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

feat: Update direct loading fp8 ckpt for LoRA training

Browse Source
This commit is contained in:
Kohya S
2024-08-27 21:40:02 +09:00
parent 0087a46e14
commit 3be712e3e0
6 changed files with 151 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
README.md
View File

@@ -9,13 +9,18 @@ __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 27, 2024 (update 2):
In FLUX.1 LoRA training, when `--fp8_base` is specified, the FLUX.1 model file with fp8 (`float8_e4m3fn` type) can be loaded directly. Also, in `flux_minimal_inference.py`, it is possible to load it by specifying `fp8 (float8_e4m3fn)` in `--flux_dtype`.
In `flux_merge_lora.py`, you can now specify the precision at save time with `fp8` (see `--help` for details). Also, if you do not specify the merge model, only the model type conversion will be performed.
Aug 27, 2024:
- FLUX.1 LoRA training now supports CLIP-L LoRA. Please remove `--network_train_unet_only`. T5XXL is not trained. The output of T5XXL is still cached, so `--cache_text_encoder_outputs` or `--cache_text_encoder_outputs_to_disk` is still required. The trained LoRA can be used with ComfyUI.
- `flux_extract_lora.py` and `convert_flux_lora.py` do not support CLIP-L LoRA.
- `--sigmoid_scale` is now effective even when `--timestep_sampling shift` is specified. Normally, leave it at 1.0. Larger values make the value before shift closer to a uniform distribution.
- __Experimental__ `--fp8_base_unet` option is added to `flux_train_network.py`. Flux can be trained with fp8, and CLIP-L can be trained with bf16/fp16. When specifying this option, the `--fp8_base` option is not required (Flux is fp8, and CLIP-L is bf16/fp16, regardless of the `--fp8_base` option).
- __Experimental__ `--fp8_base_unet` option is added to `flux_train_network.py`. Flux can be trained with fp8, and CLIP-L can be trained with bf16/fp16. When specifying this option, the `--fp8_base` option is automatically enabled.
Aug 25, 2024:
Added `shift` option to `--timestep_sampling` in FLUX.1 fine-tuning and LoRA training. Shifts timesteps according to the value of `--discrete_flow_shift` (shifts the value of sigmoid of normal distribution random number). It may be good to start with a value of 3.1582 (=e^1.15) for `--discrete_flow_shift`.