Reinforcement Learning Infrastructure

Overview

The systems engineering challenge of running RL training at scale. Distinct from pretraining infrastructure because RL has two jobs: training (forward/backward/update) and rollout generation (inference on current policy).

Key Insight: Delta Compression

Between consecutive RL checkpoints, 98%+ of bf16 weights are bit-identical. This means:

  • Ship ~2% of the model per update instead of 100%
  • Cross-region synchronization becomes practical over ordinary networks
  • No need for trainer and rollout fleet on the same RDMA fabric

This fundamentally challenges the “mega-cluster” narrative. Source

Async RL

Trade-off: accept a few minutes of policy staleness for much better compute utilization. The delta compression makes this practical by keeping policy movement small and fast.

Implications

If frontier RL doesn’t require mega-clusters, the competitive landscape changes. Teams with fragmented capacity across regions can compete. The bottleneck shifts to algorithms and product, not hardware concentration.

Connections

Vault Notes