config(ollama): reduce VRAM usage for GPU sharing with other services

Change keep-alive from infinite to 5m and max loaded models from 2 to 1,
freeing VRAM for Stable Audio, TRELLIS, and other GPU services.

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

stacks/ollama.yml

@@ -24,10 +24,10 @@ services:
       - NVIDIA_DRIVER_CAPABILITIES=all
       # Process requests sequentially to avoid batch overflow panics
       - OLLAMA_NUM_PARALLEL=1
-      # Keep models loaded in VRAM (don't unload after idle)
-      - OLLAMA_KEEP_ALIVE=-1
-      # Load multiple models concurrently (mistral-nemo + nomic-embed-text)
-      - OLLAMA_MAX_LOADED_MODELS=2
+      # Unload LLMs after 5 minutes idle (keeps VRAM free for other services)
+      - OLLAMA_KEEP_ALIVE=5m
+      # Only keep one model loaded at a time (embedding model stays, LLMs swap)
+      - OLLAMA_MAX_LOADED_MODELS=1
     healthcheck:
       test: ["CMD-SHELL", "curl -fSs http://localhost:11434/api/tags || exit 1"]
       interval: 30s