v0.17.0

The vLLM project, maintained by the open-source community, has launched version 0.17.0, marking one of its most significant updates with 699 commits from 272 contributors. This release is headlined by the integration of FlashAttention 4, a next-generation attention backend promising substantial performance gains for transformer models. It also delivers comprehensive support for the Qwen3.5 model family, including its Gated Delta Networks (GDN) architecture, FP8 quantization, and speculative decoding. The update introduces a new `--performance-mode` flag with 'balanced', 'interactivity', and 'throughput' presets, simplifying deployment optimization for common scenarios like chatbots or batch processing.

The technical foundation is upgraded to PyTorch 2.10.0, a breaking change requiring environment updates. A major milestone is the maturation of the Model Runner V2 architecture, now supporting pipeline parallelism and decode context parallelism. For scaling large models, the release introduces initial support for elastic expert parallelism, allowing dynamic addition/removal of GPUs for Mixture-of-Experts (MoE) models. Other notable additions include weight offloading with prefetching to hide latency, direct loading of quantized LoRA adapters (QLoRA), and enhanced Anthropic API compatibility for thinking blocks and tool use. The release expands model support to new architectures like Ring 2.5 and Ovis 2.6, plus ASR and multimodal models, solidifying vLLM's position as a versatile, high-performance inference engine for production AI workloads.