Datasets:
Tasks:
Text Generation
Modalities:
Image
Formats:
imagefolder
Languages:
English
Size:
< 1K
License:
What CUDA hardware was this run on?
#2
by rhambus - opened
I am interested in your numbers, but what GPU did you use on this workstation?
RTX 5070, 32GB DDR5 RAM (IIRC this set was at 5200mt/s)
It is possible to run something similar on 8GB RAM with no expert layers VRAM resident (lower speeds obviously). The general ceiling in my experience with 3.6 35a3b is ~10 on 12GB VRAM if you have real world context limits of 60-120k tok. I recommend q8 for KV quant. Start with 6-8 layers offloaded and see how things perform if you’re at 12GB