Software Engineering Courses (SECourses)•3y ago

@Dr. Furkan Gözükara do you use `no_half_vae` with the latest vae? I've noticed when it's enabled it

@Dr. Furkan Gözükara do you use

no_half_vae

no_half_vae

with the latest vae? I've noticed when it's enabled it forces the vae to be float32, which is causing my checkpoint saving to fail (it's expecting bf16)

Sstillmotion @Dr. Furkan Gözükara do you use `no_half_vae` with the latest vae? I've noticed ...

Furkan Gözükara SECourses•10/9/23, 10:40 PM

i dont use anymore

Furkan Gözükara SECourses•10/9/23, 10:40 PM

with fp16 vae

FFurkan Gözükara SECourses yes runpodctl

Stefanoid9343•10/9/23, 10:45 PM

Thanks!

FFurkan Gözükara SECourses 4090 much stronger but not working very good on windows. probably 3090 has the b...

mtnmecca•10/9/23, 10:50 PM

update - I just updated from 3060 12 gb vram to 4090 24gb vram and WOW. It is amazingly fast and visibly better quality. Switched back to Windows 10 before upgrade so I don't have a comparison to W11. Amazing! btw I haven't tried training yet since I follow your runpod advice and use 48 gb vram to train.

Mmtnmecca update - I just updated from 3060 12 gb vram to 4090 24gb vram and WOW. It is a...

Furkan Gözükara SECourses•10/9/23, 10:53 PM

i also noticed this

Furkan Gözükara SECourses•10/9/23, 10:54 PM

--medvram produces different images

thesilvermoth•10/10/23, 12:20 AM

Python is using 24GB of RAM when generating images with A1111. Is that normal?

Mmtnmecca update - I just updated from 3060 12 gb vram to 4090 24gb vram and WOW. It is a...

thesilvermoth•10/10/23, 12:24 AM

I also noticed the jump in quality. It has incredible details without hires fix

Tthesilvermoth Python is using 24GB of RAM when generating images with A1111. Is that normal?

Furkan Gözükara SECourses•10/10/23, 1:00 AM

ah ram

Furkan Gözükara SECourses•10/10/23, 1:01 AM

i think caching

Furkan Gözükara SECourses•10/10/23, 1:01 AM

so normal

FFurkan Gözükara SECourses yep

Elisoli•10/10/23, 10:10 AM

If you put of a survey of most wanted tutorials i think itll maybe easier for you to see priorities (for patreon supporters)

Mmtnmecca update - I just updated from 3060 12 gb vram to 4090 24gb vram and WOW. It is a...

Elisoli•10/10/23, 10:11 AM

so no bugs using 4090 with automatic1111\kohya ?

virginia•10/10/23, 10:14 AM

hello, all, for the training note book, i found out no place to like this part, tags

virginia•10/10/23, 10:55 AM

i mean on kaggle note book, where to input --caption_extension

virginia•10/10/23, 10:55 AM

i do not see

mayna•10/10/23, 2:12 PM

Hello, Started to follow your course on youtube for quality studio images,
when installed kohya SS gui I don't have dreambooth Lora but only LORA , should I continue ? @Dr. Furkan Gözükara

mayna•10/10/23, 2:13 PM

is that normal ?

virginia•10/10/23, 2:14 PM

hello, all what is the trigger word of lora? instance name ? if i want to use revanimated as base model to train 1.5 lora, also huggingface repo id?

Vvirginia hello, all what is the trigger word of lora? instance name ? if i want to use re...

mayna•10/10/23, 2:21 PM

Hey Virginia, does Google Collab free again ?? it was recently changed to not free anymore

Deleted User•10/10/23, 2:27 PM

Deleted User•10/10/23, 2:27 PM

Running on local URL: http://127.0.0.1:7860

To create a public link, set

share=True

share=True

launch()

launch()

Deleted User•10/10/23, 2:28 PM

How to make it public

Deleted User•10/10/23, 2:28 PM

I am using mobile device

virginia•10/10/23, 2:39 PM

no colab cannot be used

Mmayna Hello, Started to follow your course on youtube for quality studio images, when...

Sendai•10/10/23, 2:56 PM

It should only say Lora at the top

virginia•10/10/23, 3:04 PM

1024 sdxl fail again ------------------------------------------------------------
Failures:
[1]:
time : 2023-10-10_15:03:15
host : 19fe5077aa6e
rank : 1 (local_rank: 1)
exitcode : 2 (pid: 1098)
error_file: <N/A>
traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
------------------------------------------------------------
Root Cause (first observed failure):
[0]:
time : 2023-10-10_15:03:15
host : 19fe5077aa6e
rank : 0 (local_rank: 0)
exitcode : 2 (pid: 1097)
error_file: <N/A>
traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html

virginia•10/10/23, 3:04 PM

train_network.py: error: argument --max_data_loader_n_workers: expected one argument
train_network.py: error: argument --max_data_loader_n_workers: expected one argument
Traceback (most recent call last):
File "/opt/conda/bin/accelerate", line 8, in <module>
sys.exit(main())
File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/accelerate_cli.py", line 47, in main
args.func(args)
File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/launch.py", line 977, in launch_command
multi_gpu_launcher(args)
File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/launch.py", line 646, in multi_gpu_launcher
distrib_run.run(args)
File "/opt/conda/lib/python3.10/site-packages/torch/distributed/run.py", line 785, in run
elastic_launch(
File "/opt/conda/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 134, in call
return launch_agent(self._config, self._entrypoint, list(args))
File "/opt/conda/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 250, in launch_agent
raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:

Zono50•10/10/23, 3:18 PM

trying to put my face on this image, but it's not going well, tried regular lora with positive prompt, roop and auto detailer, but never gets the face well

Zono50•10/10/23, 3:18 PM

I assume it has to do with distance and darkness

aelvelis•10/10/23, 3:21 PM

Hi friends, I have this problem: Very slow image generation speed when using SDXL 1.0 on RTX 3060 Ti GPU and 16gb RAM. If i generate 1024x1024 with 20 steps then the generation time is 10 minutes, if for example 768x1024 then the generation speed is from 20 minutes. I have seen many people have generation speeds of only a couple minutes on the same GPU. I am using AUTOMATIC1111 version v1.6.0 and the start parameter --opt-sdp-attention. I don't get any errors about lack of VRAM, just a very slow generation speed. Maybe I should do a clean SD installation?

aelvelis•10/10/23, 3:22 PM

I enabled sdxl.vae and got a reduction in image generation time. But I got an error. Anyway it's not a 10 minutes wait anymore:)
When generating at 40 steps, the generation time doubles.

EElisoli so no bugs using 4090 with automatic1111\kohya ?

mtnmecca•10/10/23, 3:29 PM

none with A1111 so far and haven't tried kohya yet. Will in the next few days.

mtnmecca•10/10/23, 3:41 PM

What are other good rare tokens to use for training besides ohwx? I would like to train other subjects

Tthesilvermoth I also noticed the jump in quality. It has incredible details without hires fix

mtnmecca•10/10/23, 3:49 PM

I also just noticed the change in processing units. Never paid attention before but I was getting 1.6 sec/Iteration on 3060 and now I'm getting 0.31 sec/iteration (or 3.21 iterations/sec). 4-5x faster and better quality

aelvelis•10/10/23, 4:01 PM

I think I was able to fix the slow generation speed on the 3060 ti. I now generate a 768x1024 image in 30-40 seconds. The problem is with Nvidia drivers from version 531 and up. And also I used --xformers parameter instead of --opt-sdp-attention.
https://github.com/vladmandic/automatic/discussions/1285

GitHub

nVidia drivers change in memory management · vladmandic automatic ·...

it seems that nVidia changed memory management in the latest versions of drivers, specifically 532 and 535 new behavior is that once gpu vram is exhausted, it will actually use shared memory thus c...