Choose memory-saving checkpoints with measured tradeoffs.
What it does
Select and validate activation-recompute settings that reduce training memory while measuring the extra compute cost.
When to use it
Use for activation-memory OOMs or regressions involving selective or full recompute and activation checkpointing.
How to use it
Give it a training workload and recompute question; it returns candidate settings, their tradeoffs, and compatibility guidance.
Latest review
What you need · 0
No setup requirements listed.
Access · 0
No additional actions listed.