OOM with image 512 x 512 on V100 32GB

[Koruvika]

I've already run Restormer on V100 32GB with image 512 x 512. VRAM significantly increases after attention layers and out of memory at out_dec_level1 = self.decoder_level1(inp_dec_level1). I checked Torch 2.0.1, Torch 2.1.2+cuda121 and Torch 2.3.0+cuda121, all of them had the similarity issue.

[Koruvika]

I runned demo on colab, and seem it works normally, take ~3-4GB VRAM for image size 512 x 512

[alexwongdl]

I meet the same issue, have you found resolution.

[Mytttttttt]

这是来自QQ邮箱的假期自动回复邮件。你好，我最近正在休假中，无法亲自回复你的邮件。我将在假期结束后，尽快给你回复。

[yhao-z]

same issue, memory increases dramaticlly

[flow-specter]

same issue