Feature Request: Support GPTQ (Quotes: GPTQModel 4bit can match BF16)

[BodhiHu]

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

Hello,

Great work here :D

Is it possible that llama.cpp support GPTQ quantized models ?
The GPTQ quantized model had the advantage that it was fine-tuned with a dataset, which I think is a good reason to support GPTQ:

Quotes from the GPTQModel repo:

Quality: GPTQModel 4bit can match BF16:

https://github.com/ModelCloud/GPTQModel

Motivation

The GPTQ quantized model had the advantage that it can be fine-tuned/calibrated with a dataset, and for further fine tuning.

Possible Implementation

No response

[BodhiHu]

Hi, noticed there're some GPTQ converter PRs merged before:

#301
#423

But seems it had been deleted from the master branch ? 🤓

[ExtReMLapin]

The GPTQ quantized model had the advantage that it was fine-tuned with a dataset, which I think is a good reason to support GPTQ:

Just like imatrix

[henk717]

Llamacpp has the same techniques and in my understanding in two different ways. Either the iMatrix quants since they are generated based off a dataset or AWQ converted quants.

Thankfully on Llamacpp's side its optional since it savea so much time and hassle when you just want to make a quick quant. That does mean that if you want this that you need to be looking at GGUF's with an imatrix. I know mrademacher uses i1 to indicate that, other repackers you will have to look if an imatrix.dat was bundled or if its mentioned in the description. You only need an imatrix.dat to make gguf's not to use them so its not always there, but if one is there its a good indicator.

[BodhiHu]

Alright... so can we just run the existing GPTQ models from huggingface .etc ?

[henk717]

No because those are different formats for a different older quant method. The same models almost always. have imatrixed GGUF's available and those you want to look for.

[sorasoras]

Hi, noticed there're some GPTQ converter PRs merged before:

#301 #423

But seems it had been deleted from the master branch ? 🤓

it was replace by imatrix and that's why it was deleted