Architecture

Rollout-Level Replay Buffers

A distributed storage pattern for RL post-training of LLMs that persists individual rollouts and reasoning trajectories to secondary storage — prioritized by advantage, bounded by staleness — so pipelines recycle high-value experiences instead of regenerating them.

6 connections 2 resources 1 post

Summary

What it is

A distributed storage pattern for RL post-training of LLMs that persists individual rollouts and reasoning trajectories to secondary storage — prioritized by advantage, bounded by staleness — so pipelines recycle high-value experiences instead of regenerating them.

Where it fits

The training-side counterpart to inference memory. Because retaining trajectories in VRAM causes OOM, the pattern streams buffers to NVMe and object storage, making S3 the persistent backend of the RL loop.

Misconceptions / Traps
  • This is not classical experience replay — LLM policies drift fast, so it stores/samples per-rollout and strictly bounds staleness, or training destabilizes.
  • The compute saving (regeneration can eat 80%+ of the GPU budget) is the whole motivation.
Key Connections
  • Rollout-Level Replay Buffers depends_on Object Storage — durable rollout persistence
  • verl Hybrid Replay Buffer extends it (production implementation)
  • Rollout Routing Replay (R3) extends it for MoE routing stability

Definition

What it is

A distributed storage pattern used in reinforcement-learning post-training of LLMs that persists individual model rollouts and reasoning trajectories to secondary storage, letting training pipelines recycle high-value experiences and cut the cost of continuous on-policy data generation.

Why it exists

Dominant RL post-training methods (e.g. GRPO) are sample-inefficient — generating fresh trajectories can consume 80%+ of the post-training GPU budget. Classical experience replay struggles because policies drift and stored rollouts go stale. This pattern stores and samples per-rollout (not per-group), prioritizes by advantage magnitude, and bounds staleness via age-eviction — streaming buffers to NVMe/object storage instead of scarce VRAM.

Primary use cases

Persisting and resampling RL rollouts during LLM post-training, trading minor staleness for large compute savings, surviving training restarts.

Recent developments

Latest signals

Connections 6

Outbound 4
Inbound 2

Resources 2

Featured in