Commit Graph
1092 Commits
Author SHA1 Message Date
Booyaka101 0c8793ff47 fix: make pin_memory opt-in via dataset config (default off)
@zappazack found a real regression: pin_memory page-locks RAM that
NVIDIA's Windows driver wants for shared-memory VRAM-overflow fallback
when VRAM saturates. Page-locked pages can't be relocated, so the driver
falls back to PCIe thrashing - step times went from 1.5s to 20+s when
VRAM saturated on RTX 5090 + Qwen LoRA + 1024x1024.

The original "strict-improvement" claim was wrong. Convert to opt-in:

- Add `pin_memory: bool` field on DatasetConfig (default False)
- Only pin when both CUDA is available AND user opts in via config

Users with stable VRAM headroom can opt in for the transfer speedup.
Users at the VRAM ceiling stay safe with the default.

Refs #758.
2026-05-02 11:02:41 +08:00
Booyaka101 67e223b502 perf: enable pin_memory on CUDA DataLoaders for faster CPU->GPU transfer
torch DataLoader's pin_memory was never set, defaulting to False. Setting
pin_memory=True allocates batch tensors in page-locked memory, which lets
PyTorch issue async cuda host->device copies via DMA - typically 10-30%
faster transfer for image/video training tensors and can reduce per-step
overhead noticeably on Windows where num_workers is forced to 0 (see
6d31c6db / #770) and the main process pays the full transfer cost.

pin_memory is orthogonal to num_workers and is safe with num_workers=0,
so this benefits Windows/macOS users (who currently can't use multi-worker
loading due to pickling constraints) without changing any other behavior.
Gated on torch.cuda.is_available() so it's a no-op on CPU/MPS/ROCm-without
-CUDA-API setups.

Refs #758.
2026-05-01 23:07:43 +08:00
Jaret Burkett 0d91fcee9e Allow edit of captioning job 2026-04-30 16:00:16 -06:00
Jaret Burkett eadc9a58af Version bump 2026-04-30 06:00:58 -06:00
Jaret Burkett e9ab387dfd Fixed issue with qwen image edit models when using multiple control images when not caching text embeddings. 2026-04-30 11:59:12 +00:00
Jaret Burkett deb409085a Made it possible to use the flux 2 small decoder VAE when setting the vae_path manually for flux2 models 2026-04-30 05:42:31 -06:00
Jaret Burkett 7ccec8ec2c Add checkpointing and a proper decode for flux 2 VAEs so they can be used with DFE 2026-04-30 04:27:13 -06:00
Jaret Burkett b4f0efb025 Version Bump 2026-04-28 13:42:00 -06:00
Jaret Burkett af6458d1b5 Enable caching of ACE step latents. 2026-04-28 13:39:20 -06:00
Jaret Burkett 77b8765939 Fix issue with hover styling on select components. 2026-04-28 10:55:20 -06:00
Jaret Burkett 43989cc19e Add advanced config section to the captioner 2026-04-28 10:49:46 -06:00
Jaret Burkett f972b750e6 Performance improvements to captioner. Add ability to set a default caption for ace captioner to avoid needing to caption the audio so we only transcribe 2026-04-28 10:15:11 -06:00
Jaret Burkett acc6a36214 Scale DFE 9 to a velocity equiv weight to match flow matching gradient strength. Probably need to rework all DFEs to do this as the math checks out. 2026-04-28 09:10:02 -06:00
Jaret Burkett 1fc4ad3979 Add sapiens2 as a diffusion feature extractor 2026-04-27 15:59:03 -06:00
Jaret Burkett 67d67f8c1d Ignore hidden files when captioning 2026-04-25 17:38:31 -06:00
Jaret Burkett 998a02f30e Yet another pseudo_huber fix. Still drinking coffee. 2026-04-19 10:05:40 -06:00
Jaret Burkett fc85410c9a Fix issue with precision on pseudo_huber loss 2026-04-19 10:02:27 -06:00
Jaret Burkett 20a99258b8 Fix issue is is else on pseudo_huber loss 2026-04-19 09:59:18 -06:00
Jaret Burkett f4445cd78c Added psuedo_huber loss 2026-04-19 09:51:46 -06:00
Jaret Burkett 488878f354 Use hidden layers in the loss for DFE 7 and 8 2026-04-18 13:07:38 -06:00
Jaret Burkett beb40ae29b Add DFE8 with partial step 2026-04-17 17:40:16 -06:00
Jaret Burkett 7c4f18ce51 Fix ernie unpatchify 2026-04-17 12:03:39 -06:00
Jaret Burkett 8cb9649382 Add decode latent to ernie pipe 2026-04-17 12:01:09 -06:00
Jaret Burkett 67048df9f9 Merge branch 'main' of github.com:ostris/ai-toolkit 2026-04-17 06:05:40 -06:00
Jaret Burkett be54094704 Version bump 2026-04-17 06:05:26 -06:00
Jaret Burkett a513a1583e Fixed issue where Qwen VL MOE captioner produced nonsense 2026-04-16 21:24:43 +00:00
Jaret Burkett 22ea3dd620 Fixed issue on some systems where Logger didnt have atty 2026-04-16 21:09:52 +00:00
Jaret Burkett ab1ee4df34 Hotfix some issues with Wan models caused by diffusers and transformers updates 2026-04-16 20:53:50 +00:00
Jaret Burkett 0c18b39346 Version bump 2026-04-16 13:09:45 -06:00
Jaret Burkett afb62b1fa5 Add support for Nucleus-Image 2026-04-16 13:09:10 -06:00
Jaret Burkett 2faba22b46 Fix issue when saving advanced prompt embeds. No such file or directory error 2026-04-16 12:22:56 -06:00
Jaret Burkett 0792352dab version bump 2026-04-16 09:25:14 -06:00
Jaret Burkett 8f67f5022e Version lock peft version 2026-04-15 14:17:26 -06:00
Jaret Burkett acc3e60140 Add ernie to readme 2026-04-15 06:43:36 -06:00
Jaret Burkett dd7074a21f Fix issue with layer offloading on ernie 2026-04-14 19:19:02 -06:00
Jaret Burkett e74bc9ac7b Fix issue with concatinating advanced prompt embeds. 2026-04-14 16:04:34 -06:00
Jaret Burkett 7eb1226a6d Fix issue with loading advanced prompt configs metadata 2026-04-14 15:42:09 -06:00
Jaret Burkett 97d8c05d75 Fix issue with captioner sometimes outputting not utf-8 characters 2026-04-14 14:37:21 -06:00
Jaret Burkett (Ostris) 3e0c904054 Add support for Baidu's ERNIE-Image (#793)
* Add support for ERNIE Image

* change float64 to float32

* Version bump

* Update ERNIE defaults
2026-04-14 09:45:12 -06:00
Asaf Agami e868fca562 fix custom_flowmatch_sampler (#783) 2026-04-13 09:42:05 -06:00
Jaret Burkett 233e292256 Added some experimental low step things for zeta 2026-04-13 09:37:34 -06:00
Jaret Burkett 1058ef3513 Made AdvancedPromptEmbeds that is compatable with previous PromptEmbeds functionality, but is more streamlined and can accomidate more model embedding paradigms. 2026-04-11 10:45:47 -06:00
Jaret Burkett 0d11be41fa Adjust default sample steps. 2026-04-10 17:46:34 -06:00
Jaret Burkett 62e18427b4 Adjust default sample steps. 2026-04-10 17:46:09 -06:00
Jaret Burkett 9b4e2d1b0b More flac support 2026-04-10 12:27:09 -06:00
Jaret Burkett 0b9c365acb Add flac and ogg support 2026-04-10 12:10:54 -06:00
Jaret Burkett bfb373c8fa Prep for future breaking changes in newer versions of transformers library 2026-04-10 12:04:32 -06:00
Jaret Burkett 145144eee3 Fix issue with auto updating captions when captioning a dataset 2026-04-10 11:51:18 -06:00
Jaret Burkett 765a9d5b2e Add a download button to music samples in in the gear menu 2026-04-10 10:07:45 -06:00
Jaret Burkett d08ea8318f Change default timestep type for ace step to linear 2026-04-09 17:39:00 -06:00