vLLM v0.30.0 Release: Adds Support for DeepSeek-V4.1-Flash with MXFP8 KV Storage

Core summary

One sentence to understand this update

vLLM version 0.30.0 has been released, featuring 762 commits and adding support for new models like DeepSeek-V4.1-Flash, which utilizes MXFP8 for KV storage via FlashMLA V4.1 on SM100.

Impact & opportunity

What this could mean

Developers using vLLM can now leverage the DeepSeek-V4.1-Flash model for more efficient inference, potentially leading to performance improvements and cost savings for large language model deployments.