search

lllyasviel / stable-diffusion-webui-forge

GNU Affero General Public License v3.0
8.02k stars 781 forks source link

Forge crashes with Flux dev on 3090 #1901

Open salesperson opened 3 weeks ago

salesperson commented 3 weeks ago

I have RTX 3090 with 24gb vram and 32gb of RAM. Any other UIs are working perfectly well, but I cannot make forge working. I`ve tried changing GPU Weights to 23000, but no luck - the same error. Please help.

Here`s my setting: image

and here`s the error I get: image

Haoming02 commented 3 weeks ago

Refer to: https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981

HMRMike commented 3 weeks ago

Keep the GPU weights about 23500, and check if your Page File settings in windows allows for about 40GB, or set to auto management but expect this sort of size or more.

salesperson commented 3 weeks ago

@HMRMike thanks. Works now, but speed is very low. In ComfyUI 1 image in 20 steps take 45 sec, here it takes a couple of minutes, while system moves models and stuff to free memory. That`s odd.

HMRMike commented 3 weeks ago

@HMRMike thanks. Works now, but speed is very low. In ComfyUI 1 image in 20 steps take 45 sec, here it takes a couple of minutes, while system moves models and stuff to free memory. That`s odd.

The initial loading+first image do take a while. I tested right now with a 3090 also- (with t5 GGUF Q8 instead of fp8) and it indeed took 2:20 min. for 1024X1024 at 20 steps. But subsequent images will be faster. Changing the prompt completely- finished an image in1 minute. Different seeds per same prompt took at most 41 sec. I don't know how comfy manages that stuff but It would be quite a feat if it did all that initial loading twice as fast

Juqowel commented 3 weeks ago

You can also try https://github.com/Juqowel/GPU_For_T5 for speed up second+ image with cpu load. But t5 gguf - unsupported.

LIQUIDMIND111 commented 3 weeks ago

Keep the GPU weights about 23500, and check if your Page File settings in windows allows for about 40GB, or set to auto management but expect this sort of size or more.

bummmer, i did a w hole installation on COMFY UI to test and compare to FORGE using FLUX DEV fp8 and GGUF q8, and FORGE renders wayyyy faster for about 6/23 seconds per iteration and comfy was 9 seconds per iteration !!!