cancel
Turn on suggestions
Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.
Showing results for 
Search instead for 
Did you mean: 

AI Discussions

hinst
Adept I
a week ago

AMD AMUSE v2.1.11 potential memory leak

Good day. I come to report a potential bug in AMD AMUSE v2.1.11

System info:

  1. Windows 11 Home, version 23H2
  2. Graphics card: Radeon RX 7800 XT
  3. Graphics driver: Adrenalin Edition 24.9.1

Steps to reproduce:

  1. Load model: Juggernaut XL
  2. Set resolution: 1280 x 1280, steps=100
  3. Generate one image by clicking "Generate" button
    1. While you wait for image generation, open Windows Task Manager, go to GPU tab
  4. Observe VRAM usage over time
  5. Video memory usage is around 9 GB
  6. At the moment when image generation process is finished, the memory usage spikes up to 20 GB -> 16 dedicated + 4 shared
    1. This is not good because we have now filled all the available memory, and it stays occupied filled, despite the image generation process being already finished
  7. Generate more images with the same settings
    1. The image generation process will get slower and slower over time, eventually crashing or freezing
  8. Use batch image generation on the "Automation" tab -> the same problem will happen.
    1. The first image is always good, but next images will use too much VRAM

It seems like VRAM does not get released at some point where it probably should be released.

9GB should be enough because it is always enough for the first image, but the consequent images always struggle.

The workaround is to unload the model every time manually by clicking the red "x" button below Model Selector

Labels
0 Likes
1 Reply
Dakotapilcher
Journeyman III
Monday

You're encountering a potential memory leak with AMD AMUSE v2.1.11 during image generation, specifically with the Radeon RX 7800 XT. Based on your behavior, the VRAM is not released correctly after each image generation, leading to excessive memory usage that eventually causes performance degradation and crashes.

The fact that the memory spikes after the first image and doesn't clear itself indicates inefficient memory management within the application. This is consistent with what you've observed during single and batch image generation processes.

For now, the workaround you mentioned—manually unloading the model after each image—seems the best option. However, this is not a long-term solution. I would recommend reporting this issue directly to AMD’s support team, as it could be an underlying issue with how AMUSE handles VRAM or a driver-level problem with the Adrenalin Edition 24.9.1.

To help resolve this:
1. Check for updates: Make sure both AMUSE and your GPU drivers are current. Sometimes, new patches address these types of memory management issues.
2. Submit a bug report: Send detailed feedback to AMD through their bug reporting tool. Provide the steps you've outlined to reproduce the issue and mention the memory spikes and crashes, as this can help their team identify and resolve the problem.

This could help them develop a fix in future updates, especially if others are experiencing similar issues with Radeon GPUs and AMUSE.

0 Likes