Skip to content
ProjectZKMPublic

About

A unified scheduling framework for ZKM verifiable AI

Resources

Stars

22 stars

Watchers

1 watching

Forks

Repository files navigation

zkm-fleet

CI

Fleet schedules apps and batch jobs on shared GPU and CPU machines.

  • Apps keep containers running: a model server, or a Ziren prover with its program.
  • Batches run finite commands and release resources when cleanup finishes.

Architecture

Fleet architecture: apps and batches share machine resources

The control host runs the Rust scheduler, Postgres, an HTTPS gateway and, optionally, MCP. Each execution machine runs fleet-node and its own local Ray, and connects outward to register. Fleet handles admission, quotas, placement and lifecycle; apps own their data and pipelines. The code model follows Kubernetes' jobs, deployments and controllers, without needing Kubernetes.

Namespace owners create apps and batches and manage their priority, enabled state and scaling, within their namespace's and tag's quotas; the scheduler picks each app's node, and moves an app's start to a node with room when its own has none. Fleet copies nothing an app keeps on a node's disks, so an app that keeps data there, or runs from it, is anchored to that node with machine in its entry in the node's apps.yaml. Administrators manage namespaces, quotas, access, host storage for batches and apps' execution settings.

Two integrated app workers prove on their node's own GPU server:

  • EthProofs prepares witnesses, proves and reports every block in sequence, resuming from its saved cursor.
  • The GOAT proof builder serves Bitcoin node proof requests, proves its five circuits and keeps its task database and proofs in app storage.

Get started

  1. Install the control host and add a node.
  2. Inspect resources, scale apps and run batches.
  3. Adjust machine capacity, update and recover.

Documentation

Guide Covers
Installation Dependencies, binaries, control host, execution nodes and registration
Operations Namespace commands, tags, node capacity, app scaling, batches and logs
Maintenance Upgrades, Ray capacity changes, self-built programs, liveness and troubleshooting
Reference Architecture, workloads, proofs, access, quotas, APIs, storage and ports
EthProofs Pinned source, worker build, sequential proving and durable recovery
GOAT Pinned source, program bundle identity, packaging, cutover and recovery

About

A unified scheduling framework for ZKM verifiable AI

Resources

Stars

22 stars

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages