LocalLLaMA

3306 readers

1 users here now

Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.

Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.

As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.

Rules:

Rule 1 - No harassment or personal character attacks of community members. I.E no namecalling, no generalizing entire groups of people that make up our community, no baseless personal insults.

Rule 2 - No comparing artificial intelligence/machine learning models to cryptocurrency. I.E no comparing the usefulness of models to that of NFTs, no comparing the resource usage required to train a model is anything close to maintaining a blockchain/ mining for crypto, no implying its just a fad/bubble that will leave people with nothing of value when it burst.

Rule 3 - No comparing artificial intelligence/machine learning to simple text prediction algorithms. I.E statements such as "llms are basically just simple text predictions like what your phone keyboard autocorrect uses, and they're still using the same algorithms since <over 10 years ago>.

Rule 4 - No implying that models are devoid of purpose or potential for enriching peoples lives.

founded 2 years ago

MODERATORS

[email protected]

Do you quantize models yourself? (lemmy.ml)

submitted 1 week ago by [email protected] to c/[email protected]

7 comments fedilink hide all child comments

Hey fellow llama enthusiasts! Great to see that not all of lemmy is AI sceptical.

I'm in the process of upgrading my server with a bunch of GPUs. I'm really excited about the new Mistral / Magistral Small 3.2 models and would love to serve them for me and a couple of friends. My research led me to vLLM with which I was able to double inference speed compared to ollama at least for qwen3-32b-awq.

Now sadly, the most common quantization methods (GGUF, EXL, BNB) are either not fully (GGUF) or not at all (EXL) supported in vLLM, or multi-gpu inference thouth tensor parallelism is not supported (BNB). And especially for new models it's hard to find pre-quantized models in different, more broadly supported formats (AWQ, GPTQ).

Does any of you guys face a similar problem? Do you quantize models yourself? Are there any up-to-date guides you would recommend? Or did I completely overlook another, obvious solution?

It feels like when I've researched something yesterday, it's already outdated again today, since the landscape is so rapidly evolving.

Anyways, thank you for reading and sharing your thoughts or experience if you feel like it.

top 7 comments

sorted by: hot top controversial new old

[–] [email protected] 0 points 1 week ago* (last edited 1 week ago) (1 children)

Thank you for deciding to engage with our community here! You're in good company.

Kobold just released a bunch of tools for quant making you may want to check out.

Kcpp_tools

I have not made my own quants. I usually just find whatever imatrix gguf bartowlski or the other top makers on HF release.

I too am in the process of upgrading my homelab and opening up my model engine as a semi public service. The biggest performance gains ive found are using CUDA and loading everything in vram. So far just been working with my old nvidia 1070ti 8gb card.

Havent tried vllm engine just kobold. I hear good things about vllm it will be something to look into sometime. I'm happy and comfortable with my model engine system as it got everything setup just the way I want is but I'm always open to performance optimization.

If you havent already try running vllm with its CPU nicencess set to highest priority. If vllm can use flash attention try that too.

I'm just enough of a computer nerd to get the gist of technical things and set everything up software/networking side. Bought a domain name, set up a web server and hardened it. Kobolds webui didnt come with https SSL/TLS cert handling so I needed to get a reverse proxy working to get the connection properly encrypted.

I am really passionate about this even though so much of the technical nitty gritty under the hood behind models goes over my head. I was inspired enough to buy a p100 Tesla 16gb and try shoving it into an old gaming desktop which is my current homelab project. I dont have a lot of money so this was months of saving for the used server class GPU and the PSU to run it + the 1070ti 8gb I have later.

The PC/server building hardware side scares me but I'm working on it. I'm not used to swapping parts out at all. when I tried to build my own PC a decade ago it didnt last long before something blew so there's a bit of residual trauma there. I'm worried about things not fit right in the case, or destroying something or the the card not working and it all.

Those are unhealthy worries when I'm trying to apply myself to this cutting edge stuff. I'm really trying to work past that anxiety and just try my best to install the stupid GPU. I figure if I fail I fail thats life it will be a learning experience either way.

I want to document the upgrade process journey on my new self hosted site. I also want to open my kobold service to public use by fellow hobbyist. I'm not quite confident in sharing my domain on the public web though just yet I'm still cooking.

[–] [email protected] 0 points 1 week ago

Thanks for the tip about kobold, didn't know about that.

And yeah, I can understand that building your own rig might feel overwhelming at first, but there's tons of information online that I'm sure will help you get there!

[–] [email protected] 0 points 1 week ago* (last edited 1 week ago) (1 children)

I think they have you covered with this: https://github.com/vllm-project/llm-compressor

I always find quantizations available since I use gguf models. But I've done it before, it's not hard to do. Pick a method with good performance, it's going to affect your inference speed. I don't know which one is "best" for vllm.

[–] [email protected] 0 points 1 week ago (1 children)

Thanks that looks cool, I'll definitely try and report back. Do you happen to know what the hardware requirements are? I have access 64GB of RAM and 48GB of VRAM across 3 RTX2000e Ada GPUs.

[–] [email protected] 0 points 1 week ago* (last edited 1 week ago) (1 children)

Uh sorry, no. Since I don't use vllm, I don't know. It certainly depends on the method you choose. The activation-aware ones will use a lot of resources. Just truncating the numbers to 8bit (or whatever) uses very little resources, I did that on my laptop. Also depends on model architecture and the size of the model you feed in. Since you gave an 32b parameter model, I'd expect it to take about 64GB loaded fully into memory using 16bit floating point numbers. (32 ~~million~~ billion times 2 bytes.) But I don't really know whether quantization methods load the full thing, or if they do it in chunks. You'd have to google that or try it.

[–] [email protected] 0 points 1 week ago (1 children)

Alright thanks! I found it somewhat difficult to find information about the hardware requirements online, but yeah, maybe I just have to try it.

[–] [email protected] 0 points 1 week ago

You're welcome. If you fail and you can't just add more RAM, maybe have a look at renting cloud servers. For example you can rent a computer on runpod.io for $2 an hour with double your specs. At least that's how I do one-off big compute tasks.