## Summary
Follow-up to #8282. The docs note for `cpu_checkpointing` landed after
that PR was merged, so this cherry-picks it onto `master`.
Adds a short note under the `cpu_checkpointing` config entry covering:
- The Qwen3-8B single-H200 result: async side-stream copy matches a
blocking offload's peak reduction (up to ~14% at 32K) while staying
within ~2% of no-offload step time.
- Recommending `PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True` for
very long sequences to avoid allocator fragmentation from the
offload/restore cycle.
## Test plan
- [x] `pre-commit run --files docs/_pages/config-json.md`
- [ ] Docs preview of the Activation Checkpointing `cpu_checkpointing`
section
Made with [Cursor](https://cursor.com)
Signed-off-by: tunji-ruwase_snow <tunji.ruwase@snowflake.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Olatunji Ruwase
96ffef2cca