$ the-wire · showcase
nnsight 0.8.0 cuts import time from 7.6s to 1.6s
By RepoJournal · Filed · About Interpretability · Composed from the cited sources · methodology
nnsight v0.8.0 ships lazy-loaded model classes that take `import nnsight` from 7.6 seconds to 1.6, along with a left-padded position_ids fix and revised source-op iteration counting.
v0.8.0 ndif-team/nnsight
The release notes describe the headline change as "perf: lazy-load model classes to cut import nnsight from 7.6s to 1.6s", which every script, notebook, and CLI wrapper that imports nnsight pays for. It also derives position_ids from the attention mask for left-padded batches, so batched traced runs with left padding should no longer see position errors, and counts source-op iterations per-fire...
Merge pull request #772 from ndif-team/dev ndif-team/nnsight
The dev branch merged into main, carrying the work that became v0.8.0 upstream. If you pin to a branch rather than a tag, this is the point your checkout moves.
Action items
- → Upgrade nnsight to 0.8.0 to pick up the faster import and the left-padded position_ids fix ndif-team/nnsight [plan]