25
Multi-GPUs considerations (downonthestreet.eu)
submitted 1 day ago* (last edited 1 day ago) by Shimitar@downonthestreet.eu to c/localllama@sh.itjust.works

I have purchased an M2 to PCIe adapter and connected an additional GPU (an old RTX 6000, turing, 24gb VRAM) to my setup. This allowed me to run qwen 3.8 27B at Q8_K_XL (UD) instead of Q6_K_M.

I am doing some image transcription tasks (writing down old handwritten cursive letters) and the different is vastly HUGE.

With Q6 it would barely decyper the letters, then i had to manually recheck everything. With Q8, wow, it's almost flawless...

Using llama.cpp, with caches quantized to k8 and 200k context window. Using a subagent to to the hard work, and having the main agent just supervise. In serie, since i can run only one qwen instance at the same time anyway. I have NCCL enabled as well, which gave me a nice bump in prefill.

Setup is (all three on CPU bound PCIe lanes) two RTX A4000 (Ampere) 16+16GB VRAM on a PCIe bifurcator (16x -> 8x+8x) and the RTX 6000 (turing) 24GB VRAM on an M2 -> PCIe 4x slot. Everything at PCI 3.0, since 4.0 is unstable on the bifurcator.

Musing if i should add another RTX A4000 on my chipset bound PCIe lanes... I am worried i will lose NCCL by doing that...

Prefill is between 300 and 600 t/s, generation is around 30-40 t/s with tensor split and some massaging batch sizes.

I keep all three GPUs capped at 100W (100+100+100 = 300W) to save on electric bill... (could go up to 140+140+260 = 540W, but the t/s gain is minimal)

Now, the fun part is how the fuck to fit everything somehow.... Will try to have some fun with alluminum parts i have. And fuck cables... Long ones have signal issues, and short ones leave you with no headroom for placement....

you are viewing a single comment's thread
view the rest of the comments
[-] Shatur@discuss.tchncs.de 1 points 20 hours ago

It's strange because the multimodal projector doesn't depend on the model weights. It's downloaded separately as a BF16 file.

[-] Shimitar@downonthestreet.eu 1 points 12 hours ago

What does mtp have to do with handwriting OCR?

[-] Shatur@discuss.tchncs.de 1 points 9 hours ago

I'm not talking about MTP. I'm talking about MMPROJ, which is used for image recognition. Here's the file I'm talking about.

[-] Shimitar@downonthestreet.eu 1 points 3 hours ago

Ah, yes, understood ...

The difference is in how the model elaborate the overall images, like detecting areas which are "background" or noise, not actual writings. Side writings, signatures, and so on.

It's aneddotical, but once for example Q6 insisted in parsing two columns of text in the wrong order, where Q8 immediately understood the right order.

Actual letters and words detection also seems to be heavily affected by overall text detection, means when the model get the meaning, also can improve detection. in this case is the post recognition analysis that catches errors and fix them

[-] Shatur@discuss.tchncs.de 1 points 2 hours ago

I believe you when you say you noticed a difference ๐Ÿ™‚ I just think it might be related to other factors, like your inference engine update, drivers, or even the model configuration. You mentioned that you tried Q6 a while ago, so other things might have changed since then.

I'm not saying it's impossible for Q8 to outperform Q6, but the difference between them is usually unnoticeable in practice.

I'd suggest retesting Q6, but since you can run Q8 and don't need the extra VRAM, I don't see a reason to bother ๐Ÿ™‚

[-] Shimitar@downonthestreet.eu 2 points 2 hours ago* (last edited 2 hours ago)

Well VRAM is never enough :)

Surely I can use more :)

Parallel slots more context, a second model ...

So I will test the Q6 again, maybe downloading a newer file for it ...

Most probably you are right anyway. I upgraded adding the third GPU with 24gb VRAM and decided to fill it up. Moving the mmproj in vram is nice, and also have both ComfyUI and llamacpp loaded at the same time is.

I was hoping to increase my context enough to have parallel = 2 but not on Q8....

And i don't have enough vram anyway to run a decently quantized bigger model anyway.

[-] Shatur@discuss.tchncs.de 1 points 2 hours ago

Let us know how it goes ๐Ÿ™‚

this post was submitted on 15 Nov 2026
25 points (96.3% liked)

LocalLLaMA

5662 readers
270 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 3 years ago
MODERATORS