Modelplane v0.4.0
The theme of this release is NVIDIA open source! Both headline features came out of work with the NVIDIA team.
You can now run a new Dynamo serving stack on any Modelplane-managed cluster. It brings gang scheduling with Grove and the KAI Scheduler, and GPU-to-GPU weight transfer with ModelExpress. The stack is a per-cluster choice, so a fleet can run Dynamo and Standard side by side, and the ModelDeployment an ML team writes is the same on either.
Modelplane also now builds much of each cloud's serving stack from NVIDIA AI Cluster Runtime, which locks down driver and operator versions validated on real hardware. One tested stack per cloud, and no version fields to get wrong.
Check out https://docs.modelplane.ai/getting-started/ to get started, and the v0.4 release blog post for the full walk-through.
Notable Changes
- Add a per-cluster serving stack: Standard or Dynamo by @negz in #406
- Compose the serving stack from build-time component lists by @haarchri in #428
- Activate cloud managed resources on demand from InferenceCluster by @negz in #429
What's Changed
- feat(main): add 0.3 for hugo by @haarchri in #392
- feat(main): update configuration url in docs by @haarchri in #395
- feat(bump): use latest provider-kubernetes and provider-helm by @haarchri in #390
- fix(cel): fix the member metadata CEL rules the API server rejects by @haarchri in #400
- feat(bump): set cfg to v0.3.1 by @haarchri in #402
- feat(docs): turn the examples sections in to browsable recips and guide section by @haarchri in #405
- Remove family providers as dep by @haarchri in #397
- Update flake.lock to the actual crossplane-cli v2.5.0 commit by @negz in #415
- Set the project on the composed ProjectIAMMember by @dennis-upbound in #413
- Docs: Derive API reference grouping from XRD categories by @haarchri in #431
- Design: generating the serving stack at build time by @haarchri in #422
- Skip the default ManagedResourceActivationPolicy in the dev control plane by @haarchri in #440
Full Changelog: v0.3.0...v0.4.0
