English translation is being prepared.

English translation is being prepared.

View original

SGLang v0.5.18 is out! We have 710 PRs from 212 contributors, and 61 of them were new contributors. Welcome🧡 Highlights: 🔸 New models: Muse Glimmer and Intern-S2-Mobius, plus SANA-Video, LTX-2.5, LingBot-Video, Cosmos3, and LongCat-Image on SGLang-Diffusion 🔸 Engine startup is up to 2.38x faster: checkpoints stream from disk while CUDA graphs capture, so the two longest waits happen at once 🔸 All compiled kernels (Triton, FlashInfer, Inductor, DeepGEMM) now live in one directory: SGLANG_CACHE_DIR 🔸 NVFP4 checkpoints now run on AMD GPUs via online MXFP4 requantization, with 97.5-100% accuracy retention across MiniMax-M2.7, GLM-5.1, Kimi-K2.6, Qwen3.5-397B, and DeepSeek-R1 🔸 Kimi K3 runs much faster on AMD MI355X: 1.37-1.77x throughput and up to 2.42x faster ITL 🔸 DeepSeek-V4 decode gets faster on NVIDIA: TP LMHead all-to-all cuts LMHead time from 320us to 169us, and FlashInfer MNNVL allreduce turns on automatically for DeepSeek models Thanks again to our amazing partners and model makers: @NVIDIAAI @AIatAMD @intel @AIatMeta @intern_lm @Lightricks @AntLingAGI @robbyant_brain @Meituan_LongCat Full release note👇