Using NVIDIA GeForce RTX 3090 24G GPU, using DPM++2M Karras with steps of 201024 * 1024 to generate a graph at a speed of 2.55 it/s. Is this speed normal?

[hjj-lmx]

How can you optimize the speed? If you use Controlnet, the speed will be slower. Do you have any good suggestions?

[78Alpha]

With TensorRT you will hit a max of 5 it/s, and that is the limit of the 3090 for single image. FP8 does not give speed improvements on the Ampere tensor cores, only memory savings.

[hjj-lmx]

After TensorRT installation is completed, starting the webui will result in an error and it seems that Controlnet cannot be used

[hjj-lmx]

Every time the webui is just started, the memory is occupied by more than ten gigabytes