# vLLM v0.4.3 — Release & Upgrade Intelligence
- **Product:** vLLM (https://vllm.ai)
- **Version:** v0.4.3
- **Released:** Jun 01, 2024
- **Impact Score:** 5/10
- **Breaking Changes:** Not verified
- **Canonical Analysis:** https://changelogs.us/vllm/updates/v0.4.3/

## Executive Summary (TL;DR)
vLLM v0.4.3. Source highlight: Added support for Falcon (#5069)

## New Features
- Added support for Falcon (#5069)
- Added support for IBM Granite Code models (#4636)
- Added blocksparse flash attention kernel and Phi-3-Small model (#4799)
- Added Snowflake arctic model implementation (#4652, #4889, #4690)
- Add base class for vision-language models (#4809)
- Add fused_moe Triton configs (#4951)

## Improvements
- Improve n-gram efficiency (#4724)
- Marlin prefill performance improvement (about better on average) (#4983)
- Improve FP8 linear layer performance (#4691)
- improve paccess check (#4992)
- [Kernel] [FP8] Improve FP8 linear layer performance
- [CORE] Improvement in ranks code

## Bug Fixes
- [Bugfix] Fix `asyncio.Task` not being subscriptable
- [Bugfix] Fixed error in slice_lora_b for MergedQKVParallelLinearWithLora
- [Bug fix][Core] fixup ngram not setup correctly
- [Bugfix][Kernel] allow non-power-of-2 for prefix prefill with alibi
- [CI/Test] fix swap test for multi gpu
- [Speculative decoding] [Bugfix] Fix overallocation in ngram + spec logprobs

## Upgrade Command
```bash
python -m pip install 'vllm==0.4.3'
```