David Chanin

chore: tweak readme wording about reference papers

2

Following #57, this PR updates the wording on the README to make it more clear that those 2 papers specifically are referenced because their code was an inspiration for this...

chore: delete pdm.lock file

2

The pdm.lock file is only useful for app development to ensure that all developers and the deployed app share identical versions of all dependencies. For a library like steering-vectors, the...

Gemma 2 returns NaN when using default attn (sdpa) with padding

9

### System Info Python 3.10 Transformers 4.43.3 Linux (Colab notebook) ### Who can help? @ArthurZucker ### Information - [X] The official example scripts - [ ] My own modified scripts...

The following code works, but pyright/pylance gives a typing error: ```python from nnsight import LanguageModel model = LanguageModel('google/gemma-2-2b') with model.trace('hi'): acts = model.model.layers[0].output.save() ``` The errors are the following: ```...

I really like your hierarchical feature setup via the `Tree` class, but found it ran a bit slow due to using Python loops and lists. This PR vectorizes more of...

**Describe the bug** If you pass a 1d tensor ot tokens to `TransformerBridge.run_with_cache`, at least for Gemma models, this causes a crash with a vague message about mismatched attention shapes....