Spaces:

hf-accelerate

/

model-memory-usage

like 623

Paused

Hi, i am wondering why Gradient calculation in float16/bfloat16 costs more vram than float32? I searched on the Internet, but i cant find a explanation about this.