Ant Group’s DeepXPU team and the SGLang development team have announced a strategic collaboration to integrate Diffusion Large Language Models (dLLMs) into the SGLang ecosystem. This partnership enables day-zero support for LLaDA 2.0, a cutting-edge model co-developed by researchers at Renmin University of China and Ant Group. Unlike standard auto-regressive models such as the GPT series, which predict the next token sequentially, LLaDA 2.0 employs a diffusion process to refine text iteratively, mimicking the techniques used in modern image generation models.

Scaling these non-sequential models presents significant technical hurdles that conventional inference engines often struggle to address effectively. SGLang overcomes these challenges by implementing its proprietary Chunked-Prefill mechanism, which accommodates the unique requirements of diffusion models without requiring fundamental changes to the core architecture. This implementation allows dLLMs to benefit from existing optimization features while providing developers the flexibility to customize diffusion decoding algorithms. Such versatility is essential for researchers looking to explore the frontiers of non-sequential text generation and complex data structures.

Initial performance benchmarks demonstrate the significant efficiency gains provided by this integration within the SGLang framework. The LLaDA 2.0-flash-CAP (100B) model achieved a throughput of 935 tokens per second, a substantial improvement over the 263 tokens per second recorded by comparable large-scale models. SGLang’s proven stability and its broad compatibility with reinforcement learning ecosystems make it an ideal platform for hosting these massive next-generation models. This update is expected to accelerate the practical adoption of dLLMs and streamline the development cycle for advanced AI research teams worldwide.

Ant Group’s DeepXPU team and the SGLang development team have announced a strategic collaboration to integrate Diffusion Large Language Models (dLLMs) into the SGLang ecosystem. This partnership enables day-zero support for LLaDA 2.0, a cutting-edge model co-developed by researchers at Renmin University of China and Ant Group. Unlike standard auto-regressive models such as the GPT series, which predict the next token sequentially, LLaDA 2.0 employs a diffusion process to refine text iteratively, mimicking the techniques used in modern image generation models.

Scaling these non-sequential models presents significant technical hurdles that conventional inference engines often struggle to address effectively. SGLang overcomes these challenges by implementing its proprietary Chunked-Prefill mechanism, which accommodates the unique requirements of diffusion models without requiring fundamental changes to the core architecture. This implementation allows dLLMs to benefit from existing optimization features while providing developers the flexibility to customize diffusion decoding algorithms. Such versatility is essential for researchers looking to explore the frontiers of non-sequential text generation and complex data structures.

Initial performance benchmarks demonstrate the significant efficiency gains provided by this integration within the SGLang framework. The LLaDA 2.0-flash-CAP (100B) model achieved a throughput of 935 tokens per second, a substantial improvement over the 263 tokens per second recorded by comparable large-scale models. SGLang’s proven stability and its broad compatibility with reinforcement learning ecosystems make it an ideal platform for hosting these massive next-generation models. This update is expected to accelerate the practical adoption of dLLMs and streamline the development cycle for advanced AI research teams worldwide.