v0.7.2 - 2026-10-14
NeMo Helix v0.7.2 is a patch release for the 0.7 release line that improves deployment readiness recovery and adds configurable agent readiness probes.
What’s fixed
- Deployment startup-timeout tracking. Deployments that lose readiness
after being ready get a fresh startup-timeout window, rather than failing
immediately based on their original startup time. Status-message updates
during an uninterrupted
STARTINGperiod no longer reset its timeout. - Configurable agent readiness probes. Operators can set
agents.deployments.readiness_probe_timeout_secondsandagents.deployments.readiness_probe_failure_thresholdto tune readiness checks for newly created agent containers. Defaults remain one second and twelve consecutive failures. Restart the platform and redeploy agents to apply changes. See Readiness and startup timeouts.
Known issues
- DPO jobs fail before the first training step with
KeyError: 'checkpoint'. The generated DPO configuration still omits thepolicy.dtensor_cfg.checkpointsection required by the pinned NeMo-RL version. There is no workaround in the 0.7 release line. DPO jobs run on v0.6.0; GRPO jobs are not affected. See Customizer troubleshooting. - Hermes trials can omit the
relay_atifevidence descriptor. Use thetrace:atifdescriptor instead; it points at the same ATIF file.