v2.10.0

Huge update! This is likely going to be the last big 2.x update. This update greatly improves model implementation accuracy, and adds some of the newer Qwen models.

What's Changed

• Remove einsum in forward pass in AbstractAttention by @degenfabian in #783
• Colab compatibility bug fixes by @degenfabian in #794
• Remove einsum usage from create_alibi_bias function by @degenfabian in #781
• Actions token access by @bryce13950 in #797
• Remove einsum in apply_causal_mask in abstract_attention.py by @degenfabian in #782
• clarified arguments a bit for hook_points by @bryce13950 in #799
• Remove einsum in logit_attrs in ActivationCache by @degenfabian in #788
• Remove einsum in compute_head_results in ActivationCache by @degenfabian in #789
• Remove einsum usage in refactor_factored_attn_matrices in HookedTransformer by @degenfabian in #791
• Remove einsum usage in _get_w_in_matrix in SVDInterpreter by @degenfabian in #792
• Remove einsum usage in forward function of BertMLMHead by @degenfabian in #793
• Set default_prepend_bos to False in Bloom model configuration by @degenfabian in #806
• Remove einsum in complex_attn_linear by @degenfabian in #790
• Add a demo of collecting activations from a single location in the model. by @adamkarvonen in #807
• Add support for Qwen_with_Questions by @degenfabian in #811
• Added support for Qwen2.5 by @israel-adewuyi in #809
• Updated devcontainers to use python3.11 by @jonasrohw in #812

New Contributors

• @israel-adewuyi made their first contribution in #809
• @jonasrohw made their first contribution in #812

Full Changelog: v2.9.1...v2.10.0

v2.9.1

Minor dependency change to address a change in an outside dependency

What's Changed

• added typeguard dependency by @bryce13950 in #786

Full Changelog: v2.9.0...v2.9.1

v2.9.0

Lot's of accuracy improvements! A number of models are behaving closer to how they behave in Transformers, and a new internal configuration has been added to allow for more ease of use!

What's Changed

• fix the bug that attention_mask and past_kv_cache cannot work together by @yzhhr in #772
• Set prepend_bos to false by default for Bloom model family by @degenfabian in #775
• Fix that if use_past_kv_cache is set to True models from the Bloom family produce weird outputs. by @degenfabian in #777

New Contributors

• @yzhhr made their first contribution in #772
• @degenfabian made their first contribution in #775

Full Changelog: v2.8.1...v2.9.0

v2.8.1

New notebook for comparing models, and bug fix with dealing with newer LLaMA models!

What's Changed

• Logit comparator tool by @curt-tigges in #765
• Add support for NTK-by-Part Rotary Embedding & set correct rotary base for Llama-3.1 series by @Hzfinfdu in #764

New Contributors

• @Hzfinfdu made their first contribution in #764

Full Changelog: v2.8.0...v2.8.1

v2.8.0

What's Changed

• add transformer diagram by @akozlo in #749
• Demo colab compatibility by @bryce13950 in #752
• Add support for Mistral-Nemo-Base-2407 model by @ryanhoangt in #751
• Fix the bug that tokenize_and_concatenate function not working for small dataset by @xy-z-code in #725
• added new block for recent diagram, and colab compatibility notebook by @bryce13950 in #758
• Add warning and halt execution for incorrect T5 model usage by @vatsalrathod16 in #757
• New issue template for reporting model compatibility by @bryce13950 in #759
• Add configurations for Llama 3.1 models(Llama-3.1-8B and Llama-3.1-70B) by @vatsalrathod16 in #761

New Contributors

• @akozlo made their first contribution in #749
• @ryanhoangt made their first contribution in #751
• @xy-z-code made their first contribution in #725
• @vatsalrathod16 made their first contribution in #757

Full Changelog: v2.7.1...v2.8.0

v2.7.1

What's Changed

• Updated broken Slack link by @neelnanda-io in #742
• from_pretrained has correct return type (i.e. HookedSAETransformer.from_pretrained returns HookedSAETransformer) by @callummcdougall in #743
• Avoid warning in utils.download_file_from_hf by @albertsgarde in #739

New Contributors

• @albertsgarde made their first contribution in #739

Full Changelog: v2.7.0...v2.7.1

v2.7.0

Model 3.2 support! There is also a new compatibility added to the function test_promt to allow for multiple prompts, as well as a minor typo.

What's Changed

• Typo hooked encoder by @bryce13950 in #732
• utils.test_prompt compares multiple prompts by @callummcdougall in #733
• Model llama 3.2 by @bryce13950 in #734

Full Changelog: v2.6.0...v2.7.0

v2.6.0

Another nice little feature update! You now have the ability to ungroup the grouped query attention head component through a new config parameter ungroup_grouped_query_attention!

What's Changed

• Ungrouping GQA by @hannamw & @FlyingPumba in #713

Full Changelog: v2.5.0...v2.6.0

v2.5.0

Nice little release! This release adds a new parameter named first_n_layers that will allow you to specify how many layers of a model you want to load.

What's Changed

• Fix typo in bug issue template by @JasonGross in #715
• HookedTransformerConfig docs string: weight_init_mode => init_mode by @JasonGross in #716
• Allow loading only first n layers. by @joelburget in #717

Full Changelog: v2.4.1...v2.5.0

v2.4.1

Little update to the code usage, but huge update for memory consumption! TransformerLens now needs almost half the memory it needed previously to boot thanks to a change with how the TransformerLens models are loaded.

What's Changed

• removed einsum causing error when use_atten_result is enabled by @oliveradk in #660
• revised loading to recycle state dict by @bryce13950 in #706

New Contributors

• @oliveradk made their first contribution in #660

Full Changelog: v2.4.0...v2.4.1