mira-space/Mira

GPU Memory Requirements

YBYBZhang opened this issue · 6 comments

YBYBZhang commented

Thanks for your excellent work! How much memory GPU is required for text-to-video generation during inference?

Yuanshi9815 commented

I have tried it on different GPUs:

  • 24G A5000 ❌
  • 48G A6000 ✅

The peak memory costing is approximately 25Gb.
I guess it could work well on 24Gb GPU if you optimize the code.

  • ❤️2
YBYBZhang commented

Thanks for your detailed reply!

Yuanshi9815 commented

Update:
Moving the cond_stage_model to cpu after get_learned_conditioning() can make it work well on 24G card.
And the memory coasting will reduce from 23G to 7G during the inference. 🤯

bruinxiong commented

@Yuanshi9815 could you give more detailed instruction for reducing gpu memory cost ?

bruinxiong commented

@Yuanshi9815 It's done. It costs 13G memory during the inference for 384 res.