Skip to main content

Bottleneck Fix

Sometimes the compiler makes things slower — and the cure is to stop breaking its graph.

Key Insight​

torch.compile is fastest on one unbroken graph. When it meets code it cannot trace — a print, a data-dependent branch, an unsupported op — it inserts a graph break, splitting the model and falling back to slow eager mode in between, sometimes making the whole run slower than not compiling at all.

Why This Matters​

Finding and removing graph breaks (with TORCH_LOGS="graph_breaks") restores the speedup the compiler promised, and teaches you exactly which Python the compiler can and cannot handle — turning a bottleneck back into a win.