Brief
vLLM v0.29.0 makes Model Runner V2 the default for all models, with a ROCm exception
vLLM's v0.29.0 release notes switch Model Runner V2 to the default execution path for all models and add per-request speculative-decoding metrics. The notes are a changelog, not a migration guide, so the operational reading below rests on that single document and is flagged accordingly.
The release notes report 594 commits from 277 contributors, 91 of them new. The change with the widest operational footprint is the default: Model Runner V2 is now the default for all models, which the notes describe as completing the rollout that began with pooling models. MRV1 remains in use for a few ROCm models and features MRV2 does not yet support.
Two further items bear on running the server. CUDA graph memory profiling is added for KV cache auto-sizing, and batch-sharded sampling is said to cut per-step logits memory by 1/TP. Separately, per-request acceptance statistics can be exposed in OpenAI API responses via --per-request-spec-decode-metrics.
Our reading
Anyone running vLLM as a serving step just had the default execution path change under the same package version, which can move memory headroom, latency and throughput without any configuration edit on their side. The per-request spec-decode metrics flag is the more directly useful addition for teams that already track speculative-decoding yield, because it turns acceptance into something a dashb…
What to do or watch
The notes identify neither the affected ROCm models nor a migration test, so the precise unresolved question is which of your served models still fall under MRV1; check that against the exception before upgrading and compare KV cache sizing and per-step latency on your own hardware, watching the new spec-decode metrics for changed acceptance.
Source details and supporting facts
Each line is stated by the page named above it.
Stated by github.com
- The release features 594 commits from 277 contributors, 91 of them new.
- Model Runner V2 is now the default for all models, completing the rollout that began with pooling models.
- MRV1 remains in use for a few ROCm models and features MRV2 does not yet support.
- Batch-sharded sampling cuts per-step logits memory by 1/TP.
- Per-request acceptance stats can be exposed in OpenAI API responses via --per-request-spec-decode-metrics.
Sources
- vLLM releasesText stored 16 September 2026
How this story was checked. Written from the 1 page listed above, stored 16 September 2026; claims checked against that stored text on 16 September 2026.
What that means
- 5 of 5 reported statements were confirmed against the page that carries them; the rest were removed rather than published.
- Figures in the text were required to appear in the stored source text: yes. Identifiers: yes.
- The check reads stored text only: no claim rests on a fresh look that did not happen.
- Where the reporting was silent, the text says so instead of filling the gap.