VictorMylle/Kohya-ss-sd-scripts
1
0
Fork 1
mirror of https://github.com/kohya-ss/sd-scripts.git synced 2026-04-09 06:45:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix bug in FLUX multi GPU training

Browse Source
This commit is contained in:
kohya-ss
2024-08-22 12:37:41 +09:00
parent e1cd19c0c0
commit 98c91a7625
8 changed files with 156 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
README.md
View File

@@ -9,6 +9,12 @@ __Please update PyTorch to 2.4.0. We have tested with `torch==2.4.0` and `torchv
The command to install PyTorch is as follows:
`pip3 install torch==2.4.0 torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu124`
Aug 22, 2024:
Fixed a bug in multi-GPU training. It should work with fine-tuning and LoRA training. `--double_blocks_to_swap` and `--single_blocks_to_swap` cannot be used in multi-GPU training.
`--disable_mmap_load_safetensors` option now works in `flux_train.py`. It speeds up model loading during training in WSL2. It is also effective in reducing memory usage when loading models during multi-GPU training. Please always check if the model is loaded correctly, as it uses a custom implementation of safetensors loading.
Aug 21, 2024 (update 3):
- There is a bug that `--full_bf16` option is enabled even if it is not specified in `flux_train.py`. The bug will be fixed sooner. __Please specify the `--full_bf16` option explicitly, especially when training with 24GB VRAM.__
- Stochastic rounding is now implemented when `--fused_backward_pass` is specified. The implementation is