Skip to content

DGX Spark

KITT includes specific detection and handling for NVIDIA DGX Spark systems. This guide covers what KITT does differently on DGX hardware and any special considerations for running benchmarks on these machines.

Environment Detection

KITT automatically identifies DGX Spark systems during hardware fingerprinting. The detection checks:

  1. /etc/dgx-release -- if present and contains "spark", the environment is classified as dgx_spark.
  2. /etc/nvidia/nvidia-dgs.conf -- presence of this file also triggers dgx_spark classification.
  3. If /etc/dgx-release exists but does not contain "spark", the environment is classified as dgx (standard DGX).

You can verify the detected environment with:

kitt fingerprint --verbose

The output includes the environment type (dgx_spark, dgx, native_linux, etc.) alongside GPU, CPU, RAM, and storage details.

GPU Detection on DGX Spark (GH200)

The DGX Spark uses the NVIDIA GH200 Grace Hopper Superchip, which has a unified memory architecture. This means standard VRAM queries may not return a meaningful value:

  • pynvml: nvmlDeviceGetMemoryInfo may fail or return zero on unified memory systems. KITT catches this and logs a debug message rather than crashing.
  • nvidia-smi: The memory column may report [N/A]. KITT handles this gracefully and sets VRAM to 0 GB in the fingerprint.

Despite the memory query limitations, GPU model name and compute capability detection work normally.

Docker on DGX

DGX systems ship with Docker and the NVIDIA Container Toolkit pre-installed. KITT uses --network host for all engine containers, which works out of the box on DGX. No additional Docker configuration is required.

If you are running KITT inside a container on DGX, make sure the container has access to the GPU:

docker run --gpus all --network host ...

Tested Environments

KITT is tested on the following DGX platforms:

Platform Environment Type
DGX Spark (GH200) dgx_spark
DGX Station / DGX A100 / DGX H100 dgx

Both environment types receive enhanced diagnostic messages when GPU detection fails. If KITT cannot find a GPU on a system classified as dgx_spark or dgx, it logs a warning with specific guidance to check that NVIDIA drivers are loaded and accessible.

Build and Deployment Notes

On DGX Spark, the GH200 has ARM (Grace) CPU cores. Make sure any Docker images you use are built for linux/arm64. The standard KITT engine images (vLLM, llama.cpp, Ollama) publish multi-architecture images that include ARM support. On the DGX Spark, native mode is the default for Ollama and llama.cpp, avoiding Docker overhead entirely.

When generating deployment stacks with kitt stack generate, the generated docker-compose.yaml does not pin a platform architecture, so Docker will automatically pull the correct image for the host.