So, it's possible to train a Flux Lora with a 4070 TI Super 16gb with this tool?

[ZeroCool22]

Or only possible with kohya_ss?

Thx.

[bghira]

i'm mostly curious why you would want to when a 24G GPU is $0.15/hr on Vast or $0.22/hr on RunPod community cloud? a 5 hour training run will work better there and cost ~$0.75 to $1.00, right? what's the gain of running a low batch size on 16G and training a lower quality model?

[bghira]

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.90.07              Driver Version: 550.90.07      CUDA Version: 12.4     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4060 Ti     On  |   00000000:08:00.0 Off |                  N/A |
| 77%   66C    P2            147W /  165W |   15505MiB /  16380MiB |     32%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
+-----------------------------------------------------------------------------------------+

2024-09-03 01:48:56,730 [INFO] (helpers.training.trainer) Moving the diffusion transformer to GPU in int8-quanto precision.
2024-09-03 01:48:56,832 [INFO] (helpers.training.trainer) 
***** Running training *****
-  Num batches = 121308
-  Num Epochs = 1
  - Current Epoch = 1
-  Total train batch size (w. parallel, distributed & accumulation) = 1
  - Instantaneous batch size per device = 1
  - Gradient Accumulation steps = 1
-  Total optimization steps = 10000
-  Total optimization steps remaining = 10000
Epoch 1/1, Steps:   1%|▎                     | 117/10000 [13:44<21:45:41,  7.93s/it, lr=9.99e-5, mean_cfg=1, step_loss=0.264]

when your model is only as good as your batch size, this is just a bad idea. and you shouldn't train this way. but sure, it 'works'