Tools
SGLang 0.5.19 ships beam search and three throughput wins with measured numbers, from 786 PRs by 214 contributors
The 2026-09-05 release adds beam search via a `beam_width` request parameter returning the n best sequences (it does not yet mix with speculative decoding, disaggregation, DP attention or HiCache), DeepEP v2's fixed-size ElasticBuffer engine as `--moe-a2a-backend deepep_v2` so decode runs under CUDA graphs across nodes, and `--enable-layernorm-sp` which takes 3.5% off Qwen3-8B prefill on H100 and 5.6% on B200. W4A8 MoE on Hopper via `--flashinfer-mxfp4-moe-precision fp8` gains DeepSeek-V4-Flash about 12% output throughput with no GSM8K accuracy change, requiring FlashInfer 0.6.18.
Source
↳ Follow the thread