DiffSynth-Studio icon indicating copy to clipboard operation
DiffSynth-Studio copied to clipboard
verified publisher icon modelscope

use_gradient_checkpointing_offload occupies more VRAM when train i2v model

Open mengfanShi opened this issue 10 months ago • 2 comments

When training the I2V wanx model, using use_gradient_checkpointing_offload occupies more VRAM than using use_gradient_checkpointing. If you have time, could you please take a look? Thank you.

mengfanShi avatar Mar 13 '25 12:03 mengfanShi

@mengfanShi If you enable use_gradient_checkpointing_offload , use_gradient_checkpointing should also be enabled. Please make sure they are enabled together.

Artiprocher avatar Mar 14 '25 01:03 Artiprocher

Yes, that's the problem. Adding it makes it normal. Thank you very much.

mengfanShi avatar Mar 14 '25 02:03 mengfanShi