vLLM v0.31.0 Released with DeepSeek Optimizations and Blackwell Architecture Support

vLLM v0.31.0 introduces major performance optimizations for DeepSeek models, native support for NVIDIA Blackwell (SM100/SM103) GPUs with MXFP8 and NVFP4 quantizations, and a new `vllm preload` weight-caching daemon for fast engine restarts. The release contains 717 commits from 307 contributors, featuring FlashMLA mega attention defaults, DeepGEMM integration, and large-scale serving backends like MoonEP. As vLLM is one of the most widely used open-source LLM serving engines, these hardware and kernel optimizations substantially lower latency and operational costs for running state-of-the-art MoE models on next-generation AI infrastructure. It accelerates the deployment of low-bit precision formats like FP4 and micro-scaled FP8 on enterprise hardware. The update features dynamic layer boundary fusion for DeepSeek models, async scheduling for speculative decoding in Model Runner V2, and improved KV cache scheduling controls to prevent memory deadlocks. It also introduces breaking changes, such as gating per-request multimodal kwargs and replacing online quantization options with standard shorthands like `fp8_per_tensor`.