v1.5.0 — swarm-robustness release
Swarm-robustness release — the fixes from the June 8-site runs that repeatedly died at rounds 7-10. A single node's transient drop no longer aborts the whole run, an aborted run resumes from a mirrored global, and GPU/VPN paths self-recover.
Highlights
- Fault-tolerant swarm controllers: prune-and-continue while >= min_clients remain (#346)
- Warm-continue / auto-resume from a per-round mirrored global on every client (#347)
- VPN auto-recovery: mediswarm-vpn systemd service + 30s tun0 health watchdog (#348)
- GPU-container watchdog + fix_docker_cgroupfs.sh, GPU no longer stripped by a daemon-reload (#343)
- 24h swarm timeouts + wait-for-all (#345); per-round prediction-export throttle (#314); DataLoader worker cap 8->min(cpu,16) (#315)
- NVFlare fork -> MediSwarm-2.7.2 incl. the CCWF config-phase quorum fix (configure_min_clients)
Image: jefftud/odelia:1.5.0-dev... Full notes in CHANGELOG.md.