ESC

Type to search articles...

No articles found.

↑ ↓ Navigate ↵ Open
Esc Close
Blog Tags GitHub
All tags

parallelism-ep

2 posts tagged with "parallelism-ep"

May 31, 2026

MegaScale-MoE: 大规模通信高效的生产级混合专家模型训练系统

本文深入分析MegaScale-MoE,一个专为大规模MoE模型高效训练设计的生产系统,通过通信高效并行策略、通信-计算重叠和通信压缩,在1,440个NVIDIA Hopper GPU上训练352B MoE模型实现1.41M tokens/s吞吐量,相比Megatron-LM提升1.88倍。

communication mixture-of-experts parallelism-pp parallelism-ep parallelism-tp
April 14, 2026

Event Tensor: A Unified Abstraction for Compiling Dynamic Megakernel

现代 GPU 工作负载,特别是大语言模型(LLM)推理,受到内核启动开销和粗粒度同步的限制,阻碍了内核间并行性。

system-optimization kernel-optimization parallelism-pp parallelism-ep parallelism-tp scheduling

Navigation

  • Work

Resources

  • Lexington Themes.

Socials

  • @Mike_Andreuzza
© 2025 MicroStudio. All rights reserved.

MicroStudio is not affiliated with Stripe, Breeew, Astro, or Tailwind Labs, nor is it endorsed or sponsored by them.