SGLangProduct update
SGLang · v0.5.23
SGLang v0.5.23 adds several new LLMs, VLMs, and diffusion models and speeds up DeepSeek-class workloads. On GB300, DeepSeek-V4.1 can reach about 1.5× prefill throughput and roughly 28% lower time to first token; optional NVIDIA megakernel and AMD monokernel plugins fuse decode into persistent kernels.
Paged Experts can run oversized MoE models on a single consumer GPU, and the release adds Google TPU7x support plus LoRA on quantized MoE.
DeepSeek-V4.1 on GB300 gets up to 1.5x more prefill throughput under load and about 28% lower first-token latency on 32K to 64K prompts.
Saved . A shortened excerpt from GitHub Releases; the original publisher retains authorship.