Open Models Almanac
Methodspost-trainingrollout & RL infrastructure

specific method Β· filed under post-training

Adaptive Rollout Concurrency

Adaptive Rollout Concurrency assigns each data source a budget of concurrent rollouts based on that source's generation demand.1,2 It scales concurrency with the source's expected demand (target samples over acceptance rate) and rollout duration, so slow or heavily filtered sources get more concurrent rollouts.3

How this definition was written

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

  1. Adaptive Rollout Concurrency sets per-source budgets β€” MiMo V2 6 technical report
  2. Slowersourcesneedmoreconcurrentrolloutstosustainthesame trainingcontribution β€” MiMo V2 6 technical report
  3. Theexpectedgenerationdemandis π‘šπ‘– = 𝐡𝑖/π‘Ÿπ‘– β€” MiMo V2 6 technical report
sources
2
models
3
labs adopt it
2
strongest
used

Documented in

Evidence

2 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
used 2

Mistral-Large-4

Mistral AIadopted1 source

The episode generation concurrency adapts following backpressure signals from generators, services, and the trainer, which limits staleness.

usedunclearin Mistral Large 4Mistral AI
↑