DeepSeek-V4 on Day 0: From Fast Inference to Verified RL with SGLang and Miles
Summary
The LMSYS article announces Day-0 support for DeepSeek-V4, detailing open-source tooling (SGLang and Miles) for both inference and RL, and highlights innovations in hybrid sparse-attention, memory hierarchy, FP4 expert weights, and multi-GPU parallelism. It provides deep dives into architectural features (ShadowRadix, HiSparse, speculative decoding) and practical RL training workflows, plus kernel and deployment optimizations across modern accelerators. The piece positions DeepSeek-V4 as a scalable, open-stack solution for large-context models with robust hardware support.