$ the-wire · showcase
SAELens 6.54.5 fixes silent feature-density skips
By RepoJournal · Filed · About Interpretability · Composed from the cited sources · methodology
2 people shipped this
SAELens shipped 6.54.5 carrying a fix for feature-density evaluation that silently returned nothing, plus a training-eval memory saving operators feel immediately.
6.54.5 jbloomAus/SAELens
by semantic-release
Release 6.54.5 packages the standalone feature-density fix, so installing it is the fastest way to get working density evals rather than tracking the commit.
fix: honor standalone feature-density evaluation jbloomAus/SAELens
`run_evals` used to skip feature-density evaluation entirely when `compute_featurewise_density_statistics=True` was the only activation-based metric selected, because the dispatch condition checked only the L2, sparsity and variance flags. Runs in that configuration returned no `feature_density` or `consistent_activation_heuristic` values and also bypassed validation of zero or negative batch c...
perf(evals): skip unused feature-density masking jbloomAus/SAELens
Routine training evaluation was building a full masked copy of the feature activations even when feature-density statistics were disabled. In a CPU reproduction with activations shaped `(4, 128, 16384)`, moving that work into the branch that uses it removes a 32 MB allocation per evaluation batch, with all returned metrics matching the original code.
fix: honor standalone feature-density evaluation (#762) jbloomAus/SAELens
The merged commit carries the same standalone feature-density fix into the mainline history.
Action items
