Bump Ollama memory to 24Gi and enable flash attention

The 27B Q4_K_M model needs ~7.3 GiB system RAM for CPU-offloaded layers
but only 6.8 GiB was available within the 22Gi cgroup. Bumping to 24Gi
and enabling flash attention (reduces KV cache memory) should provide
enough headroom.

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

argocd/manifests/ollama/deployment.yaml

@@ -32,6 +32,8 @@ spec:
               value: "1"
             - name: OLLAMA_NUM_PARALLEL
               value: "1"
+            - name: OLLAMA_FLASH_ATTENTION
+              value: "1"
           volumeMounts:
             - name: models
               mountPath: /models
@@ -40,7 +42,7 @@ spec:
               memory: "512Mi"
               cpu: "500m"
             limits:
-              memory: "22Gi"
+              memory: "24Gi"
               cpu: "4000m"
               nvidia.com/gpu: "1"
           livenessProbe: