vllm Releases

The release initiates the cleanup of the V0 engine codebase by removing V0 CPU/XPU/TPU/HPU backends and various deprecated features while introducing new models and significant V1 engine improvements.

Jul 24, 2025

Minimal release notes: "Release PR https://github.com/neondatabase/neon/pull/12679. Diff with the previous release https://github.com/neondatab"

Jul 22, 2025

Minimal release notes: "Release PR https://github.com/neondatabase/neon/pull/12653. Diff with the previous release https://github.com/neondatab"

Jul 18, 2025

The final V0 release featuring Priority Scheduling in V1, embedding models support, Mamba2 in V1, and full CUDA-Graph execution for FlashAttention v3 and FlashMLA.

Jul 7, 2025

Focuses on large-scale serving with DP/EP support, RLHF workflow improvements, Hybrid Memory Allocator, and FlexAttention integration into V1.

Jun 10, 2025

A patch release containing a critical bugfix for DeepSeek family models on NVIDIA Ampere and below.

May 30, 2025

This release introduces a multi-tier KVCache offloading framework and advanced routing strategies to optimize GPU memory usage and deployment density.

May 21, 2025

Major upgrade to PyTorch 2.7, initial set of optimized kernels on NVIDIA Blackwell, and initial DP/EP/PD support for large-scale inference.

May 15, 2025

A post-release patch containing bugfixes for memory leaks and model accuracy issues.

May 2, 2025

Day 0 support for Qwen3, xgrammar structural tag feature, and important multi-modal bug fixes.

Apr 28, 2025

Includes important accuracy fixes for Llama4 models and adds support for Qwen3, Qwen3MoE, and xgrammar Enum support.

Apr 14, 2025

Day 0 Support for Llama 4 Scout and Maverick, V1 sliding window attention, and cluster scale serving features.

Apr 6, 2025

This release focuses on critical bug fixes for the V1 engine's memory usage and introduces new features like FP8 KV Cache support and the fastsafetensors loader.

Mar 23, 2025

A patch release addressing critical bug fixes for the V1 engine, including sampling parameter fixes and TPU support improvements.

Mar 19, 2025

A major release featuring the default enablement of the V1 engine, significant DeepSeek enhancements, and support for new models like Gemma 3 and Grok1.

Mar 18, 2025

Updates the rendering pipeline to support cameras with non-centered principal points and introduces texture mapping capabilities in FlameViewer.

Mar 7, 2025

This release highlights DeepSeek Multi-Token Prediction speedups and significant V1 engine features including LoRA, logprobs, and pipeline parallelism.

Feb 20, 2025

Introduces Qwen2.5-VL support, `transformers` backend support, and performance enhancements for DeepSeek models via MLA and FP8 kernels.

Feb 6, 2025

Focuses on MLA optimization for DeepSeek family models, V1 metrics and logging enhancements, and the addition of the MiniCPM-o model.

Feb 1, 2025

A patch release that restores quantization support for other MoE methods which was broken by the initial DeepSeek V3 support.

Dec 27, 2024