VAE decoding om AMD works first time only before switching to tiled or black image
- Dominant language
- Python
- Stars
- 133k
- Forks
- 15.7k
- Avg merge
- 1d 6h
- Merged PRs (30d)
- 155
Description
### Custom Node Testing
- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)
### Expected Behavior
Generate images. Not a black image and also don't fall back on tiled VAE decoding.
### Actual Behavior
Only the first generation after startup works perfectly, subsequent images have one of:
- OOM error
- Tiled (and work)
- black image
### Steps to Reproduce
Sorry, AMD GPU user here, i wish i had one of these unobtanium nvidia ones. My card is the 7900XT.
If anything is clear with AMD by now it's that literally every AMD user has a different set of commands that work. My commands arguments that work "right now": `--use-flash-attention --bf16-vae`
### Debug Logs
```powershell
During startup one of the output log lines is:
`Using split attention in VAE` (well actually twice and that on it's own is a red flag, but that aside)
```
### Other
Not to be blunt but this is a classical developer bug. Very often this is a matter of code that's being initialized and then run. And the step to run it again is often a - wise! - optimization where you keep the initialized class and re-run it. I'm guessing that's the case here too. If you point me to the files that do the VAE decode then i'd be happy to look at it.
The fact that this runs once (and does so on each restart just once) means that it must be possible to get it working.
Contributor guide
Assessment
This issue has not been assessed yet.