Performance Improvements

• LTX-2 model series (one-stage and two-stage pipelines)
• low precision (nvfp4, mxfp4, fp8)
• graph-level optimization
• optimizations on consumer-level GPUs
• kernel fusion instead of compiler-driven fusion

Features

• Improve performance for diffusers backend @DefTruth
• Serve Encoders with SGLang [Feature] Serve LLMs via SGLang runtime as encoders in multimodal generation pipelines #20032
• Batching support (dynamic batching, continuous batching) @qimcis
• Distillation support
• Better integration with RL infrastructure

Disaggregation (already supported)

• Fully disaggregated pipeline, with mooncake as transfer engine and compatibility with all models
• Extendability: make it easier to adjust and define the boundaries for disaggregated pipeline stages
• Hybrid Parallelism: support complex and advanced parallelism within pipeline stages

Model support

• image, video and 3D models
• real-time video models
• ar + diffusion hybrid model (call for help) [diffusion] model: support ar + diffusion hybrid model #23578

Community Contributions

We welcome all forms of community contribution — from bug reports and documentation improvements
to kernel optimizations, model integration, and new feature ideas.

We will be actively posting tasks in Slack channel: #collab-diffusion, feel free to take them over

If you're interested in participating, please check:

• Slack channel: #diffusion
• Slack dev channel: #collab-diffusion
• Cookbook for SGLang-Diffusion
• Documentation for SGLang-Diffusion
• SGLang-Diffusion: Two Months In

Your improvements — no matter small or large — can help make SGLang-Diffusion serving faster, easier, and more versatile.