SGLang splits sparse attention memory across 8 GPUs with decode context parallelism

SGLang (@sgl_project) says it splits sparse attention's full-history memory across 8 GPUs using decode context parallelism.

The project says the method still selects the same 2,048 tokens and allows up to 8x more room for history with no accuracy loss. The claim comes from an X post, and the report includes no benchmark results.