llama-swap: use pre-built images (:cuda, :rocm) with GPU-specific flags

- Drop custom Dockerfiles; docker-compose uses ghcr.io pre-built images
  which ship llama-swap + llama-server with no pinned versions (always latest)
- NVIDIA GTX 1660 (6GB): add -fit off --no-kv-offload --cache-type-k q4_0 --cache-type-v q4_0
  to fix OOM segfault with new llama.cpp b9014's GPU-side KV cache default
- AMD RX 6800 (16GB): flags unchanged; KV cache stays on GPU for max speed
- Both running llama-swap v211 + llama.cpp b9014 (2026-05-05)

docker-compose.yml

@@ -22,9 +22,7 @@ services:
       - LOG_LEVEL=debug  # Enable verbose logging for llama-swap
 
   llama-swap-amd:
-    build:
-      context: .
-      dockerfile: Dockerfile.llamaswap-rocm
+    image: ghcr.io/mostlygeek/llama-swap:rocm
     container_name: llama-swap-amd
     ports:
       - "8091:8080"  # Map host port 8091 to container port 8080
@@ -35,9 +33,6 @@ services:
     devices:
       - /dev/kfd:/dev/kfd
       - /dev/dri:/dev/dri
-    group_add:
-      - "985"  # video group
-      - "989"  # render group
     restart: unless-stopped
     healthcheck:
       test: ["CMD", "curl", "-f", "http://localhost:8080/health"]