fix(scheduler): fix log messages and avoid vector mutation in comparisons

- Fix log messages in allocate.go that said "requires: %v GPUs" but now
  print a full resource vector
- Rewrite lessEqualVectorsExcludingGPU to iterate and skip the GPU index
  instead of mutating input vectors via save/zero/compare/restore, which
  is not safe under concurrent access

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: enoodle

pkg/scheduler/actions/common/allocate.go

@@ -174,7 +174,7 @@ func allocateTaskToNode(ssn *framework.Session, stmt *framework.Statement, task
 }
 
 func bindTaskToNode(ssn *framework.Session, stmt *framework.Statement, task *pod_info.PodInfo, node *node_info.NodeInfo) bool {
-	log.InfraLogger.V(6).Infof("Binding Task <%v/%v> to node <%v>, requires: %v GPUs",
+	log.InfraLogger.V(6).Infof("Binding Task <%v/%v> to node <%v>, requires resources: %v",
 		task.Namespace, task.Name, node.Name, task.ResReqVector)
 
 	if err := stmt.Allocate(task, node.Name); err != nil {
@@ -185,7 +185,7 @@ func bindTaskToNode(ssn *framework.Session, stmt *framework.Statement, task *pod
 }
 
 func pipelineTaskToNode(ssn *framework.Session, stmt *framework.Statement, task *pod_info.PodInfo, node *node_info.NodeInfo, updateTasksIfExistsOnNode bool) bool {
-	log.InfraLogger.V(6).Infof("Pipelining Task <%v/%v> to node <%v> requires: %v GPUs",
+	log.InfraLogger.V(6).Infof("Pipelining Task <%v/%v> to node <%v>, requires resources: %v",
 		task.Namespace, task.Name, node.Name, task.ResReqVector)
 
 	if err := stmt.Pipeline(task, node.Name, updateTasksIfExistsOnNode); err != nil {

pkg/scheduler/api/node_info/gpu_sharing_node_info.go

@@ -71,7 +71,7 @@ func getAcceptedTaskResourceWithoutSharedGPU(task *pod_info.PodInfo) *resource_i
 	return requestedResourceWithoutSharedGPU
 }
 
-func (ni *NodeInfo) addSharedTaskResources(task *pod_info.PodInfo) {
+func (ni *NodeInfo) addSharedGPUTaskResources(task *pod_info.PodInfo) {
 	if !task.IsSharedGPUAllocation() {
 		return
 	}
@@ -80,14 +80,14 @@ func (ni *NodeInfo) addSharedTaskResources(task *pod_info.PodInfo) {
 		task.Namespace, task.Name, task.Status, ni)
 
 	for _, gpuGroup := range task.GPUGroups {
-		ni.addSharedTaskResourcesPerPodGroup(task, gpuGroup)
+		ni.addSharedGPUTaskResourcesPerPodGroup(task, gpuGroup)
 	}
 
 	log.InfraLogger.V(8).Infof("Added shared podsInfo: <%v/%v>, status: <%v>, node: <%+v>",
 		task.Namespace, task.Name, task.Status, ni)
 }
 
-func (ni *NodeInfo) addSharedTaskResourcesPerPodGroup(task *pod_info.PodInfo, gpuGroup string) {
+func (ni *NodeInfo) addSharedGPUTaskResourcesPerPodGroup(task *pod_info.PodInfo, gpuGroup string) {
 	log.InfraLogger.V(7).Infof(
 		"About to add shared podsInfo: <%v/%v>, gpuGroup: <%v> "+
 			"releasingSharedGPU: <%v> AllocatedSharedGPUsMemory <%v>, UsedSharedGPUsMemory: <%v>",

pkg/scheduler/api/node_info/node_info.go

@@ -393,14 +393,15 @@ func (ni *NodeInfo) isTaskAllocatableOnNonAllocatedResources(
 
 func (ni *NodeInfo) lessEqualVectorsExcludingGPU(a, b resource_info.ResourceVector) bool {
 	gpuIdx := ni.VectorMap.GetIndex(commonconstants.GpuResource)
-	savedA := a.Get(gpuIdx)
-	savedB := b.Get(gpuIdx)
-	a.Set(gpuIdx, 0)
-	b.Set(gpuIdx, 0)
-	result := a.LessEqual(b)
-	a.Set(gpuIdx, savedA)
-	b.Set(gpuIdx, savedB)
-	return result
+	for i := 0; i < len(a); i++ {
+		if i == gpuIdx {
+			continue
+		}
+		if a.Get(i) > b.Get(i) {
+			return false
+		}
+	}
+	return true
 }
 
 func (ni *NodeInfo) AddTask(task *pod_info.PodInfo) error {
@@ -508,7 +509,7 @@ func (ni *NodeInfo) addTaskResources(task *pod_info.PodInfo) {
 		ni.IdleVector.Sub(resourcesToTrackVector)
 	}
 
-	ni.addSharedTaskResources(task)
+	ni.addSharedGPUTaskResources(task)
 
 	log.InfraLogger.V(8).Infof("Added podsInfo: <%v/%v>, status: <%v>, node: <%+v>",
 		task.Namespace, task.Name, task.Status, ni)