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 must log in or register to comment.
Will just point out there are specialized models for OCR that would be less resource intensive. Also multiple agents batching can be more efficient usage of an instantiated model.
Still, I expect you’re having fun with hardware, and project ambition rises to use all available hardware pretty naturally. Enjoy.
OCR models seems to work mostly on printed text. Handwriting is… Wonky. In my (little) testing, Qwen 3.8 fares better than OCR models in this use case.
I am using one subagent, to split contest or each image do weight too much. Only one agent because I can only run one at time anyway…
If I where to find out how to run multiple on the same hardware…
Surprised to hear about the big difference between Q8 and Q6. Some GGUF makers (like bartowski) even recommend Q6 (but I also use Unsloth quants, they even recently released 3.1).
I might have overestimated the gain, and surely my Q8 is freshly downloaded while the Q6 is oldish… Both UD.
But there is a visible difference in image to text over handwriting recognition!
I dont have any answers for you, but I’m also curious as I am adding another p100 to my setup. I run my current card on a x4 thunderbolt dock, and the performance so far is fine.
I’m probably gonna have to build an actual case/setup for it as well.
Maybe if i come around in building something for my setup will share photos in a new post. It’s already a custom (wooden) case with an ATX metal stand mounted inside. I need to build a second layer but maybe with alluminum struts this time (i have some floating)
