sparkrun

Sparkrun is a command-line tool designed to launch, manage, and stop large language model (LLM) inference workloads on one or more NVIDIA DGX Spark systems. It eliminates the need for traditional cluster management solutions such as Slurm or Kubernetes, offering a streamlined process for deploying AI inference tasks across specialized hardware. ai.

Sparkrun features a recipe-based system where YAML configurations capture model, container, runtime, and default settings, allowing users to override parameters at launch without searching for configuration files. The tool manages container orchestration, model distribution, and networking automatically, supporting multi-node tensor parallelism where each DGX Spark contributes a GPU. It handles InfiniBand/RDMA and NCCL configuration for efficient scaling across nodes. Users can share and collaborate on recipes through git-based registries, including both community and private options, and search across all available registries.

cpp, and atlas, all accessible through the same CLI and recipe format. Sparkrun includes features like VRAM estimation, which auto-detects model architecture from HuggingFace to inform users whether a configuration fits on a single DGX Spark or requires multiple nodes before launching. Additionally, the tool offers AI-assisted inference management through the Claude Code Plugin, enabling conversational management for running, monitoring, and stopping workloads. Rich shell tab completion is available for Bash, Zsh, and Fish, providing instant completion for commands, recipe names, cluster names, and options.

Installation is handled via a single command that sets up the managed environment, configures tab completion, and initiates a setup wizard to guide cluster configuration. 0 license and is available on PyPI. The tool is intended for users operating NVIDIA DGX Spark systems who require efficient orchestration and management of LLM inference workloads.