peft VS GPTQ-for-LLaMa

There is not a lot of great content out there making this clear, but basically all that matters for basic fine tuning is how much VRAM you have -- since the 3090 / 4090 have 24GB VRAM they're both pretty decent fine tuning chips. I think you could probably fine-tune a model up to ~13B parameters on one of them with PEFT (https://github.com/huggingface/peft)

Hi everyone. Recently I've been looking into the PEFT library (https://github.com/huggingface/peft) and I was wondering if it would be possible to do prompt tuning with OpenAI's Whisper model. They have an example notebook for tuning Whisper with LoRA (https://colab.research.google.com/drive/1vhF8yueFqha3Y3CpTHN6q9EVcII9EYzs?usp=sharing) but I'm not sure how to go about changing it to use prompt tuning instead.

In the coming days, we'll work on sharing scripts to train models, optimizations for on-device inference, even nicer demos (and for more powerful models), and more. Feel free to like our GitHub repos (transformers, peft, accelerate). Enjoy!

image

I discussed part of this here: https://github.com/huggingface/peft/issues/123

It seems like the merge_and_unload function in this PEFT script might be what they are referring to: https://github.com/huggingface/peft/blob/main/src/peft/tuners/lora.py

To obtain the original llama model, one may refer to this doc. To merge a lora model with a base model, one may refer to PEFT or use the merge script provided by LMFlow.

you could use a callback, like show here, https://github.com/huggingface/peft/issues/286 and call code to merge them here.

git clone https://github.com/oobabooga/GPTQ-for-LLaMa

If you're using the model directly from ehartford, that one hasn't been quantized. Try using the GPTQ quantized version here, and use this fork of GPTQ-for-LLaMa. Load in 4-bit with --wbits 4

This worked for me on Ubuntu. If you want to use the CUDA branch instead of triton, do the same steps except clone this GPTQ-for-LLaMa fork and run python setup_cuda.py install

If you don't have triton and you use AutoGPTQ you're gonna notice a huge slow down compared to the old GPTQ-for-LLaMA cuda branch. For me AutoGPTQ gives me a whopping 1 token per second compared to the old GPTQ that gives me a decent 9 tokens per second.. both times I used a same sized model. (I think the slowdown is due to AutoGPTQ using the newer cuda branch which is much slower than the old one)

Are you using a later version of GPTQ-for-LLaMa? If so, go to ooba's CUDA fork (https://github.com/oobabooga/GPTQ-for-LLaMa). That's what I made it in and it definitely works with that. And that's what's included in the one-click-installers.

This usually happens when using models that conflict with your GPTQ installation. You should be using this fork: https://github.com/oobabooga/GPTQ-for-LLaMa. If you did the manual installation wrong, use the one click installer instead.

cd text-generation-webui # wherever you have it installed mkdir -p repositories cd repositories git clone https://github.com/oobabooga/GPTQ-for-LLaMa -b cuda GPTQ-for-LLaMa cd GPTQ-for-LLaMa python setup_cuda install

I'm on GPTQ for LLaMA folder under repositories says it's pointed at https://github.com/oobabooga/GPTQ-for-LLaMa.git. But I've run through the instructions and also applied the monkey patch to train and apply 4 bit lora which may come into play. No idea.

git clone https://github.com/oobabooga/GPTQ-for-LLaMa.git -b cuda