# vLLM v0.6.2 — Release & Upgrade Intelligence
- **Product:** vLLM (https://vllm.ai)
- **Version:** v0.6.2
- **Released:** Sep 25, 2024
- **Impact Score:** 7/10
- **Breaking Changes:** Source flagged
- **Canonical Analysis:** https://changelogs.us/vllm/updates/v0.6.2/

## Executive Summary (TL;DR)
vLLM v0.6.2. Source highlight: Support Llama 3.2 models (#8811, #8822)

## Breaking Changes & Deprecations
- ⚠️ You will see the following error now, this is breaking change!
- [Misc] Limit to ray[adag] 2.35 to avoid backward incompatible change by @ruisearch42 in https://github.com/vllm-project/vllm/pull/8509

## New Features
- Introduce `MQLLMEngine` for API Server, boost throughput 30% in single step and 7% in multistep (#8157, #8761, #8584)
- Add output streaming support to multi-step + async (#8335)
- Add flashinfer backend (#7928)
- Add cuda graph support during decoding for encoder-decoder models (#7631)
- [Gemma2] add bitsandbytes support for Gemma2
- [multi-step] add flashinfer backend

## Improvements
- Beam search have been soft deprecated. We are moving towards a version of beam search that's more performant and also simplifying vLLM's core. (#8684, #8763, #8713)
- Enhancements: pp for qwen2-vl (#8696), multiple images for qwen-vl (#8247), mistral function calling (#8515), bitsandbytes support for Gemma2 (#8338), tensor parallelism with bitsandbytes quantization (#8434)
- Multi-step scheduling enhancements
- Prompt logprobs support in Multi-step (#8199)
- [Frontend] Improve Nullable kv Arg Parsing
- [doc] improve installation doc

## Bug Fixes
- Fix the AMD weight loading tests
- [Bugfix]: Fix the logic for deciding if tool parsing is used
- [Bugfix] Fix InternVL2 inference with various num_patches
- [BugFix] lazy init _copy_stream to avoid torch init wrong gpu instance
- [BugFix] Fix Duplicate Assignment of Class Variable in Hermes2ProToolParser
- [Bugfix] Offline mode fix

## Upgrade Command
```bash
python -m pip install 'vllm==0.6.2'
```