$ cat interpretability/week/2026-11-02.log
the week in review · Nov 2 – Nov 8, 2026
nnsight 0.8.2 fixes silent wrong values under vLLM
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
SAELens ships 6.55.0 on transformer-lens 4.x but keeps its 3.x pin, and its eval metrics get two dtype and sparse-tensor fixes.
v0.8.2 ndif-team/nnsight
The 0.8.2 bug-fix release frees GPU memory that stopped traces, failed backward passes and edits kept allocated, and corrects cases where vLLM returned wrong values without raising under tensor parallelism and with concurrent requests. It also adds tensor parallelism support on transformers 5.21.
v6.55.0 jbloomAus/SAELens
The release adds a transformer-lens 4.x path, tests it in CI, and adds model_class_name="TransformerBridge" for loading models, but keeps the transformer-lens<4.0.0 pin. Running on 4.x is opt-in until the pin lifts.
fix: promote low-precision activations before computing eval metrics jbloomAus/SAELens
get_sparsity_and_variance_metrics still does its input-space arithmetic in the activation dtype, so mse, the l2_ratio and l2_norm metrics, relative_reconstruction_bias and explained_variance_legacy computed in float16 could report numbers that do not match the real activations. Low-precision activations are now promoted before the arithmetic.
$ ls interpretability/week/ # the briefings behind this review
Keep up with Interpretability in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.
Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.