vLLM v0.30.0 Released with DeepSeek-V4 Support and Fast Weight-Caching Daemon
vLLM version 0.30.0 introduces native support for new model architectures such as DeepSeek-V4.1-Flash and GLM-5.3-Flash alongside a host-resident HiSparse decoding tier. It also adds a persistent per-GPU weight-cache daemon using CUDA IPC that bypasses disk reads during engine restarts for significantly faster initialization. This release substantially reduces cold-start latency and engine initialization overhead for LLM serving infrastructure in production environments. Furthermore, enhanced support for sparse Multi-head Latent Attention (MLA) and low-precision data formats helps eliminate memory bandwidth bottlenecks on high-end GPUs like the NVIDIA H200 and Blackwell architectures. Freezing Python garbage collection during CUDA graph capture reduces engine initialization on H200 GPUs from 28.9 seconds to 8.2 seconds. The release also integrates MXFP8 KV-caching, Expert Parallelism Load Balancing (EPLB) for MoE routing, and dual-key Gumbel-max watermarking compatible with speculative decoding.
## BACKGROUND
vLLM is a popular open-source LLM inference engine designed for high throughput and efficient memory utilization via techniques like PagedAttention. Advanced MoE (Mixture-of-Experts) architectures and sparse MLA (Multi-head Latent Attention) formats rely on specialized CUDA kernels such as FlashMLA and dynamic load balancers like EPLB to maintain high throughput during large-scale deployment.