Skip to main content

HPC-MCP: LLM-Assisted HPC Utility Server

HPC-MCP is a service-based ICICLE software component that gives HPC users a natural-language interface to backend HPC utilities. The current release provides distributed LLM training-time prediction for Vista and Perlmutter through HTTP endpoints and MCP clients.

GitHub Repo License: MIT

References​

  • Deployed service: https://hpcmcpservercpu.pods.icicleai.tapis.io
  • MCP endpoint: https://hpcmcpservercpu.pods.icicleai.tapis.io/mcp
  • OpenAPI specification: docs/openapi.json
  • Live OpenAPI specification: https://hpcmcpservercpu.pods.icicleai.tapis.io/openapi.json
  • Component metadata file: component.yaml
  • Tapis Pod example: tapis-pod.example.json
  • Upstream estimator: https://github.com/ICICLE-ai/distributed_training_estimator_of_LLM

Acknowledgements​

National Science Foundation (NSF) funded AI institute for Intelligent Cyberinfrastructure with Computational Learning in the Environment (ICICLE) (OAC 2112606)

Issue reporting​

Report bugs, documentation issues, or release questions through GitHub Issues:

https://github.com/ICICLE-ai/hpc-mcp-server/issues

Release and ICICLE Training Catalog contacts:

  • Carlos Guzman: guzman.109@osu.edu
  • Amit Vyas: vyas.154@osu.edu