Sign In

Use QPOLA optim in AI-toolkit

0

Use QPOLA optim in AI-toolkit

How to use QPOLA optim in AI-toolkit.

What? QPOLA (紹介):
https://civitai.red/articles/32677/v104-optim-qpola-latest-optim-qpola-released

Rewrite the optimizer.py file located in “.AI-Toolkit\toolkit” as shown below.

".AI-Toolkit\toolkit" に配置されている "optimizer.py" を以下のように書き換えます

https://github.com/ostris/ai-toolkit/commit/676b4f3c4c6055bd2795384859bb3890e386e453#commitcomment-194533639

Place qpola.py and qpola_kernel.ptx into “.AI-Toolkit\toolkit\optimizers”.
Launch AI-Toolkit, open “Show Advanced”, and change the setting to [optimizer: "qpola"].
After that, you can use it as usual.

".AI-Toolkit\toolkit\optimizers" へ qpola.py、qpola_kernel.ptx、 を配置します
AI-Toolkit を起動し 「Show Advanced」 を開き [optimizer: "qpola"] へ変更します
あとは通常の使い方で大丈夫です

:Note (注意):

Since AI-Toolkit requires you to rewrite optimizer.py every time it is updated, you can reduce the effort by keeping a duplicate and simply switching the file name back when needed.
If Ostris incorporates this update, any optimizer will become usable just by placing it in the folder.
If someone is able to reach out on Discord, please ask Ostris to consider implementing this.

AI-toolkit の更新のたびに、optimizer.py の書き換えが必要になるため、複製をしておいて ファイル名変更 で戻すようにすると手間を減らせます、
Ostris氏がこのupdateを取り入れてくれたら、どんなoptimもフォルダに配置するだけで使えるようになります、
もしDiscordでお願いできる方いれば Ostris氏へ検討を依頼してみてください

:Verification (検証):

With the Anima model, both 4bitconvrot and 8bitconvrot, QPOLA successfully completes training in a stable manner.
On 4bitconvrot, the resolution was 1024px, and VRAM usage stayed below 3.8 GB (works even on a laptop!).
We confirmed that anyone can easily perform LoRA training, so please give it a try.

Anima モデル、4bitconvrot、8bitconvrot、両方で QPOLA は安定的に学習完了します
4bitconvrot では、解像度:1024px、VRAM:3.8GB以下でした(ノートPCでもOK!)
誰でも簡単に LoRA 学習を行えることが確認できました、ぜひお試しください

0