Speaker
Description
Training and tuning ML models for the EIC needs computation spread across grid, HPC and cloud sites, and running that work takes a lot of operator effort. This talk shows how PanDA and iDDS run and increasingly automate distributed ML workflows. iDDS describes training and hyperparameter optimization (HPO) campaigns as workflows of dependent steps, and PanDA runs those steps on GPU resources at many sites. PanDA MCP, a Model Context Protocol server, gives LLM agents permission-controlled access to PanDA and iDDS. Agents built on it monitor jobs, diagnose failures from logs and source code, and propose fixes for a human to approve. Agentic warm start lets new HPO campaigns begin from the best trials of similar past jobs stored in a knowledge base, so they don't waste their first trials on random points. We close with lessons learned and how the EIC community can adopt these tools.