SGLang v0.5.18 adds diffusion models, faster startup, and AMD NVFP4 support

SGLang v0.5.18, published on August 22, 2026, includes 710 pull requests from 212 contributors. The release adds support for several new model families, including diffusion models such as SANA-Video, LingBot-Video-MoE, LTX-2.5, Cosmos3 Edge and Distilled, and LongCat-Image, plus autoregressive models Muse Glimmer and Intern-S2-Mobius. Cookbook recipes were added for the Qwen3.8 family, Ling-3.0, Nemotron 3.5 Lightning, Dots3-Note, and DeepSeek-V4-Pro-0813.

Faster startup and decode

Checkpoint staging now overlaps with CUDA graph capture at startup. With the opt-in flag named startup-weight-load-mode overlap, Qwen3-32B on H100 starts 8.6 to 11.7 percent faster than serial with prefetch, and 2.38 times faster (35.6 seconds versus 84.8 seconds) than the plain default. This change lets checkpoint pages stage from storage while CUDA graphs capture.

For decode, the TP LMHead path replaces allgather plus scatter with a single all-to-all operation for pure data-parallel dp-attention. On DeepSeek-V4-Pro with B200, LMHead time drops from 320 microseconds to 169 microseconds, and TPOT improves from 36.97 milliseconds to 35.67 milliseconds.

FlashInfer MNNVL now supports pure allreduce sites, reusing the MNNVL workspace instead of falling back to NCCL. On DeepSeek-V4-Flash with TP4 decode on Blackwell, this gains up to 6.9 percent at small batches. It is auto-enabled for DeepSeek-V3, V3.2, and V4, and available elsewhere with the flag enable-flashinfer-pure-allreduce.

AMD and Kimi K3 tuning

NVFP4 checkpoints can now run on AMD GPUs. The quantization mode quark_mxfp4 dequantizes ModelOpt and Quark NVFP4 weights and requantizes to MXFP4 at load, without holding a full-precision copy. GSM8K recovery is 97.5 to 100.2 percent versus the NVFP4 reference across MiniMax-M2.7, GLM-5.1, Kimi-K2.6, Qwen3.5-397B, and DeepSeek-R1.

Kimi K3 gets tuned kernels for the MI355X. A grouped-head MLA verify kernel replaces the MHA-shaped split-KV path that re-read the shared latent once per head, delivering 1.37 to 1.77 times throughput and 1.45 to 2.42 times ITL at concurrency 2 to 32. The AITER MLA prefill kernel now accepts K3's 12-head shape, with TTFT improvements up to 14.9 percent. A gfx950-tuned decode stage-1 geometry adds 46 to 73 percent ITL at 68k input, opt-in via the environment variable SGLANG_MLA_DECODE_TUNE set to 1. GSM8K holds at 0.951 to 0.957.

Cache and dependencies

All compiled-kernel caches, including Triton, FlashInfer, Inductor, DeepGEMM, and CUDA driver caches, now live under a single directory specified by SGLANG_CACHE_DIR. The first launch after upgrading recompiles once. Dependencies include torch 2.13.0 with triton 3.7.1, flashinfer 0.6.17, CuTeDSL 4.6.2 (fixing an FA4 startup regression on Blackwell), DeepEP installed from released sgl-deep-ep wheels, and sgl-kernel 0.4.6.post1.

Other notable fixes include speculative decoding crashes with DP-Attention, silent KV corruption when speculative draft tokens exceed 4, and a bug where stop or EOS inside a spec accept run beat the max_new_tokens finish. The release also includes a fix for DFlash sliding attention causality defaults and a budget for the DFLASH draft KV pool based on its own attention geometry.