← The AI Systems Map
L0 · LabsFramework / tooling

DeepSpeed / Megatron

ZeRO + tensor/pipeline parallelism for trillion-param training

Microsoft DeepSpeed (ZeRO, ZeRO-Offload) and Nvidia Megatron-LM provide the distributed-training primitives — tensor, pipeline, sequence, and expert parallelism — used to train frontier dense and MoE models. Megatron-Core is the reference for large-scale GPU training; the two are frequently combined. They are the invisible plumbing behind most >100B-parameter runs.

Key idea

ZeRO sharding + 4D parallelism

Owners

Microsoft / Nvidia

How it fits the stack

DeepSpeed / Megatron with what it depends on (above) and what it feeds (below). The figure renders as a crawlable diagram and upgrades to an interactive 3D graph as it scrolls into view.

depends onusesusesDeepSpeed / MegatronLabsNvidia Data-Center GPU(Blackwell/Rubin)chokepointMeta AI (FAIR)OpenAI
DeepSpeed / MegatronDepends on ↑Feeds ↓

DeepSpeed / Megatron in the AI stack. DeepSpeed / Megatron with its immediate upstream dependencies (top) and downstream dependents (bottom) in the AI value chain. Hover a node in 3D, or read the full relationships below.

Graph data (text) — 4 entities, 3 relationships
  • DeepSpeed / Megatrondepends onNvidia Data-Center GPU (Blackwell/Rubin)
  • Meta AI (FAIR)usesDeepSpeed / Megatron
  • OpenAIusesDeepSpeed / Megatron