Abdur Raheem Ali

icon indicating a website link abdur-raheem.com

icon company Independent (formerly Microsoft) icon location London, UK icon location counterfactual oracle predicting subspaces of the residual stream

Results 4 comments of Abdur Raheem Ali

Add a helper function to display vectors of logits nicely

https://www.brendangregg.com/blog/2017-02-06/flamegraphs-vs-treemaps-vs-sunburst.html for a non-interactive visualization, flame graphs do pretty well (I'm @sheikheddy from an alt-account)

[Bug Report] convert_llama_weights fails if I already quantized the weights to 4 bits

The error originates from an incorrect use of the einops.rearrange function in the convert_llama_weights function from the transformer_lens package. Specifically, the error message "Shape mismatch, can't divide axis of length...

[Bug Report] convert_llama_weights fails if I already quantized the weights to 4 bits

this may be addressed by https://github.com/neelnanda-io/TransformerLens/pull/486. I'll check it out.

Tracking mood/emotions

> Log manually on a scale every day, possibly several times a day. I use Daylio as a source of ground truth for this (premium version allows export). Apparently this...