Decreasing render speed with number of lora's added Flux dev, is this normal?

[alaman64]

Hi I noticed that my render speed decreases based on the number of lora I add. I'm on Flux Dev, RTX 4090. As you can see the speed drops to about 1/2 with 2 lora's loaded. So 20 second render with 0 lora takes 40 second with 2 lora's. IS this normal? or are there parameters I can change to improve this? Thanks.

flux1-dev-fp8.safetensors, automatic FP16, GPU Weight 21568

0, Lora = 3.16 It/sec
1, Lora = 1.88
2, Lora = 1.44
3, Lora = 1.18

[FifthDream]

Oh, i hope someone figures out the power behind that. On my end: 0 loras: 9-13 s/it. One lora: 70 - ∞ s/it. Sigh.

[alaman64]

What card are you working with?

[FifthDream]

3060 Ti

[SwedishViking404]

I have a 4090, and it is usual to see an increase in seconds per iteration as more loras are added as well as a commensurate increase in GPU memory usage per lora. In your case, I would say that it is a memory issue. You could try experimenting with the Swap Method, Swap Location and GPU Weights.

[iqddd]

If you can, it's best to use the Automatic method rather than the Automatic (FP16 LORA) method. This one merges Loras with the model. If you have enough RAM for the operation, you'll barely notice any slowdown when using Loras. Plus, it doesn't depend on the number of Loras. Just a heads-up, though: the process can take a while and happens every time you change Loras or their weights.

[alaman64]

Thanks, switching to automatic appears to have fixed the issue.