APRIL Strategy

**A**ctive **P**artial **R**ollout for **I**efficient generation with **L**ong-tail handling.

Strategy

1. Oversample: Generate more rollouts than needed
2. Abort: Cancel long-running generations
3. Reuse: Save and reuse partial trajectories

Configuration

rollout:
  april:
    oversample_ratio: 1.5
    batch_timeout: 30.0
    partial_reuse_threshold: 0.5

Benefits

• Reduces waiting for slow generations
• Improves GPU utilization
• Maintains training throughput