search

FMInference / DejaVu

268 stars 32 forks source link

Training MLP layer Issue #6

Open iidsample opened 11 months ago

iidsample commented 11 months ago

Hi, We are trying to train MLP layers for reproducing results for the OPT model.

We followed the instructions provided in /sparse_predictor/Readme.md.

When running the MLP trainer we see that it is looking for hard coded 64 layers and is not able to find those layers. We see that the files which are being looked at are - mlp_x_43.mmap which do not exist for the smaller models. Can this hard coding be fixed ? It will be great if you can take a look at this. Thank you.

Jimskns commented 8 months ago

I've encountered the same issue when adapting the code to opt-6.7b, can anyone help with this? Many thanks!