HN Explorer

3 comments

[ 2.9 ms ] story [ 17.5 ms ] thread
danielhanchen
Excited to have collabed on this! Thanks electroglyph for the contrib!
storystarling
Do the memory savings carry over to inference or is this strictly optimizing the backward pass? I'm running embedding pipelines via Celery and being able to squeeze this into lower VRAM would help the margins quite a bit.