在 NVIDIA DGX Spark 上提升 MoE 推理性能的三种方法
原标题：3 Ways to Boost MoE Inference on NVIDIA DGX Spark
AI 摘要
NVIDIA 技术博客介绍在 DGX Spark 上优化 Ling-3.0-flash 混合专家模型推理的三种方法：用 Humming MXFP4 MoE 内核替换 FlashInfer 后端、对输出层 LM head 做在线 FP8 量化、以及基于 NEXTN 的多 token 预测投机解码。前两项分别带来 9.3% 和累计 21.7% 的解码吞吐提升。文章还指出，同时启用 FP8 LM head 与 MTP 时，若草稿模型未共享完整 LM head 模块，框架会静默回退，导致吞吐低于基线。
正文节选
Generating tokens faster starts with identifying what limits inference performance. In the batch-size-1 evaluation of Ling-3.0-flash from Inclusion AI on NVIDIA DGX Spark, profiling showed that decoding was memory-bound: reading model weights drove execution time. Two operations stood out—the mixture-of-experts (MoE) computation and the output-layer projection that produces scores across the model’s vocabulary. Those findings pointed to two ways to improve performance: reduce the weight traffic