Summary performance (regenerated by tools/benchmark_summary.py)¶

One-call tl.summary(model, x) cost decomposed against a plain uninstrumented forward (min-of-N, CPU, single process). The RATIO column is the published figure the summary memo's engine-floor trigger reads: a second capture engine is considered only if the optimized single-engine summary stays above ~5x a plain forward.

model	forward	capture	postprocess (in capture)	render	cleanup	summary total	ratio	postprocess share
toy	21.6 ms	87.1 ms	14.8 ms	2.1 ms	10.1 ms	99.2 ms	4.6x	17%
resnet18	1320.1 ms	4271.8 ms	327.6 ms	35.7 ms	367.2 ms	4674.7 ms	3.5x	8%
gpt2	1307.4 ms	6450.2 ms	1567.2 ms	252.1 ms	989.0 ms	7691.3 ms	5.9x	24%

Notes:

postprocess is measured INSIDE capture (the wrapped forward plus the 26-step pipeline); its share is the summary-scoped optimization target the summary memo measured at ~64% of capture cost on the pre-speedening baseline (f74e4a9e). The post-speedening share is what THIS page reports.
render is the rebuilt view ladder + hairline renderer.
Numbers vary by host AND by concurrent load (both the plain forward and the capture inflate together under load, compressing the ratio less than either absolute); regenerate on a quiet host with python tools/benchmark_summary.py and commit the page.
disposition: a second hook-based capture engine goes to a maintainer decision only if the optimized single-engine ratio stays above ~5x.