SGLang supported models: Diffusion language models

Read 2026-10-03. This is the strongest engine-side evidence that diffusion LMs need a DIFFERENT SERVING SHAPE rather than a different kernel: the page is a separate top-level docs section from 'Large Language Models', with its own launch flags (--dllm-algorithm LowConfidence or JointThreshold, --dllm-algorithm-config, --no-dllm-fdfo) and its own scheduling primitive, 'First-Done-First-Out (FDFO) Scheduling ... enabled by default: each request leaves the batch as soon as its block is resolved, instead of advancing in lockstep'. Supported set is FOUR families, none of them a repo model: LLaDA2.0 (mini, flash; LLaDA2.0-flash is a 100B MoE diffusion LM), SDAR/JetLM (SDAR-8B-Chat dense, SDAR-30B-A3B-Chat MoE), and DiffusionGemma (google/diffusiongemma-26B-A4B-it, uniform-state renoising block-diffusion multimodal MoE, 25.2B total / 3.8B active, served with the Gemma4Renoise sampler). LLaDA-8B-Instruct (the 8B dense LLaDA in this repo's models) is NOT in the table, nor is Dream 7B.