v4.2.0

What's Changed

• fix sparse probe control seeding by @MdSadiqMd in #1821
• Add tie-aware roc_auc and average_precision to sparse probe metrics by @akshpatel0910 in #1820
• fix(bridge): honor stop_at_eos=False in hf_generate by @emerardd in #1826
• fix(projection_kernel): derive the default rank tolerance from the compute dtype by @janmenjayap in #1829
• fix sparse probe partial results by @MdSadiqMd in #1828
• Forward the requested revision to every remote-code force-import by @YHC66 in #1819
• fix(sparse_probing): add std_floor to standardization by @VictorAraujopy in #1827
• Fix get_neuron_results projection shapes for integer neuron slices by @yuanwuyuan9 in #1837
• Add Muse Glimmer architecture adapter by @almogtavor in #1822
• fix(typing): correct annotations that current beartype rejects by @YHC66 in #1838
• Fix torch.stack and torch.cat recursion for CompositionScores by @yuanwuyuan9 in #1840
• Wire CLIP vision-layer hook aliases for LLaVA adapters by @koriyoshi2041 in #1823
• fix(bridge): preserve seq2seq state in streaming generation by @emerardd in #1825
• test(nemotron_h): use a real NemotronHConfig in stateful-cache tests by @YHC66 in #1845
• Fix stack_neuron_results for an integer neuron_slice by @Zhuoxi2000 in #1844
• Fix per-layer head result cache validation by @yuanwuyuan9 in #1843
• Use the ln2 scale for neurons in fused LN+projection resid decomposition by @Zhuoxi2000 in #1847
• Loosen float32 tolerance in edge-effects mutation test by @YHC66 in #1850
• Fix FactoredMatrix ellipsis indexing by @yuanwuyuan9 in #1849
• fix: align per-example logit attribution across positions by @nehal-a2z in #1835
• Built in Logit Lens by @jlarson4 in #1861
• perf: reduce temporary memory when computing attention head results by @kiteretsu903 in #1854
• fix: preserve attention masks in Inspect HF forwards by @emerardd in #1852
• fix(bridge): slice pos before dropping batch dim in get_caching_hooks by @JoeyTan21 in #1859
• Fixing CI test flakiness by @jlarson4 in #1867
• added unreduced patching by @jlarson4 in #1866
• fix(bridge): set rmsnorm_uses_offset on the Qwen3.5, Qwen3Next and Nemotron adapters by @willgosnold in #1869
• fix(bridge): key get_caching_hooks entries under the matched alias names by @Mudassiruddin7 in #1862
• fix(bridge): keep the batch dimension in the caching hooks when it is larger than 1 by @Mudassiruddin7 in #1865
• fix: honor vLLM logits processor transforms in reconstruction by @emerardd in #1853
• fix(bridge): let a hook return None under run_with_hooks(remove_batch_dim=True) by @Mudassiruddin7 in #1864
• fix: keep synthetic output biases out of the optimizer by @kiteretsu903 in #1871
• fix(bridge): rotate Cohere and Cohere2 with their own RoPE convention by @Mudassiruddin7 in #1873
• fix(bridge): update the KV cache in MPT's attention reconstruction by @Mudassiruddin7 in #1874
• fix(bridge): apply HunYuan's QK-norm after RoPE by @Mudassiruddin7 in #1872
• Prep for release v4.2.0 – Contribution follow ups by @jlarson4 in #1875
• fixing slow tests by @jlarson4 in #1877
• Release v4.2.0 by @jlarson4 in #1876

New Contributors

• @akshpatel0910 made their first contribution in #1820
• @VictorAraujopy made their first contribution in #1827
• @yuanwuyuan9 made their first contribution in #1837
• @Zhuoxi2000 made their first contribution in #1844
• @nehal-a2z made their first contribution in #1835
• @kiteretsu903 made their first contribution in #1854
• @JoeyTan21 made their first contribution in #1859
• @willgosnold made their first contribution in #1869
• @Mudassiruddin7 made their first contribution in #1862

Full Changelog: v4.1.0...v4.2.0