vLLM v0.28.0 raises the default batched token budget and moves bitsandbytes out of tree
- Sources: primary, discussion
- Summary: The vLLM v0.28.0 release notes list 584 commits. The release raises the default batched token budget to 16384 and moves bitsandbytes support out of the tree. Four features are removed or relocated in this release.
- Why it matters: Operators upgrading inherit new scheduling defaults and four removed or relocated features that need checking before rollout.