HELP!
Maybe it's because I have a 2080ti but this model takes like 5mins to make 20 step image and I don't know why. I could run SD3 heaviest model with almost no problem. Any help would be appreciated. I don't think it's my system causing it to run THIS slow.
My specs.
2080ti
32gb of DDR5
14900k (latest and greatest)
As you can see, my PC was recently upgraded and I have the latest hardware (outside of my 2080ti but the 2080ti is still 12gb vram)
I've installed everything correctly (I think.)
- Flux_dev and flux_dev8 are both installed in UNET folder. Both run slow but dev8 is a little faster.
- weight type - tried all 3 with no noticeable difference. (default, fp8_e4m3fn, & e5m2)
- Clip1 - t5xxl_fp16. Also tried fp8.
- Clip2 - Clip_I
- type -Flux
No matter what settings I change, it still runs SOOOOOOOOOOOO slow. Even using the 8 model which doesn't take all my VRAM. Which makes me think it's not a VRAM of GPU problem but something with my comfyUI software.
The only weird message I get is: "Model doesn't have a device attribute." This message seems like a problem but I don't know what could be causing it....
Even though it's only a 2080ti, I've NEVER had this type of problem trying to run any model before. Merging 3 different models in comfyui - no problem. Running the SD3 equivalent of this model - no problem.
Any help would be so amazingly appreciated! If you can help me fix it, I'll gift you buzz in Civitai!
Outside of the last message that I mentioned that happened at the start, this is what it says after making the photo. (it does make a photo but it takes FOREVER.)
Using pytorch attention in VAE
Model doesn't have a device attribute.
Requested to load AutoencodingEngine
Loading 1 new model
Prompt executed in 281.95 seconds (281.95 seconds!!!!! 4.7 mins for a 20step image!!!! (sad face emoji))
That's definitely too slow. My 4070 12G works for ~2.5s/it so under 1 min for 20 steps, and both t5 and unet in 16 bit.
Have you checked GPU usage during execution? You can see it via nvidia-smi.
A lot of people with the latest updates have the same "Model doesn't have a device attribute." message. Some people experience that it takes about twice as much time to generate one image. I have the same error, but I can't say that I notice much difference speedwise.
Have the same issue on 3070ti with 12Gb, and threadripper 7970x with 64 cores and 128Gb DDR5. 28min for 1 image with 50 steps. This is around 50 times slower than Stable Diffusion 3 with similar settings. What could be the problem?
There was a new update for Comfy earlier, it solved the issue for me.
https://github.com/comfyanonymous/ComfyUI/commit/e9589d6d9246d1ce5a810be1507ead39fff50e04