Quantization of larger models on smaller GPUs using CPU offloading about llm-awq HOT 6 CLOSED

Hi, may I know if you are loading and applying the pre-computed scales, or are you doing the AWQ search yourself?

I think our current implementation does not use techniques like offloading to reduce GPU memory usage during the quantization stage. We can definitely add that to reduce GPU memory usage. What do you think @Sakits ?

from llm-awq.

Hey @tonylins,

I was trying to do AWQ search. Since AWQ works layer by layer, the entire model should be loaded into RAM and only the layer that is being quantized should be moved to GPU and then moved back to RAM after quantization.

from llm-awq.

Hi @abhinavkulkarni ,
We have added the CPU offloading support for run_awq in the dev/more_models branch. Now you should able to run awq for opt-6.7b on a single RTX 3060-12G. Welcome to try it out and feel free to bring up any issues you might encounter!

Thanks for your interest in our work!

from llm-awq.

Hi @Sakits, @tonylins,

Thanks for adding support for CPU offloading. I had to make additional changes on top of your branch to run all the steps - run AWQ search for scale and clip values, evaluate using fake quantization, dump AWQ weights, and run AWQ evaluation using quantized weights.

You can view the changes in my forked branch here.
If those look okay to you, I'd be happy to raise a PR.

I had to hard code the value of max_memory according to my GPU (12GB of VRAM) here, but I'd be happy to add --max-memory argument to the command line and remove this hard-coded value.

from llm-awq.

Hi @abhinavkulkarni,

Thank you for your contributions on enhancing the CPU offloading and --max-memory argument!
We'd definitely welcome a pr from you!

from llm-awq.

Thanks @Sakits, I have opened a PR: #22

I was able to run all the steps for mosiacml/mpt-7b-instruct model on RTX 3060 (12GB of VRAM).

from llm-awq.