feat(ollama): add docker swarm configuration for rocm gpu support

Introduce a new compose.yml to deploy Ollama with AMD GPU hardware acceleration on Docker Swarm.

- Uses the ROCm-enabled Ollama image.
- Mounts necessary GPU devices (/dev/kfd and /dev/dri) with root permissions for hardware access.
- Sets HSA_OVERRIDE_GFX_VERSION to 10.3.0 to support Navi 21 (6900 XT) cards.
- Configures persistent storage on local SSD and integrates with an external proxy network.
- Includes deployment constraints to target nodes labeled with GPUs.

Local Homelab/ollama/compose/compose.yml

@@ -0,0 +1,38 @@
+version: '3.8'
+name: ollama
+
+services:
+  ollama:
+    image: ollama/ollama:rocm
+    deploy:
+      replicas: 1
+      placement:
+        constraints:
+          - node.labels.gpu == true
+      restart_policy:
+        condition: on-failure
+    
+    # We use "host" mode for devices in Swarm to avoid permission headaches
+    # This allows the container to see the /dev/kfd and /dev/dri paths we mount below
+    user: root
+    
+    volumes:
+      - /mnt/local-ssd/docker/ollama:/root/.ollama
+      # Mount the Kernel Fusion Driver (Required for ROCm)
+      - /dev/kfd:/dev/kfd
+      # Mount the Direct Rendering Interface (The actual cards)
+      - /dev/dri:/dev/dri
+
+    networks:
+      - proxy
+      
+    environment:
+      # Force support for Navi 21 (6900 XT)
+      - HSA_OVERRIDE_GFX_VERSION=10.3.0
+      # Tell ROCm to verify usage
+      - ROCR_VISIBLE_DEVICES=all
+      - OLLAMA_HOST=0.0.0.0
+
+networks:
+  proxy:
+    external: true