From c72ede4b6d4e41650d359012afb2f9731fbbe149 Mon Sep 17 00:00:00 2001
From: Drew Viles <drew@hudson-viles.uk>
Date: Mon, 8 Jan 2024 14:12:18 +0000
Subject: [PATCH] fix: Added GPU node selector

This prevents a GPU destined pod ending up on the non-gpu based nodes. This has been seen in the wild even though the resources request has a GPU specified.
---
 charts/dogkat/templates/gpu-pod.yaml | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/charts/dogkat/templates/gpu-pod.yaml b/charts/dogkat/templates/gpu-pod.yaml
index 1817626..1b929da 100644
--- a/charts/dogkat/templates/gpu-pod.yaml
+++ b/charts/dogkat/templates/gpu-pod.yaml
@@ -14,13 +14,14 @@ metadata:
     {{- include "e2e-testing.gpu.labels" . | nindent 4 }}
 spec:
   restartPolicy: OnFailure
+  nodeSelector:
+    nvidia.com/gpu.present: "true"
   containers:
     - name: nvidia-vectoradd
       image: {{.Values.gpu.image.repo}}:{{.Values.gpu.image.tag}}
       terminationMessagePath: /dev/termination-log
       terminationMessagePolicy: File
       imagePullPolicy: IfNotPresent
-
       {{- if .Values.gpu.resources }}
       resources:
         "nvidia.com/gpu": {{ .Values.gpu.numberOfGPUs }}