is there a easy way i can follow to deploy and run any hugging face model on my dgx spark?
Check out the playbooks at DGX Spark which walk through many common setups.
For example there’s one for running models using vllm in Docker:
https://build.nvidia.com/spark/vllm
https://build.nvidia.com/spark/vllm/instructions
https://build.nvidia.com/spark/vllm/agent-ready-models
The bottom page has some links to the vllm recipes that help you build a command, for ex:
There isn’t really one easy way, especially if you want good performance, in my experience after using my Spark for about 2 months. Some models will require hyper specific setups that are found buried deep within some repository with multiple “pinned” specific versions of some dependency that may get outdated later. There is a website with spark benchmarks that includes recipes you can try, but don’t always expect them to work:
Maybe just use this to then try to track down the obscure repos within which they reside.
For using VLLM, I would just grab GitHub - eugr/spark-vllm-docker: Docker configuration for running VLLM on dual DGX Sparks · GitHub as a baseline. For llama.cpp, you can follow the nvidia instructions I guess. They did work for me. llama.cpp just usually has worked very slowly for me on a spark at higher contexts.
There is also a lot of hubbub on here about getting DeepseekV4 Flash running on a single spark and that isn’t even using any of the common binaries. It’s using some hyper-specific custom built engine from antirez that’s literally only made for that purpose. You probably can’t throw a stone without encountering a DSV4F topic on this subforum right now so I’m not going to bother linking that.
I currently use this:
It’s pretty easy to run. Not the best for coding, but I like its general intelligence.
The easiest way is probably to just get one good agent running, link it to Hermes or Opencode and then just ask it to build a runtime for the model you want to run.
Latest commits to eugr’s spark-vllm-docker were made to make it more agent friendly.
In addition to the NVIDIA playbooks, you can also check out the OpenZeka Blog. We publish beginner-friendly, step-by-step DGX Spark tutorials covering different models, inference frameworks, UIs, and deployment scenarios, from single-Spark setups to 8-Spark clusters.
The tutorials include the exact Docker commands, model configurations, launch parameters, and expected terminal output at each step. Many also include the ready-to-use container images we built and tested, which makes reproducing the setups much easier.
For more detailed technical material, we also maintain OpenZeka Whitepapers, with deployment notes, benchmarks, cluster configurations, scaling tests, and performance analysis.
Both are updated regularly as we test new models and configurations on DGX Spark. If you run into any issues while trying one of the setups, feel free to reach out and we’ll be happy to help.