From 68849b86f973db7dec46f683f1f849f623ebedc0 Mon Sep 17 00:00:00 2001
From: dvirlabs <dvirlabs@gmail.com>
Date: Sun, 1 Mar 2026 17:24:03 +0200
Subject: [PATCH] fix: target NVIDIA device plugin to GPU node k3s-worker-5-gpu
 with proper tolerations

---
 .../nvidia-device-plugin.yaml                     | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/manifests/nvidia-device-plugin/nvidia-device-plugin.yaml b/manifests/nvidia-device-plugin/nvidia-device-plugin.yaml
index 409a610..c50cc06 100644
--- a/manifests/nvidia-device-plugin/nvidia-device-plugin.yaml
+++ b/manifests/nvidia-device-plugin/nvidia-device-plugin.yaml
@@ -28,13 +28,16 @@ spec:
       labels:
         name: nvidia-device-plugin-ds
     spec:
-      # nodeSelector removed for initial deployment to avoid chicken-egg problem
-      # The nvidia.com/gpu.present label is applied BY the device plugin itself
-      # FAIL_ON_INIT_ERROR=false ensures graceful skip on nodes without GPU
-      # After first deployment, you can optionally re-enable with:
-      # nodeSelector:
-      #   nvidia.com/gpu.present: "true"
+      # Schedule ONLY on the GPU node by hostname
+      nodeSelector:
+        kubernetes.io/hostname: k3s-worker-5-gpu
       tolerations:
+      # Tolerate the gpu=true:NoSchedule taint on the GPU node
+      - key: gpu
+        operator: Equal
+        value: "true"
+        effect: NoSchedule
+      # Also tolerate nvidia.com/gpu taint if present
       - key: nvidia.com/gpu
         operator: Exists
         effect: NoSchedule