.NodeInfo{}, &device.PodDevices{})
+			assert.Equal(t, tc.wantOK, ok, tc.name)
+			if tc.wantOK {
+				assert.Equal(t, tc.wantCores, got[AMDDevice][0].Usedcores, tc.name)
+			}
+		}
+	})
+
 	t.Run("retains the registered product type in the allocation", func(t *testing.T) {
 		const productType = "AMD_Instinct_MI300X_VF"
 		devices := []*device.DeviceUsage{