---
tags: systems
---
# torch-compile

torch.compile with max-autotune, as a stronger floor.

## What happened

- 2026-09-13: 1.12 speed-up (+0.12 on the baseline)

## Next

Hand-written kernels should beat this.