LeanProfiler

5.7. Start with the symptom🔗

Symptom

First tool

Why

one Lean declaration is slow to build

lean --profile or trace.profiler

the cost occurs during elaboration or compilation

a running Lean command has a slow phase

LeanProfiler

the application owns the meaningful phase names

a PyTorch operator dominates CPU time

PyTorch Profiler

operators are discovered and grouped automatically

a CUDA model waits unexpectedly

LeanProfiler boundary plus PyTorch/CUPTI or Nsight detail

one view locates the application call; the other resolves kernels and runtime activity

a candidate build regressed

LeanProfiler comparison over repeated controlled captures

the summary schema and threshold policy are explicit