Are MTP layers included?

#1
by InfernalDread - opened

Hello,

Just wanted to start by saying great work and thank you for creating this version of an already awesome model! Just wanted to ask, since this was a concern with the Air version as well, are the MTP layers included in this version as well?

Thank you for your time.

Cerebras org

Hi @InfernalDread , we did not add MTP layers here (and set num_nextn_predict_layers: 0 in model config, so conversion to llama.cpp should work out of the box). We're working on an update where we calibrate and prune the MTP layers according to the REAP criterion, but for now MTP is just omitted.

oh ok, sounds good, once again, great work!

Sign up or log in to comment