Commit Graph

568 Commits

Author SHA1 Message Date
Jaret Burkett
0d8a33dc16 Offload ARA with the layer if doing layer offloading. Add support to offload the LoRA. Still needs optimizer support 2025-10-21 06:03:27 -06:00
Jaret Burkett
76ce757e0c Added initial support for layer offloading wit Wan 2.2 14B models. 2025-10-20 14:54:30 -06:00
Jaret Burkett
1f81bc4060 Fix issue where text encoder could be the wrong quantization and fail when using memory manager 2025-10-15 11:01:30 -06:00
Jaret Burkett
7abf5e20be Add conv3d to memory management excluded modules 2025-10-15 10:12:06 -06:00
Jaret Burkett
1bc6dee127 Change auto_memory to be layer_offloading and allow you to set the amount to unload 2025-10-10 13:12:32 -06:00
Jaret Burkett
55b8b0e23e Fix issue where ARA was not working when using memory manager 2025-10-07 13:39:44 -06:00
Jaret Burkett
c9f982af83 Add support for using quantized models with ramtorch 2025-10-06 13:46:57 -06:00
Jaret Burkett
dc1cc3e78a Fixed issue where multi control samples didnt work when not caching 2025-10-05 14:38:53 -06:00
Jaret Burkett
4e5707854f Initial support for RamTorch. Still a WIP 2025-10-05 13:03:26 -06:00
Jaret Burkett
3086a58e5b git status 2025-10-01 14:12:17 -06:00
Jaret Burkett
3b1f7b0948 Allow user to set the attention backend. Add method to recomver from the occasional OOM if it is a rare event. Still exit if it ooms 3 times in a row. 2025-09-27 08:56:15 -06:00
Jaret Burkett
be990630b9 Remove dropout from cached text embeddings even if used specifies it so blank prompts are not cached. 2025-09-26 11:50:53 -06:00
Jaret Burkett
1069dee0e4 Added ui sopport for multi control samples and datasets. Added qwen image edit 5209 to the ui 2025-09-25 11:10:02 -06:00
Jaret Burkett
454be0958a Initial support for qwen image edit plus 2025-09-24 11:39:10 -06:00
Jaret Burkett
f74475161e Add stepped loss type 2025-09-22 15:50:12 -06:00
Jaret Burkett
28728a1e92 Added experimental dfe 5 2025-09-21 10:48:52 -06:00
Jaret Burkett
390e21bec6 Integrate dataset level trigger words and allow them to be cached. Default to global trigger if it is set. 2025-09-18 03:29:18 -06:00
Jaret Burkett
3cdf50cbfc Merge pull request #426 from squewel/prior_reg
Dataset-level prior regularization
2025-09-18 03:03:18 -06:00
squewel
e27e229b36 add prior_reg flag to FileItemDTO 2025-09-18 02:09:39 +03:00
max
e4ae97e790 add dataset-level distillation-style regularization 2025-09-18 01:11:19 +03:00
Jaret Burkett
218f673e3d Added support for new concept slider training script to CLI and UI 2025-09-16 10:22:34 -06:00
Jaret Burkett
3666b112a8 DEF for fake vae and adjust scaling 2025-09-12 18:09:08 -06:00
Jaret Burkett
b95c17dc17 Add initial support for chroma radiance 2025-09-10 08:41:05 -06:00
Jaret Burkett
af6fdaaaf9 Add ability to train a full rank LoRA. (experimental) 2025-09-09 07:36:25 -06:00
Jaret Burkett
f699f4be5f Add ability to set transparent color for control images 2025-09-02 11:08:44 -06:00
Jaret Burkett
85dcae6e2b Set full size control images to default true 2025-09-02 10:30:42 -06:00
Jaret Burkett
7040d8d73b Preperation for audio 2025-09-02 07:26:50 -06:00
Jaret Burkett
9ef425a1c5 Fixed issue with training qwen with cached text embeds with a batch size more than 1 2025-08-28 08:07:12 -06:00
Jaret Burkett
1f541bc5d8 Changes to handle a different DFE arch 2025-08-27 11:05:16 -06:00
Jaret Burkett
ea01a1c7d0 Fixed a bug where samples would fail if merging in lora on sampling for unquantized models. Quantize non ARA modules as uint8 when using an ARA 2025-08-25 09:21:40 -06:00
Jaret Burkett
f48d21caee Upgrade a LoRA rank if the new one is larger so users can increase the rank on an exiting training job and continue training at a higher rank. 2025-08-24 13:40:25 -06:00
Jaret Burkett
5c27f89af5 Add example config for qwen image edit 2025-08-23 18:20:36 -06:00
Jaret Burkett
bf2700f7be Initial support for finetuning qwen image. Will only work with caching for now, need to add controls everywhere. 2025-08-21 16:41:17 -06:00
Jaret Burkett
8ea2cf00f6 Added training to the ui. Still testing, but everything seems to be working. 2025-08-16 05:51:37 -06:00
Jaret Burkett
3413fa537f Wan22 14b training is working, still need tons of testing and some bug fixes 2025-08-14 13:03:27 -06:00
Jaret Burkett
be71cc75ce Switch to unified text encoder for wan models. Pred for 2.2 14b 2025-08-14 10:07:18 -06:00
Jaret Burkett
e12bb21780 Quantize blocks sequentialls without a ARA 2025-08-14 09:59:58 -06:00
Jaret Burkett
3ff4430e84 Fix issue with fake text encoder unload 2025-08-14 09:33:44 -06:00
Jaret Burkett
85bad57df3 Fix bug that would use EMA when set false 2025-08-13 11:39:40 -06:00
Jaret Burkett
77b10d884d Add support for training with an accuracy recovery adapter with qwen image 2025-08-12 08:21:36 -06:00
Jaret Burkett
bb6db3d635 Added support for caching text embeddings. This is just initial support and will probably fail for some models. Still needs to be ompimized 2025-08-07 10:27:55 -06:00
Jaret Burkett
5d8922fca2 Add ability to designate a dataset as i2v or t2v for models that support it 2025-08-06 09:29:47 -06:00
Jaret Burkett
9da8b5408e Initial but untested support for qwen_image 2025-08-04 13:29:37 -06:00
Jaret Burkett
9dfb614755 Initial work for training wan first and last frame 2025-08-04 11:37:26 -06:00
Jaret Burkett
f453e28ea3 Fixed deprecation of lumina pipeline error 2025-07-29 08:26:51 -06:00
Jaret Burkett
ca7c5c950b Add support for Wan2.2 5B 2025-07-29 05:31:54 -06:00
Jaret Burkett
cefa2ca5fe Added initial support for Hidream E1 training 2025-07-27 15:12:56 -06:00
Jaret Burkett
77dc38a574 Some work on caching text embeddings 2025-07-26 09:22:04 -06:00
Jaret Burkett
c5eb763342 Improvements to VAE trainer. Allow CLIP loss. 2025-07-24 06:50:56 -06:00
Jaret Burkett
f500b9f240 Add ability to do more advanced sample prompt objects to prepart for a UI rework on control images and other things. 2025-07-17 07:13:35 -06:00