Windecay

Results 14 comments of Windecay

> The pinned_memory by accident fixes the slowdown when memory estimation is a bit off on Windows, but we will work out that issue soon as well since it can...

> > > The pinned_memory by accident fixes the slowdown when memory estimation is a bit off on Windows, but we will work out that issue soon as well since...

> this is by design, image prompts are intentionally cover current inpainting zone (outpainting is just a form of inpainting) thank you !

> > since I'm doing tests with loras ( changing their weights ) then it makes sense it is always apply this. my problem is more the time it takes,...

loaded completely 9.5367431640625e+25 1560.802734375 True ..

> > [@rattus128](https://github.com/rattus128) No oom, but vae decode node crash ,no logs ,just crash. --use-sage-attention --fast pinned_memory > > ![Image](https://github.com/user-attachments/assets/09f878c5-e87c-46a4-b09a-9f60e07e4e74) > > The same problem. But I use pytorch2.10 nightly...

> 钩子的实现方式是和非flux版本的pulid插件一样的 谢谢你,我再研究对比下显存占用曲线。要在低配机器上防止OOM,要合理分配。

comfy开启后默认状态约0.5GB占用,pulid生成一次后,显存残留占用2.3GB。 修改代码加入del eva_clip后,生成一次后占用1.5GB。后续每次一样,运作正常。 加入del pulid_flux清理无效。还是1.5GB 应该是"pulid_model": pulid_flux,被引用太多次了。研究了一天没弄好~

> > comfy开启后默认状态约0.5GB占用,pulid生成一次后,显存残留占用2.3GB。 修改代码加入del eva_clip后,生成一次后占用1.5GB。后续每次一样,运作正常。 加入del pulid_flux清理无效。还是1.5GB 应该是"pulid_model": pulid_flux,被引用太多次了。研究了一天没弄好~ > > 你说的2.3G和1.5G指的是加了模型卸载节点之后占用量,还是运行后你显存的剩余量(不知道总显存多少),我在最后加了一个模型卸载节点,运行后,显存占用始终是3.5G 我的数据是使用了简单采样工作流后总残留的显存数量(监控python.exe进程)。普通的模型卸载节点对这剩余数据没有改变。

> ![Image](https://github.com/user-attachments/assets/f7fc1116-0a6e-48ac-91ea-3fe213fb5a0c) > > 你如果像运行完之后完全卸载掉数据,可以用我另一个插件的这个节点打开第一个开关,缺点就是每次执行都会重新执行加载模型节点 这确实是一个很彻底的办法。 万不得已才使用。 你的新版节点我已经同步更新~有空继续研究。