From 4e37f64553728bab5b59ae7d7e8a50ee91c23232 Mon Sep 17 00:00:00 2001 From: Arsolitt Date: Sat, 18 Apr 2026 07:19:05 +0300 Subject: [PATCH] docs(gpu-operator): document tolerate-all on compat DaemonSet Explain why tolerations: [{operator: Exists}] is safe on the driver compat DaemonSet: the nodeSelector already confines scheduling to GPU nodes, so the blanket toleration only kicks in when those nodes carry the dedicated=gpu / nvidia.com/gpu taints that the GPU Operator's default policy and many deployments apply. Signed-off-by: Arsolitt --- .../system/gpu-operator/examples/nvidia-driver-compat.yaml | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/packages/system/gpu-operator/examples/nvidia-driver-compat.yaml b/packages/system/gpu-operator/examples/nvidia-driver-compat.yaml index 63e07c4f..bad8609e 100644 --- a/packages/system/gpu-operator/examples/nvidia-driver-compat.yaml +++ b/packages/system/gpu-operator/examples/nvidia-driver-compat.yaml @@ -35,6 +35,13 @@ spec: # cluster uses to mark GPU hosts. nodeSelector: nvidia.com/gpu.present: "true" + # Tolerate-all is intentionally broad: the nodeSelector above already + # confines scheduling to GPU nodes, so the "Exists" toleration only + # matters when those nodes carry extra taints (dedicated=gpu:NoSchedule, + # nvidia.com/gpu:NoSchedule from the GPU Operator's default policy, + # etc.). This mirrors the stance of the upstream nvidia-driver-daemonset + # and nvidia-device-plugin DaemonSets — blanket tolerations plus a + # narrow nodeSelector. tolerations: - operator: Exists initContainers: