Writing — Eleven graph breaks: chasing a 24× inference regression
< All posts
FEB 2026 · GPU OPTIMIZING · COMPILATION · 1 MIN

Eleven graph breaks: chasing a 24× inference regression

Placeholder body copy. The model compiled, the logs said so, and inference was twenty-four times slower than the eager baseline. Every graph break is a place where the compiler gives up and hands control back to Python, and there were eleven of them.

Placeholder continuation. Walk through how each break was found, which ones were in the model and which ones were in glue code, and what the trace looked like once the graph was whole again.

About aaaallleen.github.io
aaaallleen.github.io
Version 1.0 · built 2026-09-14
ENGINE
Astro 7.3.2, fully static
HOSTING
GitHub Pages
DESIGN
Retro desktop, mocked up in Claude Design
TYPE
Silkscreen & Verdana
OWNER
Allen Lu · San Diego, CA