@kimbochen Some folks are going to be at https://t.co/ONKP0yoIyw in Seoul, honestly everyone I have gone to has been great - I’ll try and swing by pastel! Thanks for the tip
where my mxfp6 heads at? Yall feel like commenting on -> https://t.co/9SF3HP6Njo
I like x4 but I can't really see a use case on the hot path where the interesting data movement doesn't require unpack and repack in either case
Training with fewer bits is faster, but quantization error can interfere with learning.
Today on Connectionism, we detail how we trained Inkling 2 Small with a native 4-bit forward pass from the first step of pretraining through the final RL rollout.
https://t.co/CIMECSEpD0