logoalt Hacker News

embedding-shapetoday at 9:40 AM0 repliesview on HN

It shouldn't? Unless you're using BF16 for all weights (I'm using NVFP4 for the text encoder, otherwise everything BF16 (and audio F32)) you'll fit it all within 96GB VRAM, bugs non-with-standing :) I've been fitting this within 96GB VRAM without issues.