single source · post-training
Data Scheduler
It addresses GPU idleness in distributed MoE training by assigning new prompts to GPUs with load balancing.1
Trade-off It significantly accelerates RL training without sacrificing model quality.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
implement a Data Scheduler to seamlessly schedule fine-grained sequences instead of micro-batches
— MiMo-V2-Flash Technical Reportwe significantly accelerate RL training without sacrificing model quality
— MiMo-V2-Flash Technical Report
Filed under post-training › rollout & RL infrastructure