v4.4.0

What's Changed

• fix(moe): avoid incorrect Granite grouped expert execution on MPS by @janmenjayap in #1887
• Fork the CUDA RNG explicitly in deterministic_rng by @YHC66 in #1943
• fix: honor in-place edits on native normalization hooks by @emerardd in #1947
• Updating internal-testing Qwen 3.5 model by @jlarson4 in #1950
• Whisper adapter by @dipo101 in #1944
• fix: respect parallel MLP inputs in residual decomposition by @emerardd in #1946
• fix: take cfg.device from a pre-loaded hf_model instead of get_device() by @tivris in #1949
• docs(tutorials): index the 4.x demo notebooks by @Sonica-B in #1948
• fix: preserve square unembedding layout during final norm folding by @emerardd in #1951
• fix: map head and neuron results through post-sublayer norms in residual decompositions by @flyingdoubleG in #1953
• fix: honor padding_side for token input in forward() by @ska2704 in #1952
• fix: resolve square MLP weight layout when folding and centering by @YHC66 in #1955
• feat: support KV-group destinations in direct path patching by @emerardd in #1956
• fix: keep float64 weights in float64 during weight processing by @Sonica-B in #1960
• fix: handle zero projections and low precision in norm-matched placebo by @emerardd in #1959
• fix: handle in-memory tokenizers in get_tokenizer_with_bos by @janmenjayap in #1962
• fix: stop after the last block for stop_at_layer=n_layers by @ska2704 in #1964
• fix: return raw model logits from hooked generation by @emerardd in #1967
• feat(bridge): generate() for Whisper by @dipo101 in #1970
• fix(evals): run the eval helpers on the model's own device by @abhishekcpp007 in #1965
• fix(ActivationCache): allow an identity slice on a batchless cache by @jontb in #1975
• Reject negative temperature and non-positive repetition_penalty in sample_logits by @praann07 in #1977
• docs(demos): add Natural Language Autoencoders round-trip demo by @janmenjayap in #1963
• fix: keep the causal mask for broadcast 4D padding masks by @YHC66 in #1971
• feat(bridge): weight processing for Whisper by @dipo101 in #1972
• fix: fire hook_out on blocks that return a list by @ska2704 in #1973
• fix: honor prepend_bos in the batched generation attention mask by @vayllon301 in #1984
• feat(bridge): accept level and prepend in add_hook by @vayllon301 in #1985
• docs: add a weight-edit audit demo with TransformerBridge by @emerardd in #1983
• Fix tensor detection patterns for prompt lists by @koriyoshi2041 in #1982
• fix(bridge): reject misplaced hook and cache execution arguments by @emerardd in #1990
• Whisper model verification by @dipo101 in #1989
• fix(ActivationCache): reject a negative remainder layer in stack_head_results by @abhishekcpp007 in #1976
• Pre release v4.4.0 by @jlarson4 in #1995
• Make residual-decomposition APIs work on SSM / hybrid bridges by @SBrouck in #1987
• Reduce the total number of garbage collection calls in CI by @jlarson4 in #1999
• Release v4.4.0 by @jlarson4 in #1997

New Contributors

• @Sonica-B made their first contribution in #1948
• @jontb made their first contribution in #1975
• @praann07 made their first contribution in #1977
• @vayllon301 made their first contribution in #1984
• @SBrouck made their first contribution in #1987

Full Changelog: v4.3.0...v4.4.0