Open Models Almanac
Methodspost-trainingrollout & RL infrastructure

single mention · post-training

Compute-node sharding into scale units

Compute-node sharding partitions compute nodes into multiple shards, which are referred to as scale units.1 This sharding accommodates a large number of machines and reduces blast radius by isolating workloads from different experiments.2

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. we partition compute nodes into multiple shards (so-called scale units — DeepSeek V41 Tech Report
  2. To accommodate a large number of machines, we partition compute nodes into multiple shards (so-called scale units). Such sharding also allows us to reduce blast radius by isolating workloads from different experiments — DeepSeek V41 Tech Report

Filed under post-training › rollout & RL infrastructure