While I am lucky to have a 3090 at home, my work laptop has a Quadro P5200 with 16gb vram, die based on GP104, almost a gtx1080. Old but can still generate. XL is nice and nippy but Krea is sluggish to say the least when it doesn't OOM.
I've attempted to use various attentions but had no luck with sageattn, flashattn, flash-legacy or kitchenattn... xformers works but I am wondering if there is anything else I might be able to add to my install? (Forge Neo). Pytorch 2.7, cu128. Or any advice on utilising the gpu with anything optimised for this era of card for SD.
Thanks in advance, any positive insight is greatly appreciated, I have spent days tweaking (or attempting to and failing), my limited knowhow is exhausted!
ADDED Oct 7th. Some tests using variants.
FORGE NEO
Quadro P5200 16gb Vram, 32gb ram
qwen3vl_4b_fp8_scaled.safetensors - text encoder
Qwen2D_VAE.safetensors - VAE
Xformers 0.0.30
1024x768
Prompt, no loras, "dog on the beach"
completion:
GGUF Q5
Time taken: 5 min. 6.5 sec.
NVFP4
Time taken: 4 min. 48.8 sec.
GGUF Q4
Time taken: 4 min. 41.6 sec
W4A8_convrot
Time taken: 2 min. 6.3 sec.
INT8_CONVROT
Time taken: 1 min. 26.1 sec.