Skip to content

Train, serve and govern models on HPC as one loop

ExaMLOps runs the whole model lifecycle — train, register, promote, serve, observe, retrain — on the schedulers European HPC centres actually run: Slurm, Flux, or none at all. You operate it from the exa CLI, a web dashboard, or in plain English through the Skipper agent, all over the same code paths.

Explore the platform Start in five minutes

Every kind of work travels its own line through shared stations: a prediction on the data line, a retrain on the control line, a decision on the decision line, a cluster job on the compute line, and telemetry and evidence on the signal line. Open any station on the map for what it does.

Start here

If you want to… Go to
Get a stack running and train a model Quick Start
Understand how the pieces fit Architecture
Find the command for a task exa CLI — Command Guide
Drive the platform from a browser Dashboard Usage Guide
Talk to the platform in plain English Management Agent
Add your own model Add a New Model 
make bootstrap                                   # dev stack + all dependencies
exa status                                       # what is running, what is in production
exa pipeline run --model JPCP --dataset PM100Dataset --dummy

What it covers

  • Training pipelines — Prefect flows that auto-discover every registered model × dataset, with YAML-driven per-environment overlays.
  • HPC orchestration — a scheduler abstraction over Slurm, Flux and a mock backend, plus a fleet layer that discovers, approves, places and accounts for clusters.
  • Model lifecycle — MLflow registry with multi-stage aliases, metric-gated promotion, lineage, diffing, and signed artifacts.
  • Serving — Ray Serve multi-model routing with version-selectable inference, traffic splits and champion–challenger.
  • GenAI / LLMOps — an OpenAI-compatible gateway, prompt registry, RAG, vector store, guardrails, and evaluation with calibrated judges.
  • Observability — Prometheus, Grafana, Loki, Tempo tracing, drift detection (output and input-embedding), and a closed-loop autopilot.
  • Governance — a hash-chained audit trail, policy-as-code, secrets, supply-chain signing, and EU AI Act / NIST RMF mappings.
  • FinOps & Green AI — GPU-hour cost attribution and carbon accounting, both with swappable calculation providers.

Everything above is reachable from the CLI. Every capability lists all of it by lifecycle area, searchable; the command guide gives each command's purpose and a runnable example; the roadmap shows what is shipped and what is under construction.

Interfaces

ExaMLOps deliberately exposes the same capabilities four ways, over shared code paths:

Surface Entry point
Command line exa — command guide
Web Dashboard
Conversational exa chat — Skipper agent
Agent-callable exa mcp serve — MCP tools, resources and prompts

See All Interfaces for how they relate.