Restore vision projector for medium profile

compose.yaml

@@ -153,12 +153,15 @@ services:
     environment:
       NVIDIA_VISIBLE_DEVICES: ${MEDIUM_GPU_DEVICES:-0,1}
       NVIDIA_DRIVER_CAPABILITIES: compute,utility
-      # Temporary llama.cpp cache workaround: keep Medium text-only. Current
-      # multimodal builds can discard the complete KV state on a later turn.
-      # The model split, context and MTP settings remain unchanged.
+      MTMD_BACKEND_DEVICE: CUDA1
     command:
       - --model
       - "/models/${MEDIUM_MODEL_FILE:?MEDIUM_MODEL_FILE is required}"
+      - --mmproj
+      - "/models/${VISION_PROJECTOR_FILE:?VISION_PROJECTOR_FILE is required}"
+      - --mmproj-offload
+      - --mmproj-device
+      - CUDA1
       - --alias
       - qwen-medium
       - --ctx-size