Skip to main content
The inference and clustering core of AINode is powered by vLLM with community patches. Specifically, we inherit the robust NCCL implementation from eugr/spark-vllm-docker (dgxspark-3node-ring branch), which has proven most reliable for Grace-Blackwell unified memory and fabric (RoCE / 400G switch) topologies. We are grateful to eugr and the broader vLLM / Ray / Spark community whose hard work on real hardware made this project possible.

Key dependencies

ProjectRole
eugr/spark-vllm-dockerGB10-patched vLLM containers, NCCL 3-node ring fix, cluster recipes
vLLMHigh-performance inference engine
RayCross-node distributed orchestration
NCCLGPU-to-GPU collective communication
Hugging FaceModel hub, Transformers, PEFT, Datasets
aiohttpAsync HTTP server for the API + web UI