Release v4.3.0

What's Changed

• docs(svd_circuits): add demo notebook, slow name-mover parity, and tool docs by @janmenjayap in #1831
• fix: preserve sampling penalties in stateful generation fallback by @emerardd in #1888
• feat(relevance_lens): RelevanceLens.fit + provenance + integration + docs by @janmenjayap in #1834
• feat(attribution_patching): faithfulness (edge ablation) on TransformerBridge by @janmenjayap in #1830
• feat(sparse_probing): support group-aware probe splits by @lorenzozanee in #1824
• Adjusting post norm MLP to be properly differentiated by @jlarson4 in #1890
• Fix T5 bridge forward with transformers >= 5.15 by @YHC66 in #1882
• Fix left-padded BOS removal along the last tensor axis by @AaravMehta-07 in #1891
• fix: include Granite embedding scale in residual decomposition by @emerardd in #1878
• fix: apply MLA complex RoPE to rope slices that are not aligned in storage by @YHC66 in #1898
• Add subspace interchange interventions by @dipo101 in #1892
• fix: keep RMSNorm offsets out of Gemma3 vision LayerNorm by @emerardd in #1897
• Fixing a bug with Jais2 by @jlarson4 in #1900
• fix: preserve prompt attention masks in generation caches by @emerardd in #1899
• Add distributed alignment search for interchange interventions by @dipo101 in #1893
• Adding Guards to our tools/analysis modules by @jlarson4 in #1901
• Add trained controls for distributed alignment search by @dipo101 in #1894
• Analysis Tools Review & Re-contextualization by @jlarson4 in #1907
• fix: preserve final RMS predictions during weight processing by @emerardd in #1903
• Preserve float64 precision in bootstrap and sign-flip statistics by @AaravMehta-07 in #1905
• fix: preserve retained hooks during bridge helper cleanup by @emerardd in #1908
• Count only the cost warning in the coordinate_patch_hooks warn-once test by @YHC66 in #1910
• Fix level-scoped permanent hook cleanup by @koriyoshi2041 in #1914
• fix: keep created biases in the model dtype during weight processing by @tivris in #1912
• fix: preserve autograd through equal-valued scaled hooks by @emerardd in #1915
• Restore num_experts when an HF config aliases it via attribute_map by @YHC66 in #1916
• Add neuron bases for interchange interventions by @dipo101 in #1921
• fix: compute FactoredMatrix and SVDInterpreter SVDs in float32 for 16-bit weights by @tivris in #1919
• Apply the causal mask in MLAAttentionBridge when HF passes no 4D mask by @YHC66 in #1920
• fix(ActivationCache): build layer-0 stacks in the cache's dtype and d… by @avgvi in #1925
• fix(test_prompt): read each answer's first token regardless of the tokenizer's padding side by @avgvi in #1923
• fix: apply dtype and device changes to gated MLPs in compatibility mode by @tivris in #1929
• Fixing flaky Backwards Lens Notebook test by @jlarson4 in #1930
• Fix/slow tier memory budget by @jlarson4 in #1928
• fix(statistics): compute the binomial CDF in log space so it does not underflow by @bercedev in #1931
• feat(bridge): RoBERTa masked-LM adapter by @Canonik in #1895
• fix: validate reused coordinate decompositions by pursuit algorithm by @emerardd in #1932
• Pre-release preparation for v4.3.0 by @jlarson4 in #1934
• docs(demos): add SwiGLU gate-sign demo for double-negative echoes by @emerardd in #1936
• fix: register unmapped leaf parameters so module conversions reach them by @tivris in #1940
• fix(bridge): discover non-standard block lists in _enumerate_blocks by @LightWork666 in #1855
• fix(jacobian_lens): ablate token directions jointly by @flyingdoubleG in #1941
• fix: preserve selected J-space directions across atom scales by @emerardd in #1942
• v4.3.0 by @jlarson4 in #1937

New Contributors

• @lorenzozanee made their first contribution in #1824
• @AaravMehta-07 made their first contribution in #1891
• @dipo101 made their first contribution in #1892
• @tivris made their first contribution in #1912
• @avgvi made their first contribution in #1925
• @bercedev made their first contribution in #1931
• @flyingdoubleG made their first contribution in #1941

Full Changelog: v4.2.0...v4.3.0