[mkaic]

Note

Code and PR description are LLM-written. If this PR has been marked Ready For Review, code has
been reviewed and tested in staging by Kai

Diff breakdown

Added and removed lines. Configuration changes count as altered logic.

Category	Lines
Genuinely new or altered logic	12
Moving or rearranging existing logic	0
Tests	0
Documentation	0
Generated file / lockfile churn	2

Description

Prepare inference-models==0.39.7.dev2026102801 for OWLv2 staging tests, with matching CPU, GPU, CUDA 13, Jetson, and OpenVINO pins. This PR targets kai/owlv2-upload and stacks on #3288, so the feature PR keeps its existing version. The lockfile changes only the local package version. Dependency versions stay unchanged.

Type of change

Release preparation.

How has this change been tested?

Package metadata and all five exact runtime pins agree. The offline lock consistency and diff checks passed. No test suites ran. Package publication, image builds, and staging validation remain pending.

Any specific deployment considerations

Kai handles publication and deployment. Keep this PR unmerged while #3288 is open to keep version changes out of the feature PR.

In Publish Inference Models Wheels to PyPi, select kai/owlv2-staging-release under Use workflow from and enable publish.

After publication succeeds, run Build and Push CPU Container and Build and Push GPU Container. Select this release branch for both workflows. Set the same unique custom_tag and enable force_push for both builds.

Use the full image tag, VERSION-custom_tag, from the build output.

Update both staging CPU and GPU tags in async-serverless/k8s/values/inference-versions.yaml on a separate deployment branch. In Deploy Async Serverless, select that branch, target staging: all US, and set components to:

consumer,cpu-direct

This revision supersedes the already published 0.39.7.dev20261028 package. That package and the owlv2-staging-1 images contain the removed separate-adapter implementation. Use custom_tag=owlv2-staging-2 for the new builds. This new version is not published by the agent.

Coordinate with Grzegorz before deploying: staging currently runs an unmerged rf-serving branch. The standard-image deployment branch would replace that configuration. Testing the v2 serving stack also requires releasing the model-manager change from #3288 with that stack.