Flux Krea using 40GB+ VRAM with SDNQ - is this normal? #4104

SAC020 · 2025-08-09T05:03:07Z

SAC020
Aug 9, 2025

Invalid issue, I had batch size 16. Works fine. Can't figure out how to delete the post.

===

Trying to use Flux Krea on a 4090 - I'm not sure what optimization should I configure, or if everything is working as expected

With SDNQ int4, it still consumes 40GB+ VRAM, making it unusable.

07:54:34-138856 DEBUG    Sampler: "Default" cls=FlowMatchEulerDiscreteScheduler config={'num_train_timesteps': 1000,
                         'shift': 3.0, 'use_dynamic_shifting': True, 'base_shift': 0.5, 'max_shift': 1.15,
                         'base_image_seq_len': 256, 'max_image_seq_len': 4096, 'invert_sigmas': False, 'shift_terminal':
                         None, 'use_karras_sigmas': False, 'use_exponential_sigmas': False, 'use_beta_sigmas': False,
                         'time_shift_type': 'exponential', 'stochastic_sampling': False}
07:54:34-158289 INFO     Processing modifiers: apply
07:54:34-170190 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:54:34-175150 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:54:34-178126 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:54:34-181598 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:54:34-191531 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:54:34-196571 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:54:34-200032 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:54:34-203617 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:54:42-110565 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:54:44-589340 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:54:44-592663 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:54:44-595616 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:54:46-954211 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:54:49-144542 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:54:49-147813 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:54:49-152294 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:54:51-544762 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:54:53-572740 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:54:53-576669 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:54:53-581871 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:54:55-762722 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:54:57-695860 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:54:57-700733 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:54:57-706728 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:54:59-883796 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:01-639662 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:01-643178 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:01-647253 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:03-904140 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:05-784039 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:05-788007 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:05-792674 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:07-927709 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:09-719841 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:09-723663 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:09-727631 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:11-874752 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:13-749382 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:13-753047 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:13-757085 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:15-919180 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:17-618687 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:17-622010 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:17-625573 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:19-803216 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:21-493184 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:21-496657 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:21-500128 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:23-633098 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:25-554942 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:25-558315 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:25-563598 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:27-763807 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:29-549752 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:29-554712 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:29-558708 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:31-687908 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:33-404766 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:33-407861 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:33-411059 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:35-667095 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:37-361461 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:37-364437 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:37-369398 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:39-470091 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:41-274566 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:41-278326 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:41-281708 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:43-390081 TRACE    Server: alive=True requests=154 memory=23.14/63.92 status='running' task='Load'
                         timestamp='20250809075224' current='task(1iagcw0lq9l74uz)' id='d06d22e01827487' job=0 jobs=1
                         total=1 step=0 steps=0 queued=0 uptime=206 elapsed=199.09 eta=None progress=0
07:55:43-404024 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:45-210176 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:45-214804 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:45-220421 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
07:55:45-245913 INFO     Base: pipeline=FluxPipeline task=TEXT_2_IMAGE batch=1/1x16 set={'prompt_embeds':
                         torch.Size([16, 512, 4096]), 'pooled_prompt_embeds': torch.Size([16, 768]),
                         'negative_prompt_embeds': None, 'guidance_scale': 4, 'generator': 'cuda:[84358502, 84358503,
                         84358504, 84358505, 84358506, 84358507, 84358508, 84358509, 84358510, 84358511, 84358512,
                         84358513, 84358514, 84358515, 84358516, 84358517]', 'num_inference_steps': 50,
                         'true_cfg_scale': 0, 'output_type': 'latent', 'width': 1024, 'height': 1024, 'parser':
                         'native'}
07:55:45-299832 DEBUG    Module: name=transformer cls=FluxTransformer2DModel size=12.089 params=11901408320
                         quant=QuantizationMethod.SDNQ
07:55:45-305396 DEBUG    Module: name=text_encoder_2 cls=T5EncoderModel size=7.370 params=4762310656
                         quant=QuantizationMethod.SDNQ
07:55:45-308408 DEBUG    Module: name=text_encoder cls=CLIPTextModel size=0.229 params=123060480 quant=None
07:55:45-311852 DEBUG    Module: name=vae cls=AutoencoderKL size=0.156 params=83819683 quant=None
Progress ?it/s                                              0% 0/50 00:00 ? Base07:57:43-979053 TRACE    Server: alive=True requests=266 memory=45.1/63.92 status='running' task='Load'
                         timestamp='20250809075224' current='task(1iagcw0lq9l74uz)' id='d06d22e01827487' job=0 jobs=1
                         total=1 step=0 steps=0 queued=0 uptime=326 elapsed=319.68 eta=None progress=0

SAC020 · 2025-08-09T05:13:22Z

SAC020
Aug 9, 2025
Author

My mistake, I somehow didn't realize I had batch size to 16. Works fine

0 replies

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

Flux Krea using 40GB+ VRAM with SDNQ - is this normal? #4104

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Replies: 1 comment

Uh oh!

{{title}}

Uh oh!

Select a reply

Uh oh!

Uh oh!

Flux Krea using 40GB+ VRAM with SDNQ - is this normal? #4104

Uh oh!

Uh oh!

SAC020 Aug 9, 2025

Replies: 1 comment

Uh oh!

SAC020 Aug 9, 2025 Author

SAC020
Aug 9, 2025

SAC020
Aug 9, 2025
Author