Note on Decentralized Automated Scientific Inference

Undescribed Horrific Abuse, One Victim & Survivor of Many gmkarl at
Mon Jan 9 03:46:37 PST 2023

The galactica model is trained on scientific documents. Its large form
has 120B parameters.

Each layer is about 10GB in float16 precision. The bigscience/petals
system for p2p inference splits by layers, so each node would need 5GB
vram for 8 bit inference, 20GB vram for full precision.

bigscience/petals currently requires manual coding for the client and
server model interfaces.

galactica inference alone is unlikely to be easy to add new full
documents to without adaptation or finetuning, as like most models it
was trained with a context size of only 2k words. adaptation and
finetuning can be done p2p.

More information about the cypherpunks mailing list