v0.3.3

Major changes

• StarCoder2 support
• Performance optimization and LoRA support for Gemma
• 2/3/8-bit GPTQ support
• Integrate Marlin Kernels for Int4 GPTQ inference
• Performance optimization for MoE kernel
• [Experimental] AWS Inferentia2 support
• [Experimental] Structured output (JSON, Regex) in OpenAI Server

What's Changed

• Update a comment in benchmark_serving.py by @ronensc in #2934
• Added early stopping to completion APIs by @Maxusmusti in #2939
• Migrate MistralForCausalLM to LlamaForCausalLM by @esmeetu in #2868
• Use Llama RMSNorm for Gemma by @WoosukKwon in #2974
• chore(vllm): codespell for spell checking by @mspronesti in #2820
• Optimize GeGLU layer in Gemma by @WoosukKwon in #2975
• [FIX] Fix issue #2904 by @44670 in #2983
• Remove Flash Attention in test env by @WoosukKwon in #2982
• Include tokens from prompt phase in counter_generation_tokens by @ronensc in #2802
• Fix nvcc not found in vllm-openai image by @zhaoyang-star in #2781
• [Fix] Fix assertion on Mistral YaRN model len by @WoosukKwon in #2984
• Port metrics from aioprometheus to prometheus_client by @hmellor in #2730
• Add LogProbs for Chat Completions in OpenAI by @jlcmoore in #2918
• Optimized fused MoE Kernel, take 2 by @pcmoritz in #2979
• [Minor] Remove gather_cached_kv kernel by @WoosukKwon in #3043
• [Minor] Remove unused config file by @esmeetu in #3039
• Fix using CuPy for eager mode by @esmeetu in #3037
• Fix stablelm by @esmeetu in #3038
• Support Orion model by @dachengai in #2539
• fix get_ip error in pure ipv6 environment by @Jingru in #2931
• [Minor] Fix type annotation in fused moe by @WoosukKwon in #3045
• Support logit bias for OpenAI API by @dylanwhawk in #3027
• [Minor] Fix StableLMEpochForCausalLM -> StableLmForCausalLM by @WoosukKwon in #3046
• Enables GQA support in the prefix prefill kernels by @sighingnow in #3007
• multi-lora documentation fix by @ElefHead in #3064
• Restrict prometheus_client >= 0.18.0 to prevent errors when importing pkgs by @AllenDou in #3070
• Support inference with transformers-neuronx by @liangfu in #2569
• Add LoRA support for Gemma by @WoosukKwon in #3050
• Add Support for 2/3/8-bit GPTQ Quantization Models by @chu-tianxiang in #2330
• Fix: AttributeError in OpenAI-compatible server by @jaywonchung in #3018
• add cache_config's info to prometheus metrics. by @AllenDou in #3100
• Support starcoder2 architecture by @sh0416 in #3089
• Fix building from source on WSL by @aliencaocao in #3112
• [Fix] Don't deep-copy LogitsProcessors when copying SamplingParams by @njhill in #3099
• Add guided decoding for OpenAI API server by @felixzhu555 in #2819
• Fix: Output text is always truncated in some models by @HyperdriveHustle in #3016
• Remove exclude_unset in streaming response by @sh0416 in #3143
• docs: Add tutorial on deploying vLLM model with KServe by @terrytangyuan in #2586
• fix relative import path of protocol.py by @Huarong in #3134
• Integrate Marlin Kernels for Int4 GPTQ inference by @robertgshaw2-neuralmagic in #2497
• Bump up to v0.3.3 by @WoosukKwon in #3129

New Contributors

• @Maxusmusti made their first contribution in #2939
• @44670 made their first contribution in #2983
• @jlcmoore made their first contribution in #2918
• @dachengai made their first contribution in #2539
• @dylanwhawk made their first contribution in #3027
• @ElefHead made their first contribution in #3064
• @AllenDou made their first contribution in #3070
• @jaywonchung made their first contribution in #3018
• @sh0416 made their first contribution in #3089
• @aliencaocao made their first contribution in #3112
• @felixzhu555 made their first contribution in #2819
• @HyperdriveHustle made their first contribution in #3016
• @terrytangyuan made their first contribution in #2586
• @Huarong made their first contribution in #3134

Full Changelog: v0.3.2...v0.3.3