新闻
Accelerating Dropless MoE Training in JAX with NVIDIA Transformer Engine
NVIDIA Developer · 发布于 · 阅读约3分钟
30秒读懂
- 发生了什么
- NVIDIA Transformer Engine with JAX achieves 10.4x throughput improvement for Mixture of Experts training, reaching 1,068 TFLOPS/GPU on DeepSeek-V3.
- 为何重要
- Matters for engineers training large MoE models at scale who need to maximize GPU utilization and reduce inter-GPU communication overhead.
- 注意
- Dropless MoE requires specialized kernels for variable token counts; standard capacity-based MoE may remain simpler for some use cases despite efficiency tradeoffs.
- mixture of experts
- qwen
- deepseek
这条新闻背后的模式
每个模式都讲清楚技术如何运作、何时值得投入，以及在哪里会失效。
The Agent Architect
每周一个模式、一个权衡、一个生产事故案例。为构建智能体系统的人准备的每周简报。
每周一封邮件,一键退订。您的地址仅用于发送简报。