RTX 4090 speed 1.22-1.6 it/s FLUX [dev] 09-16s for 1 pic

[TimmekHW]

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:15<00:00,  1.33it/s]
To load target model IntegratedAutoencoderKL███████████████████████████████████████████| 20/20 [00:14<00:00,  1.25it/s]
Begin to load 1 model
[Unload] Trying to free 8579.84 MB for cuda:0 with 0 models keep loaded ...
[Unload] Current free memory is 11478.45 MB ...
[Memory Management] Current Free GPU Memory: 11478.45 MB
[Memory Management] Required Model Memory: 159.87 MB
[Memory Management] Required Inference Memory: 8372.00 MB
[Memory Management] Estimated Remaining GPU Memory: 2946.58 MB
Moving model(s) has taken 0.78 seconds
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:16<00:00,  1.22it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:16<00:00,  1.25it/s]
A whimsical scene unfolds: a young woman, her features reminiscent of a fox, sits regally on a weathered tree stump, verdant foliage enveloping her in emerald hues. Her vibrant orange fur, tipped with white and fluffed to perfection, cascades down her back like a fiery mantle. Soft, warm lighting bathes the scene, with subtle bokeh blur gently obscuring the lush surroundings, focusing attention on the enchanting subject.
Steps: 20, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Distilled CFG Scale: 3.5, Seed: 1354291294, Size: 896x1152, Model hash: 06f96f89f6, Model: flux_dev, Version: f2.0.1v1.10.1-previous-450-gdb6448df

Driver 560.94

I use FLUX[DEV] fp8 because it gives better quality and variety than Q4, while consuming the same amount.
Clip_L and T5 are also already sewn in there, which saves video memory. With 1024 x 1024 and a long prompt, consumption does not exceed 14 GB of VRAM

If anything, write t.me/timmek1

[TimmekHW]

UPDATE: I read the lllyasviel article and realized that I was using the wrong NF4. So in theory, NF4 should be better