Extract triples with Ollama or vLLM, store them in a graph database, and explore them in a GPU-accelerated web UI
The Hardware platform column shows where an issue is most relevant. "All hardware platforms" applies to every supported platform.
| Symptom | Hardware platform | Cause | Fix |
|---|---|---|---|
| Ollama performance issues | All hardware platforms | Suboptimal Ollama settings | Set environment variables: OLLAMA_FLASH_ATTENTION=1, OLLAMA_KEEP_ALIVE=30m, OLLAMA_MAX_LOADED_MODELS=1, OLLAMA_KV_CACHE_TYPE=q8_0 |
| Memory pressure when switching Ollama models | DGX Spark | Unified memory buffer cache not released | Flush buffer cache (see UMA note below) |
| VRAM exhausted or memory pressure | DGX Station | GPU memory fragmentation | Clear GPU memory: nvidia-smi --gpu-reset or restart Docker containers |
| Slow triple extraction | All hardware platforms | Large model or large context window | Reduce document chunk size or use a faster model |
| ArangoDB connection refused | All hardware platforms (Ollama stack) | Service not fully started | Wait ~30s after ./start.sh, then verify with docker ps |
ArangoDB exits with <jemalloc>: Unsupported system page size | DGX Station systems with 64 KB page-size kernels | The upstream ArangoDB image may include jemalloc built for smaller pages | Use ./start.sh --neo4j for Neo4j + Ollama |
Qdrant exits or crash-loops with <jemalloc>: Unsupported system page size | DGX Station systems with 64 KB page-size kernels using --vector-search | The upstream Qdrant image may include jemalloc built for smaller pages | Leave vector search disabled, or retry after the Qdrant image is updated for 64 KB pages |
| Container fails to start with GPU error | All hardware platforms | NVIDIA Container Toolkit not configured | Run nvidia-ctk runtime configure --runtime=docker and restart Docker |
| Port already in use | All hardware platforms | Previous instance still running | Run ./stop.sh (with the same stack flags) or docker compose down |
| Need another graph or LLM stack | All hardware platforms | Default stack is not the one you want | Use ./start.sh --neo4j for Neo4j + Ollama or ./start.sh --vllm for Neo4j + vLLM |
| vLLM takes long to become ready | All hardware platforms (vLLM stack) | Model load can take 30+ minutes | The UI may show an initializing banner while the model loads. Check progress: docker logs vllm-service -f |
The 64 KB page-size rows above apply to affected DGX Station systems only; DGX Spark is not affected.
NOTE
Unified memory (UMA). On hardware platforms with unified memory, GPU and CPU share memory dynamically. Some applications have not yet been updated for UMA, so you may hit memory issues even within capacity. If that happens, manually flush the buffer cache:
sudo sh -c 'sync; echo 3 > /proc/sys/vm/drop_caches'
NOTE
Model size vs. memory. Larger models generally improve triple quality. If you hit memory limits, reduce context window size, use a quantized variant, or choose a smaller model for your hardware platform.