•
vLLM 0.29.0 (~9 Sep 2026): Model Runner V2 default for all models—CUDA graph memory profiling, batch-sharded sampling (logits mem ~1/TP). MRV1 remains for some ROCm/unsupported features. Deprecated architectures removed; prefer vllm serve. Not transformers-5-17 Hub. Not investment advice.
vLLM 0.29.0 (~9 September 2026) is the after-work hook if you serve open-weight models on a GPU box—not a Transformers new-models checklist. Model Runner V2 (MRV2) is now the default for all models, ending the opt-in phase. Highlights called out in community notes: CUDA graph memory profiling and batch-sharded sampling (logits memory roughly 1/TP). MRV1 remains for some ROCm / unsupported feature paths. Deprecated architectures (Arctic, Chameleon, MPT, and others) are removed; prefer vllm serve over python -m vllm.entrypoints.openai.api_server. This note does not bundle the transformers-5-17 new-models Hub post (even if Hy4 appears in release chatter). Not investment advice.
The only course CTA is Introduction to Python Programming.
| Topic | 0.29 takeaway |
|---|---|
| Default runner | MRV2 for all models (upgrade flips the path) |
| Memory path | CUDA graph profiling + batch-sharded sampling → logits mem ~1/TP |
| Fallbacks | MRV1 still used where ROCm / features are unsupported |
| Removals | Deprecated architectures gone (Arctic, Chameleon, MPT, …) |
| Entry point | Prefer vllm serve <model>; old api_server module path deprecated |
Plain-English roundups: AI/TLDR v0.29.0, freedom.tech note.
pip install -U vllm
vllm serve <your-model>
1. Upgrade and confirm the serve path uses vllm serve (not the deprecated python -m …api_server).
2. Note TP / logits memory on one small model after MRV2 becomes default.
3. Grep configs for removed architectures (Arctic / Chameleon / MPT / …) before you bump production pins.
Stop after one of three.
Upgrade one local serve script to 0.29 + vllm serve, smoke-test a known-good model, then stop. Environment stuck? Introduction to Python Programming.