Software Engineering Courses (SECourses)•5mo ago

Dataset just 20 images like last time. Command in Musumi-TUner to start: python3 src/musubi_tuner/

Dataset just 20 images like last time.

Command in Musumi-TUner to start:

python3 src/musubi_tuner/qwen_image_cache_latents.py --dataset_config dataset/dataset_qwen_test.toml --vae models/qwen_image/vae/vae/diffusion_pytorch_model.safetensors

python3 src/musubi_tuner/qwen_image_cache_text_encoder_outputs.py --dataset_config dataset/dataset_qwen_test.toml --text_encoder models/qwen_image/text_encoders/split_files/text_encoders/qwen_2.5_vl_7b.safetensors

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True accelerate launch --num_cpu_threads_per_process 1 src/musubi_tuner/qwen_image_train_network.py \
--dataset_config dataset/dataset_qwen_test.toml \
--dit models/qwen_image/diffusion_models/split_files/diffusion_models/qwen_image_bf16.safetensors \
--text_encoder models/qwen_image/text_encoders/split_files/text_encoders/qwen_2.so_vl_7b.safetensors \
--vae models/qwen_image/vae/vae/diffusion_pytorch_model.safetensors \
--output_dir output_qwen_advanced \
--output_name Qwen-LoRA-Arek-Advanced \
--mixed_precision bf16 \
--sdpa \
--optimizer_type adamw8bit \
--learning_rate 2e-4 \
--lr_scheduler cosine_with_restarts \
--lr_warmup_steps 300 \
--lr_scheduler_num_cycles 3 \
--network_module networks.lora_qwen_image \
--network_dim 32 \
--network_alpha 16 \
--network_args "loraplus_lr_ratio=4" \
--max_train_steps 3000 \
--save_every_n_steps 250 \
--gradient_checkpointing \
--timestep_sampling qinglong_qwen \
--dynamo_backend INDUCTOR \
--fp8_base \
--fp8_scaled \
--blocks_to_swap 20

AiInfluence•8/13/25, 6:41 PM

well , i still like krea more with the examples you shown yesterday

Sskyrrr What's your take on buckets ? Do you notice better results without ? I use diffe...

Furkan Gözükara SECourses•8/14/25, 10:27 AM

i dont prefer bucketing

Furkan Gözükara SECourses•8/14/25, 10:27 AM

but if you have certain aspect ratio and you will use same after training

Furkan Gözükara SECourses•8/14/25, 10:27 AM

it helps

Kkeyser_soze anyone trained wan 2.2 in msubai trainer , particularly in h100 80gb if you are...

Furkan Gözükara SECourses•8/14/25, 10:27 AM

i am coding an app now for qwen. it will also have wan 2.2. it is based on bmaltais implementation

Furkan Gözükara SECourses•8/14/25, 10:28 AM

Sakami•8/14/25, 10:29 AM

thank you soooo much

Ccyberbol Hey everyone, Kicking off a new training run for my Qwen-Image likeness LoRA to...

Furkan Gözükara SECourses•8/14/25, 10:30 AM

coding the interface to start testing hopefully

DDr. Odin I got a question..in one trainer.. when setting the captions from one txt file (...

Furkan Gözükara SECourses•8/14/25, 10:30 AM

it is entire image

Furkan Gözükara SECourses•8/14/25, 10:30 AM

so whatever in your training image entirely

Ccyberbol Dataset just 20 images like last time. Command in Musumi-TUner to start: pytho...

Furkan Gözükara SECourses•8/14/25, 10:31 AM

i think results can be made better

Furkan Gözükara SECourses•8/14/25, 10:31 AM

recently i compared optimizers on flux

Furkan Gözükara SECourses•8/14/25, 10:32 AM

and our workflow was most realistic one

FFurkan Gözükara SECourses so whatever in your training image entirely

Dr. Odin•8/14/25, 10:32 AM

oo thanks
also
I started cloud training today on OneTrainer, but I’m getting an SSH key authentication error in PowerShell. My public keys on RunPod and my PC match — I’ve verified this multiple times. I even used ChatGPT to help troubleshoot, but the issue persists.

Currently, I have no pod running before training. When I start training on OneTrainer, I get the authentication error in PowerShell, but on RunPod, the pod launches and shows as running and ready with no errors in the logs.

I’ve been advised to confirm that the SSH keys match, and they do. The last time I trained, everything worked fine, but today I’m facing this repeated error. On RunPod’s side, everything seems okay — the issue appears only in OneTrainer.
ive been at this since last night... ive gotten no where
Could you please assist in resolving this?

FFurkan Gözükara SECourses i am coding an app now for qwen. it will also have wan 2.2. it is based on bmalt...

n0tmad•8/14/25, 10:38 AM

wan 2.2 lora training?

n0tmad•8/14/25, 10:39 AM

will it include 14b parameter model @Furkan Gözükara SECourses

Timson•8/14/25, 10:46 AM

@Furkan Gözükara SECourses are flux trainign in main branch of kohya_ss now?

AiInfluence•8/14/25, 10:46 AM

you mean qwen?

Timson•8/14/25, 10:47 AM

My scripts broke because sd3-flux.1 branch is removed now

Timson•8/14/25, 10:47 AM

we used to use it

DDr. Odin oo thanks also I started cloud training today on OneTrainer, but I’m getting an ...

Furkan Gözükara SECourses•8/14/25, 10:49 AM

i recommend massed compute

Furkan Gözükara SECourses•8/14/25, 10:49 AM

it is installed there

Furkan Gözükara SECourses•8/14/25, 10:49 AM

just run update and use

Nn0tmad wan 2.2 lora training?

Furkan Gözükara SECourses•8/14/25, 10:49 AM

yes

TTimson @Furkan Gözükara SECourses are flux trainign in main branch of kohya_ss now?

Furkan Gözükara SECourses•8/14/25, 10:49 AM

yes for a long time

Furkan Gözükara SECourses•8/14/25, 10:49 AM

qwen image is in musubi tuner

Furkan Gözükara SECourses•8/14/25, 10:49 AM

i am making gui based on that

Furkan Gözükara SECourses•8/14/25, 10:49 AM

Furkan Gözükara SECourses•8/14/25, 10:49 AM

adding more info the gui settings right now

FFurkan Gözükara SECourses yes for a long time

Timson•8/14/25, 10:52 AM

I'ts been a while... do we need to manually reinstall torch and xformers still or is it fixed too in main?

FFurkan Gözükara SECourses and our workflow was most realistic one

cyberbolOP•8/14/25, 11:05 AM

Cant wait to see your reasult and advice.. Want to to my FINAL qwen and flux krea LORA or full model

FFurkan Gözükara SECourses i am making gui based on that

cyberbolOP•8/14/25, 11:08 AM

Dr.Furkan, if you making GUI please add all new function like :
--timestep_sampling option with qwen_shift
--qinglong_qwen
---dynamo_backend INDUCTOR

etc..

I personally liked to use command but will be great and easier when we have GUI

cyberbolOP•8/14/25, 11:09 AM

WIll be great also see later optimal config files for example 16gb vram, 24gb vram etc...

FFurkan Gözükara SECourses i think results can be made better

cyberbolOP•8/14/25, 11:11 AM

SO you thing my example pic are bad for some reason ?? Can you share you honest opinion ?

AiInfluence•8/14/25, 11:11 AM

@Furkan Gözükara SECourses when will you drop gui?

FFurkan Gözükara SECourses i recommend massed compute

Dr. Odin•8/14/25, 11:46 AM

whats the problem on runpod.. is OT not updated.. been using runpod and deposited what i had there.. i guess im not getting a refund but what update is missing

TTimson I'ts been a while... do we need to manually reinstall torch and xformers still o...

Furkan Gözükara SECourses•8/14/25, 12:24 PM

i think all in main

Furkan Gözükara SECourses•8/14/25, 12:24 PM

my installer properly installs

Ccyberbol Dr.Furkan, if you making GUI please add all new function like : --timestep_sampl...

Furkan Gözükara SECourses•8/14/25, 12:25 PM

check everyuthing and let me know what missing

Furkan Gözükara SECourses•8/14/25, 12:25 PM

still improving gui adding more info

Furkan Gözükara SECourses•8/14/25, 12:25 PM