lora VS kohya-trainer

Diffusion unet has an "extended" version nowadays that applies to the resnet part as well as the cross-attention: https://github.com/cloneofsimo/lora

Absolutely. But that doesn't matter because you only have to train it at scale, once. There are papers released already that show it's possible to update weights in small sections. You won't have to wait for the next monolithic LLM to drop to get up to date information. It will start to learn in bits and pieces.

No, it's a deep learning thing

Low Rank Adaptation of Large Language Models.

Recently, I have seen the LoRA technique (Low-Rank Adaptation of Large Language Models) as a popular method for fine-tuning LLMs and other models.

Enter LoRA, a method proposed for adapting pre-trained models to specific tasks[2]. By freezing pre-trained model weights and injecting trainable rank decomposition matrices into the transformer architecture, LoRA can reduce the number of trainable parameters and the GPU memory requirement, making the adaptation of LLMs for downstream tasks more feasible.

Open-source LLM projects have largely solved this using Low-Rank Adaptation of Large Language Models (LoRA): https://arxiv.org/abs/2106.09685

Apparently an RTX 4090 running overnight is sufficient to produce a fine-tuned model that can spit out new Harry Potter stories, or whatever...

Alpaca was trained for $600 ($100 for the smaller model) and offers outputs competitive with ChatGTP. https://arxiv.org/abs/2106.09685

This longer colab notebook: I did use this one (or one of the slight derivatives of it) and got out a safetensors file, but the lora didn't work at all--I'd use it a increase it's weight but I just would see no effect

It isn't what you are asking for (sry) but I struggled with this thing for way too long until I found out about the Kohya Trainer. https://github.com/Linaqruf/kohya-trainer So much easier with a lot of videos by the various YT folks. Standalone WebUI that just works. Life is good here!

It's seem that he is stable diffusion model creators. In that space, it's less knowing about the code and more experimenting on what would happen in the training. The stable diffusion community has repertoire of fine-tuning tools that is accessible for someone who have no single idea on the code behind it, no different than using application like kohya.

Have you tried out one of the colabs from https://github.com/Linaqruf/kohya-trainer ? The colabs themselves are pretty long, but you just have to read each step and then usually push the button to run that cell, then move on to the next one.

** https: //github.com/linaqruf/kohya-trainer**

There are a lot of variables that affect what kind of settings to use, but afaik the best solution to finding the right step count for what your training is still just to save multiple epochs and then run a x/y/z plot comparison. If you can't do that locally because of your 4gb card, you could try using Lora colabs that include inference capabilities.

You seem to be a little confused. You wont find an ipynb of a model. You would reference a model via a content portal like hugginface. If your model is hosted there, you dont have to download it to your computer or gdrive first. You just reference it with the hugginface-style reference, ie runwayml/stable-diffusion-v1-5. Some colabs will let you also reference a URL to pull down the model. Example. https://github.com/Linaqruf/kohya-trainer/blob/main/kohya-LoRA-dreambooth.ipynb. In that case, you can get the direct url to a checkpoint, for example at civit.ai. If you're decent at messing around with code, you can deconstruct that code block to use in a different colab. As for gdrive, it's only a couple dollars to get 100G.

Try this colab to train Lora, it can generate image without the UI too