From 6dfa76565db613178688b99489e65c9b8125d05c Mon Sep 17 00:00:00 2001
From: Zachary Zhao <ongarabazanadi77@gmail.com>
Date: Sun, 25 Oct 2026 03:07:24 +0800
Subject: [PATCH] fix(api): accept the binder-translated NvFractions limit next
 to gpu-fraction

In NvFractions mode the binder turns gpu-memory.portion.limit into the
nvidia.com/container.<c>.gpu-memory.limit annotation at bind time and keeps
gpu-fraction on the pod. ValidateGPUFractionRequest then rejected that update
as "cannot combine ... limit annotation with gpu-fraction", so every bind of a
pod with a portion limit failed and the pod cycled between Scheduled and
BindingError.

Exempt the translated limit when the portion-limit annotation is present,
mirroring the existing gpu-memory.limit shorthand exemption.

Ref kai-scheduler/KAI-Scheduler#2490

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Signed-off-by: Zachary Zhao <ongarabazanadi77@gmail.com>
---
 .../unreleased/fixed-20261025-025826.yaml     |  7 +++
 utilities/resources/gpu_sharing_validation.go |  6 +-
 .../resources/gpu_sharing_validation_test.go  | 61 +++++++++++++++++++
 3 files changed, 73 insertions(+), 1 deletion(-)
 create mode 100644 .changes/unreleased/fixed-20261025-025826.yaml

diff --git a/.changes/unreleased/fixed-20261025-025826.yaml b/.changes/unreleased/fixed-20261025-025826.yaml
new file mode 100644
index 0000000..1d49007
--- /dev/null
+++ b/.changes/unreleased/fixed-20261025-025826.yaml
@@ -0,0 +1,7 @@
+kind: Fixed
+body: |-
+  Allow the binder-translated NvFractions limit next to gpu-fraction when gpu-memory.portion.limit is set
+
+custom:
+    Author: ZeXi0308
+    Issue: "28"
diff --git a/utilities/resources/gpu_sharing_validation.go b/utilities/resources/gpu_sharing_validation.go
index 2ae5fe2..d933424 100644
--- a/utilities/resources/gpu_sharing_validation.go
+++ b/utilities/resources/gpu_sharing_validation.go
@@ -56,7 +56,11 @@ func ValidateGPUFractionRequest(pod *v1.Pod) error {
 
 	// A legacy gpu-memory request may accompany a translated shorthand limit.
 	_, hasMemoryLimitShorthand := pod.Annotations[constants.GpuMemoryLimit]
-	if req.Limit != nil && (hasLegacyFraction || (hasLegacyMemory && !hasMemoryLimitShorthand)) {
+	// A legacy gpu-fraction request may accompany the NvFractions limit the binder
+	// derives from gpu-memory.portion.limit at bind time.
+	_, _, hasPortionLimit := ExtractGpuMemoryPortionLimitAnnotation(pod)
+	if req.Limit != nil &&
+		((hasLegacyFraction && !hasPortionLimit) || (hasLegacyMemory && !hasMemoryLimitShorthand)) {
 		return fmt.Errorf("cannot combine %s limit annotation with %s or %s annotation",
 			constants.NvFractionsAnnotationPrefix, constants.GpuFraction, constants.GpuMemory)
 	}
diff --git a/utilities/resources/gpu_sharing_validation_test.go b/utilities/resources/gpu_sharing_validation_test.go
index 6d4c59a..9cf9ff9 100644
--- a/utilities/resources/gpu_sharing_validation_test.go
+++ b/utilities/resources/gpu_sharing_validation_test.go
@@ -177,6 +177,67 @@ func TestValidateGpuMemoryPortionLimitAnnotation(t *testing.T) {
 	}
 }
 
+func TestValidateGPUFractionRequest_TranslatedLimits(t *testing.T) {
+	nvRequestKey := CalcGpuFractionAnnotationForContainer("container-0")
+	nvLimitKey := CalcGpuFractionLimitAnnotationForContainer("container-0")
+	portionLimitKey := CalcGpuMemoryPortionLimitAnnotationForContainer("container-0")
+
+	tests := []struct {
+		name              string
+		annotations       map[string]string
+		wantErrContaining string
+	}{
+		{
+			name: "gpu-fraction with portion limit accepts the binder-translated NvFractions limit",
+			annotations: map[string]string{
+				constants.GpuFraction: "0.5",
+				portionLimitKey:       "0.8",
+				nvRequestKey:          "1000Mi",
+				nvLimitKey:            "1600Mi",
+			},
+		},
+		{
+			name: "gpu-fraction without portion limit rejects an NvFractions limit",
+			annotations: map[string]string{
+				constants.GpuFraction: "0.5",
+				nvRequestKey:          "1000Mi",
+				nvLimitKey:            "1600Mi",
+			},
+			wantErrContaining: "cannot combine",
+		},
+		{
+			name: "gpu-memory with gpu-memory.limit accepts the translated NvFractions limit",
+			annotations: map[string]string{
+				constants.GpuMemory:      "1000",
+				constants.GpuMemoryLimit: "1600Mi",
+				nvRequestKey:             "1000Mi",
+				nvLimitKey:               "1600Mi",
+			},
+		},
+		{
+			name: "gpu-memory without gpu-memory.limit rejects an NvFractions limit",
+			annotations: map[string]string{
+				constants.GpuMemory: "1000",
+				nvRequestKey:        "1000Mi",
+				nvLimitKey:          "1600Mi",
+			},
+			wantErrContaining: "cannot combine",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			pod := &v1.Pod{
+				ObjectMeta: metav1.ObjectMeta{Annotations: tt.annotations},
+				Spec:       v1.PodSpec{Containers: []v1.Container{{Name: "container-0"}}},
+			}
+
+			err := ValidateGPUFractionRequest(pod)
+			assertErrorContains(t, err, tt.wantErrContaining)
+		})
+	}
+}
+
 func TestValidateGPUFractionRequest_ComputeSharingMode(t *testing.T) {
 	tests := []struct {
 		name        string