文本扩散模型大幅提升生成效率,适合需要低延迟批量推理的 AI 应用开发者,建议立即在 SGLang 中体验。
SGLang 宣布 Day-0 支持 Google 的 DiffusionGemma 模型,这是 Gemma 4 的文本扩散变体(26B A4B MoE)。与传统逐 token 解码不同,DiffusionGemma 通过并行去噪 token 块实现极低批处理生成速度。该模型支持离散文本扩散、多模态输入(文本、图像、视频)输出文本、稀疏 MoE 架构(8/128 专家)以及可配置思考模式。开发者现在即可通过 SGLang 运行该模型。
🎉 SGLang has Day-0 support for DiffusionGemma, a t…
🎉 SGLang has Day-0 support for DiffusionGemma, a text-diffusion variant of @googlegemma 's Gemma 4 (26B A4B MoE), built for blazing low-batch generation speed!
Instead of token-by-token decoding, it denoises blocks of tokens in parallel for much faster generation. 1️⃣ Discrete text diffusion: block-parallel multi-canvas sampling 2️⃣ Multimodal: text, image & video in, text out 3️⃣ Sparse MoE (8 of 128 experts): strong reasoning, low memory 4️⃣ Configurable thinking mode
Run it now with SGLang!