Replies: 3 comments 2 replies
-
Am I the only one who has this problem? |
Beta Was this translation helpful? Give feedback.
-
Same here (RTX 3060 12 GB); for a 768x1024 it drops from 1.8 it/s to 13 s/it before returning to normal on the next generation (even if the next generation is the same lora and prompt). One pattern I notice is that the Model Memory shows 0.0 only for these slow generations. Other than that there's no difference to the output from the commandline. [Memory Management] Current Free GPU Memory (MB) = 3601.97705078125 When I check Task Manager (this is Windows), the CPU usage is low and the GPU usage high, so it is using the GPU to compute, but maybe it is reading the model or writing the output from main memory? |
Beta Was this translation helpful? Give feedback.
-
Turn on the resource manager when generating. This is a known bug of Forge - sometimes VRAM leaks when LoRa is connected. |
Beta Was this translation helpful? Give feedback.
-
I'm encountering significantly increased image generation times (upwards of 5 minutes) when loading certain LoRAs for the first time, then after few generations, the speed is normal again (around 15 seconds).
Any Ideas on how to resolve this issue.
Hardware:
CPU: [Intel Core i7-8700K]
GPU: [NVIDIA GeForce RTX 4070Ti]
VRAM: [12GB]
RAM: [32GB]
Beta Was this translation helpful? Give feedback.
All reactions