search

ai-safety-foundation / sparse_autoencoder

Sparse Autoencoder for Mechanistic Interpretability
https://ai-safety-foundation.github.io/sparse_autoencoder/
MIT License
191 stars 39 forks source link

Make the demo notebook more visible in the README #142

Closed ArthurConmy closed 11 months ago

ArthurConmy commented 11 months ago

Sadly, I'm seeing this run out of RAM even on V100s for all setups I've tried so far

alan-cooney commented 11 months ago

Sadly, I'm seeing this run out of RAM even on V100s for all setups I've tried so far

Are you running out of memory on train or validation?

The demos use the Anthropic Towards Monosemanticity defaults, which include a very large batch size (8k). Coupled with the fact that we're training there on a larger residual stream (d_model=512), it is pretty resource hungry so I'm not surprised a single V100 (esp if 16GB) can't take it. We can add a note for a suggestion to reduce the batch size - or just put a default size that is a bit smaller (but we'll need to do a new sweep to find sensible defaults if so for l1 coefficient).