v0.2.3

Major changes

• Refactoring on Worker, InputMetadata, and Attention
• Fix TP support for AWQ models
• Support Prometheus metrics
• Fix Baichuan & Baichuan 2

What's Changed

• Add instructions to install vllm+cu118 by @WoosukKwon in #1717
• Documentation about official docker image by @simon-mo in #1709
• Fix the code block's format in deploying_with_docker page by @HermitSun in #1722
• Migrate linter from pylint to ruff by @simon-mo in #1665
• [FIX] Update the doc link in README.md by @zhuohan123 in #1730
• [BugFix] Fix a bug in loading safetensors by @WoosukKwon in #1732
• Fix hanging in the scheduler caused by long prompts by @chenxu2048 in #1534
• [Fix] Fix bugs in scheduler by @linotfan in #1727
• Rewrite torch.repeat_interleave to remove cpu synchronization by @beginlner in #1599
• fix RAM OOM when load large models in tensor parallel mode. by @boydfd in #1395
• [BugFix] Fix TP support for AWQ by @WoosukKwon in #1731
• [FIX] Fix the case when input_is_parallel=False for ScaledActivation by @zhuohan123 in #1737
• Add stop_token_ids in SamplingParams.repr by @chenxu2048 in #1745
• [DOCS] Add engine args documentation by @casper-hansen in #1741
• Set top_p=0 and top_k=-1 in greedy sampling by @beginlner in #1748
• Fix repetition penalty aligned with huggingface by @beginlner in #1577
• [build] Avoid building too many extensions by @ymwangg in #1624
• [Minor] Fix model docstrings by @WoosukKwon in #1764
• Added echo function to OpenAI API server. by @wanmok in #1504
• Init model on GPU to reduce CPU memory footprint by @beginlner in #1796
• Correct comments in parallel_state.py by @explainerauthors in #1818
• Fix OPT weight loading by @WoosukKwon in #1819
• [FIX] Fix class naming by @zhuohan123 in #1803
• Move the definition of BlockTable a few lines above so we could use it in BlockAllocator by @explainerauthors in #1791
• [FIX] Fix formatting error in main branch by @zhuohan123 in #1822
• [Fix] Fix RoPE in ChatGLM-32K by @WoosukKwon in #1841
• Better integration with Ray Serve by @FlorianJoncour in #1821
• Refactor Attention by @WoosukKwon in #1840
• [Docs] Add information about using shared memory in docker by @simon-mo in #1845
• Disable Logs Requests should Disable Logging of requests. by @MichaelMcCulloch in #1779
• Refactor worker & InputMetadata by @WoosukKwon in #1843
• Avoid multiple instantiations of the RoPE class by @jeejeeli in #1828
• [FIX] Fix docker build error (#1831) by @allenhaozi in #1832
• Add profile option to latency benchmark by @WoosukKwon in #1839
• Remove max_num_seqs in latency benchmark by @WoosukKwon in #1855
• Support max-model-len argument for throughput benchmark by @aisensiy in #1858
• Fix rope cache key error by @esmeetu in #1867
• docs: add instructions for Langchain by @mspronesti in #1162
• Support chat template and echo for chat API by @Tostino in #1756
• Fix Baichuan tokenizer error by @WoosukKwon in #1874
• Add weight normalization for Baichuan 2 by @WoosukKwon in #1876
• Fix the typo in SamplingParams' docstring. by @xukp20 in #1886
• [Docs] Update the AWQ documentation to highlight performance issue by @simon-mo in #1883
• Fix the broken sampler tests by @WoosukKwon in #1896
• Add Production Metrics in Prometheus format by @simon-mo in #1890
• Add PyTorch-native implementation of custom layers by @WoosukKwon in #1898
• Fix broken worker test by @WoosukKwon in #1900
• chore(examples-docs): upgrade to OpenAI V1 by @mspronesti in #1785
• Fix num_gpus when TP > 1 by @WoosukKwon in #1852
• Bump up to v0.2.3 by @WoosukKwon in #1903

New Contributors

• @boydfd made their first contribution in #1395
• @explainerauthors made their first contribution in #1818
• @FlorianJoncour made their first contribution in #1821
• @MichaelMcCulloch made their first contribution in #1779
• @jeejeeli made their first contribution in #1828
• @allenhaozi made their first contribution in #1832
• @aisensiy made their first contribution in #1858
• @xukp20 made their first contribution in #1886

Full Changelog: v0.2.2...v0.2.3