# SGLang Diffusion：加速视频与图像生成 - 来源：LMSYS：Blog（Chatbot Arena 团队） - 发布时间：2025-11-07 00:00 - AIHOT 链接：https://aihot.virxact.com/items/cmnxbjke5006zsln0u1opdfy9 - 原文链接：https://www.lmsys.org/blog/2025-11-07-sglang-diffusion ## AI 摘要 SGLang 推出 Diffusion 推理引擎，将高性能 LLM 服务架构扩展至视频与图像生成领域。该系统支持 Wan、Hunyuan、Qwen-Image、Flux 等主流开源扩散模型，在 H100/H200 GPU 上实现 1.2-5.9 倍加速。通过 ComposedPipelineBase 架构抽象与统一序列并行技术，提供 OpenAI 兼容 API、CLI 及 Python 接口，并与 FastVideo 合作覆盖从训练到部署的全流程。 ## 正文 Contents Why Diffusion in SGLang? Architecture Model Support Usage Install CLI Demo Text to Video: Wan-AI/Wan2.1 Image to Video: Wan-AI/Wan2.1-I2V Text to Image: FLUX Text to Image: Qwen-Image Image to Image: Qwen-Image-Edit Performance Benchmark Roadmap and Diffusion Ecosystem Acknowledgment Learn more SGLang Diffusion: Accelerating Video and Image Generation We are excited to introduce SGLang Diffusion, which brings SGLang's state-of-the-art performance to accelerate image and video generation for diffusion models. SGLang Diffusion supports major open-source video and image generation models (Wan, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux) while providing fast inference speeds and ease of use via multiple API entry points (OpenAI-compatible API, CLI, Python interface). SGLang Diffusion delivers 1.2x - 5.9x speedup across diverse workloads. In collaboration with the FastVideo team, we provide a complete ecosystem for diffusion models, from post-training to production serving. The code is available here. Documentation is available here. SGL Diffusion Performance Benchmark on an H100 GPU SGLang Diffusion Performance Benchmark on an H200 GPU Why Diffusion in SGLang? With diffusion models becoming the backbone for state-of-the-art image and video generation, we have heard strong community demand for bringing SGLang's signature performance and seamless user experience to these new modalities. We built SGLang Diffusion to answer this call, providing a unified, high-performance engine for both language and diffusion tasks. This unified approach is crucial, as the future of generation lies in combining architectures. Pioneering models are already fusing the strengths of autoregressive (AR) and diffusion-based approaches—from models like ByteDance's Bagel and Meta's Transfusion that use a single transformer for both tasks, to NVIDIA's Fast-dLLM v2 which adapts AR models for parallel generation. SGLang Diffusion is designed to be a future-proof, high-performance solution ready to power these innovative systems. Architecture SGLang Diffusion is engineered for both performance and flexibility, built upon SGLang's battle-tested serving architecture. It inherits the powerful SGLang scheduler and reuses highly-optimized sgl-kernel for maximum efficiency. At its core, our architecture is designed to accommodate the diverse structures of modern diffusion models. We introduce ComposedPipelineBase, a flexible abstraction that orchestrates a series of modular PipelineStages. Each stage encapsulates a common diffusion function—such as the denoising loop in DenoisingStage or VAE decoding in DecodingStage—allowing developers to easily combine and reuse these components to construct complex, customized pipelines. ComposedPipelineBase PipelineStage DenoisingStage DecodingStage To achieve state-of-the-art speed, we integrate advanced parallelism techniques. It supports Unified Sequence Parallelism (USP)—a combination of Ulysses-SP and Ring-Attention—for the core transformer blocks, alongside CFG-parallelism and tensor parallelism (TP) for other model components. To accelerate development and foster a powerful ecosystem, our system is built on an enhanced fork of FastVideo, and we are collaborating closely with their team. This partnership allows SGLang Diffusion to focus on delivering cutting-edge inference speed, while FastVideo provides comprehensive support for training-related tasks like model distillation. Model Support We support various popular open-source video & image generation models, including: Video models: Wan-series, FastWan, Hunyuan Image models: Qwen-Image, Qwen-Image-Edit, Flux For full list of supported models, reference here. Usage For a seamless user experience, we provide a suite of familiar interfaces, including a CLI, a Python engine API, and an OpenAI-compatible API, allowing users to integrate diffusion generation into their workflows with minimal effort. Install SGLang Diffusion can be installed via multiple ways: # with pip or uv uv pip install 'sglang[diffusion]' --prerelease=allow # from source git clone https://github.com/sgl-project/sglang.git cd sglang uv pip install -e "python[diffusion]" --prerelease=allow # with pip or uv uv pip install 'sglang[diffusion]' --prerelease=allow # from source git clone https://github.com/sgl-project/sglang.git cd sglang uv pip install -e "python[diffusion]" --prerelease=allow CLI Launch a server and then send requests: sglang serve --model-path black-forest-labs/FLUX.1-dev --port 3000 curl http://127.0.0.1:3000/v1/images/generations \ -o >(jq -r '.data[0].b64_json' | base64 --decode > example.png) \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $OPENAI_API_KEY" \ -d '{ "model": "black-forest-labs/FLUX.1-dev", "prompt": "A cute baby sea otter", "n": 1, "size": "1024x1024", "response_format": "b64_json" }' sglang serve --model-path black-forest-labs/FLUX.1-dev --port 3000 curl http://127.0.0.1:3000/v1/images/generations \ -o >(jq -r '.data[0].b64_json' | base64 --decode > example.png) \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $OPENAI_API_KEY" \ -d '{ "model": "black-forest-labs/FLUX.1-dev", "prompt": "A cute baby sea otter", "n": 1, "size": "1024x1024", "response_format": "b64_json" }' Or, generate an image without launching a server: sglang generate --model-path black-forest-labs/FLUX.1-dev \ --prompt "A Logo With Bold Large Text: SGL Diffusion" \ --save-output sglang generate --model-path black-forest-labs/FLUX.1-dev \ --prompt "A Logo With Bold Large Text: SGL Diffusion" \ --save-output Reference install guide and cli guide for more installation methods. Demo sglang generate --model-path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \ --prompt "A curious raccoon" \ --save-output sglang generate --model-path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \ --prompt "A curious raccoon" \ --save-output Fallback link: Download the video sglang generate --model-path=Wan-AI/Wan2.1-I2V-14B-480P-Diffusers \ --prompt="Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." \ --image-path="https://github.com/Wan-Video/Wan2.2/blob/990af50de458c19590c245151197326e208d7191/examples/i2v_input.JPG?raw=true" \ --num-gpus 2 --enable-cfg-parallel --save-output sglang generate --model-path=Wan-AI/Wan2.1-I2V-14B-480P-Diffusers \ --prompt="Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." \ --image-path="https://github.com/Wan-Video/Wan2.2/blob/990af50de458c19590c245151197326e208d7191/examples/i2v_input.JPG?raw=true" \ --num-gpus 2 --enable-cfg-parallel --save-output Fallback link: Download the video sglang generate --model-path black-forest-labs/FLUX.1-dev \ --prompt "A Logo With Bold Large Text: SGL Diffusion" \ --save-output sglang generate --model-path black-forest-labs/FLUX.1-dev \ --prompt "A Logo With Bold Large Text: SGL Diffusion" \ --save-output sglang generate --model-path=Qwen/Qwen-Image \ --prompt='A curious raccoon' \ --width=720 --height=720 --save-output sglang generate --model-path=Qwen/Qwen-Image \ --prompt='A curious raccoon' \ --width=720 --height=720 --save-output sglang generate --model-path=Qwen/Qwen-Image-Edit \ --prompt="Convert 2D style to 3D style" --image-path="https://github.com/lm-sys/lm-sys.github.io/releases/download/test/TI2I_Qwen_Image_Edit_Input.jpg" \ --width=1536 --height=1024 --save-output sglang generate --model-path=Qwen/Qwen-Image-Edit \ --prompt="Convert 2D style to 3D style" --image-path="https://github.com/lm-sys/lm-sys.github.io/releases/download/test/TI2I_Qwen_Image_Edit_Input.jpg" \ --width=1536 --height=1024 --save-output Performance Benchmark As shown in the chart at the top of this post, we compared the performance of SGLang Diffusion: Against a popular open-source baseline, Hugging Face Diffusers. SGLang Diffusion delivers state-of-the-art performance, significantly accelerating both image and video generation. Under different parallelism setups. Both CFG-Parallel and USP deliver significant speedups compared to the single-GPU setup. Roadmap and Diffusion Ecosystem Our vision is to build a comprehensive diffusion ecosystem in collaboration with the FastVideo team, providing an end-to-end solution from model training to high-performance inference. The SGLang Diffusion team is centered on continuous innovation in performance and model support: Model support and optimizations Optimize Wan, FastWan, Hunyuan, Qwen-Image series, FLUX Support LongCat-Video Optimize Wan, FastWan, Hunyuan, Qwen-Image series, FLUX Support LongCat-Video Kernel support and fusions Quantization kernels Rotary embedding kernels Flash Attention 4 integration in sgl-kernel for blackwell Quantization kernels Rotary embedding kernels Flash Attention 4 integration in sgl-kernel for blackwell More server features Configurable cloud storage upload of generated files Batching support More parallelism methods Quantization Configurable cloud storage upload of generated files Batching support More parallelism methods Quantization General architecture: Simplify the effort of supporting new models Enhance cache and attention backend supports Simplify the effort of supporting new models Enhance cache and attention backend supports Building this ecosystem is a community effort, and we welcome and encourage all forms of contribution. Join us in shaping the future of open-source diffusion generation. Acknowledgment SGLang Diffusion Team: Yuhao Yang, Xinyuan Tong, Yi Zhang, Ke Bao, Ji Li, Xi Chen, Laixin Xie, Yikai Zhu, Mick FastVideo Team: Peiyuan Zhang, William Lin, Yongqi Chen, Kevin Lin, Wenxuan Tan, Wei Zhou, Runlong Su, Jinzhe Pan, Hangliang Ding, Matthew Noto, You Zhou, Jiali Chen, Hao Zhang Special thanks to NVIDIA and Voltage Park for their compute support. Learn more Roadmap: Diffusion (2025 Q4) Slack channel: #diffusion (join via slack.sglang.io)