Ahmadzei's picture
added 3 more tables for large emb model
5fa1a76
So it would need at least 4x model size GPU memory, even with mixed precision as gradient updates are in fp32.