SGLang

Facts

license

Apache-2.0source, accessed 2026-08-28

hosted by

hosted under the non-profit open-source organization LMSYS; joined the PyTorch Ecosystem in March 2025source, accessed 2026-08-28

kv cache reuse

RadixAttention, which keeps a radix tree of prefixes so unrelated requests can share KV cache blockssource, accessed 2026-08-28

paper throughput claim

up to 6.4x higher throughput than the inference systems it was compared against at publicationsource, accessed 2026-08-28

deployment claim

the project's own README claims deployments on over 400,000 GPUs; the figure is self-reported and not independently verified heresource, accessed 2026-08-28