# vLLM v0.10.1 — Release & Upgrade Intelligence
- **Product:** vLLM (https://vllm.ai)
- **Version:** v0.10.1
- **Released:** Aug 18, 2025
- **Impact Score:** 7/10
- **Breaking Changes:** Source flagged
- **Canonical Analysis:** https://changelogs.us/vllm/updates/v0.10.1/

## Executive Summary (TL;DR)
vLLM v0.10.1. Source highlight: **NOTE: This release deprecates V0 FA3 support and as a result FP8 kv-cache in V0 may have issues**

## Breaking Changes & Deprecations
- Breaking change**: Removed AQLM quantization support (#22943) - users should migrate to alternative quantization methods.
- Important**: As part of the ongoing V0 engine cleanup, several breaking changes have been introduced:

## New Features
- **Dependency optimizations**: Removed xformers requirement for Mistral-format Pixtral and Mistral3 models (#21154), deprecation warnings added for old DeepGEMM version (#22194).
- [Feat] Allow custom naming of vLLM processes
- [Docs] Add Expert Parallelism Initial Documentation
- [Docs] Add `requirements/common.txt` to run unit tests
- Add H20-3e fused MoE kernel tuning configs for Qwen3-Coder-480B-A35B-Instruct
- [Frontend] Add request_id to the Request object so they can be controlled better via external load balancers

## Improvements
- **Enhanced existing models**: NemotronH support (#22349), Ernie 4.5 Base 0.3B model name change (#21735), GLM-4.5 series improvements (#22215), Granite models with fused MoE configurations (#21332) and quantized checkpoint loading (#22925), Ultravox support for Llama 4 and Gemma 3 backends (#17818), Mamba1 and Jamba model support in V1 (without CUDA graphs) (#21249)
- **CUDA graph performance**: Full CUDA graph support with separate attention routines, adding FA2 and FlashInfer compatibility (#20059), plus 6% end-to-end throughput improvement from Cutlass MLA (#22763).
- **Speculative decoding optimizations**: N-gram speculative decoding with single KMP token proposal algorithm (#22437), explicit EAGLE3 interface for enhanced compatibility (#22642).
- **Default behavior improvements**: Pooling models now default to chunked prefill and prefix caching (#20930), disabled chunked local attention by default for Llama4 for better performance (#21761).
- **Performance optimizations**: Improved startup time by disabling C++ compilation of symbolic shapes (#20836), enhanced headless models for pooling in Transformers backend (#21767).
- **NVIDIA Blackwell (SM100) optimizations**: CutlassMLA as default backend (#21626), FlashInfer MoE per-tensor scale FP8 backend (#21458), SM90 CUTLASS FP8 GEMM with kernel tuning and swap AB support (#20396).

## Bug Fixes
- **Hardware compatibility and kernels**: ARM CPU build fixes for systems without BF16 support (#21848), Machete memory-bound performance improvements (#21556), FlashInfer TRT-LLM prefill attention kernel support (#22095), optimized `reshape_and_cache_flash` CUDA kernel (#22036), CPU transfer support in NixlConnector (#18293).
- [Bugfix][ROCm] Fix for warp_size uses on host
- [TPU][Bugfix] fix moe layer
- [Misc] Fix duplicate FusedMoEConfig debug messages
- [Bugfix] Fix CUDA arch flags for MoE permute
- [Fix] Update mamba_ssm to 2.2.5

## Upgrade Command
```bash
python -m pip install 'vllm==0.10.1'
```