# vLLM v0.9.1 — Release & Upgrade Intelligence
- **Product:** vLLM (https://vllm.ai)
- **Version:** v0.9.1
- **Released:** Jun 10, 2025
- **Impact Score:** 5/10
- **Breaking Changes:** Not verified
- **Canonical Analysis:** https://changelogs.us/vllm/updates/v0.9.1/

## Executive Summary (TL;DR)
vLLM v0.9.1. Source highlight: Progress in large scale serving

## New Features
- Add FlexAttention to vLLM V1 (#16078)
- Add DeepSeek-R1-0528 function call chat template (#18874)
- Add H20-3e fused MoE kernel tuning configs for DeepSeek-R1/V3 (#19205), Qwen3-235B-A22B (#19315)
- Blackwell: Add Cutlass MLA backend (#17625), Tunings for SM100 FP8 CUTLASS kernel (#18778), Use FlashInfer by default on Blackwell GPUs (#19118), Tune `scaled_fp8_quant` by increasing vectorization (#18844)
- FP4: Add compressed-tensors NVFP4 support (#18312), FP4 MoE kernel optimization (#19110)
- ROCm: Add AITER grouped topk for DeepSeekV2 (#18825)

## Improvements
- CPU: V1 support for the CPU backend (#16441)
- TPU: Initial support of model parallelism with single worker using SPMD (#18011), Multi-LoRA Optimizations for the V1 TPU backend (#15655)
- Platform: Make torch distributed process group extendable (#18763)
- [Core] Improve Tensor serialisation
- [CI] improve embed testing
- improve the robustness of parsing vlms config in AutoRound

## Bug Fixes
- DP Attention + Expert Parallelism: CUDA graph support (#18724), DeepEP dispatch-combine kernel (#18434), batched/masked DeepGEMM kernel (#19111), CUTLASS MoE kernel with PPLX (#18762)
- Update SECURITY.md with link to our security guide (#18961), Add security warning to bug report template (#19365)
- [CI/Build] [TPU] Fix TPU CI exit code
- [Bugfix] Disable prefix caching by default for benchmark
- [Build] Fixes for CMake install
- [rocm] Fix wrong attention log

## Upgrade Command
```bash
python -m pip install 'vllm==0.9.1'
```