5fa1a76
1
So it would need at least 4x model size GPU memory, even with mixed precision as gradient updates are in fp32.
