Jaret Burkett
3530876838
Merge branch 'main' into flux3-image
2026-10-14 13:41:20 -06:00
Jaret Burkett
7045370346
Keep cached pixel tensors as uint8 in the dataloader workers and only materialize to fp16 on the main thread
2026-10-14 13:40:49 -06:00
Jaret Burkett
7a64fdafad
Merge in from main
2026-10-14 11:27:16 -06:00
Jaret Burkett
2fa0bd87e6
Add support for Iris-3B
2026-10-09 05:13:39 -06:00
Jaret Burkett
cb51e9a511
Add flux3 image
2026-10-09 03:08:46 -06:00
Jaret Burkett
a949f21f8e
Add ability to merge dataset buckets. Useful when using numerous different datasets.
2026-10-08 14:04:30 -06:00
Jaret Burkett
0cbac2c1f3
Add threadding to bucket resizing of a batch. Helpful to speed up large batch gathering.
2026-10-08 11:26:48 -06:00
Jaret Burkett
6468a2ffa7
Fix issue loading new comfy checkpoints with ming
2026-09-26 13:32:05 -06:00
Jaret Burkett
16ae8d2999
Fixed issue on DGX spark that caused unnessary moves of models causing slow down on unified memory
2026-09-26 11:56:42 -06:00
Jaret Burkett
07abdbedc7
Add match_target_res for Qwen Image 2.1 so reference images are scaled to the target's pixel area with their own aspect ratio, consistently across text-embedding caching, training, and sampling, and key the embedding cache on it.
2026-09-22 13:57:49 -06:00
Jaret Burkett
a35d833abf
Fix text encoder unload skipping models whose pipeline does not expose the text encoder (Qwen-Image-2.1, Z-Image, LTX2), so cached-embedding runs actually free the TE.
2026-09-22 12:08:29 -06:00
Jaret Burkett
543a0d74d0
Fix issue with saving device state when a text encoder is unloaded.
2026-09-22 12:03:09 -06:00
Jaret Burkett (Ostris)
3a39d8cfc8
Merge pull request #1021 from hitoshic1982/fix/cache-text-encoder-device-preset
...
Fix cache_text_encoder device preset missing in BaseModel
2026-09-22 11:46:24 -06:00
Jaret Burkett
a8dfcf7d7e
Save thumbnail with alpha if there is one.
2026-09-20 15:16:15 -06:00
Jaret Burkett
086b663bba
Add support for training with an alpha channel for Qwen Image 2.1
2026-09-20 14:49:53 -06:00
Jaret Burkett
c2622ed250
Add support for Qwen Image 2.1
2026-09-20 09:50:57 -06:00
Jaret Burkett
cc98c3b690
Add ability to train with just music or vocals on yue2. Doesnt work as expected though
2026-09-17 20:19:32 -06:00
Jaret Burkett
a9e7752178
Add MOSS music audio captioner
2026-09-17 10:45:37 -06:00
Jaret Burkett
080873bec9
Harden ace step transcription engine to detect and break out of loops. Add melbanroformer to seperate out the vocals and music for better transcriptions. Much better results.
2026-09-17 09:49:19 -06:00
Jaret Burkett
1db6083645
Show more information about LoRAs when loading loras in generate page
2026-09-17 09:35:43 -06:00
Jaret Burkett
410206e41c
Improve speed of sage sheet decoder with yue2. Recover from tracks that fail
2026-09-16 12:24:39 -06:00
Jaret Burkett
ce64e65c39
Add support for Qwen 2.5 Omni and add lora loading for captioning models
2026-09-16 10:36:19 -06:00
Jaret Burkett
2977672e4b
Point omni vllm to real repo
2026-09-16 09:37:21 -06:00
Jaret Burkett
db8dbd68c2
Add support for training VLLMs, omni models, and LLMs
2026-09-15 17:47:02 -06:00
Jaret Burkett
f9b23b40bb
Load sage sheet and use it on training with 50% dropout so it generates sheet music and handles without it.
2026-09-14 17:01:36 -06:00
Jaret Burkett
e43f138ae4
Initial experimental support for yue2
2026-09-14 05:07:11 -06:00
Jaret Burkett
82137c8ca0
Fix album art issue with music files
2026-09-13 18:08:37 -06:00
Jaret Burkett
881143cbe8
Fixed issue with EMA climbing decay instead of reducing it.
2026-09-12 21:00:31 -06:00
Jaret Burkett
737ee20733
Add lokr support to inference engine
2026-09-12 14:17:44 -06:00
Jaret Burkett
732c761231
Add lora on generate
2026-09-10 18:23:23 -06:00
Jaret Burkett
449488f6b7
Initial inference engine build out
2026-09-10 14:22:49 -06:00
Jaret Burkett
f56b5a1d40
Change hidream to be proper x0 prediction and use a custom weighing scheme for x0 models.
2026-09-10 11:49:54 -06:00
Jaret Burkett
b1bf3e4381
Add a config to force pure noise step at a set rate
2026-09-09 11:26:28 -06:00
Jaret Burkett
4a41f146a3
Fix feedback when doing an ema with feedback so it uses its own scale
2026-09-09 06:43:14 -06:00
Jaret Burkett
b764c4a94e
DFE cleanup
2026-09-09 06:42:24 -06:00
Jaret Burkett
af82197907
Fix gradient checkpointing for sapiens DFE
2026-09-08 12:24:56 -06:00
Jaret Burkett
af89a27e0f
Fixed issue with partial offloading torchao quants
2026-09-08 05:15:26 -06:00
Jaret Burkett
3e1cb5f67f
Rework DFEs to work better with x0 models. Show their loss in the graph
2026-09-07 13:26:01 -06:00
Jaret Burkett
7690ea62c8
Add variable precision adam convrot optimizer with stochastic rounding and accumulation.
2026-09-06 10:51:35 -06:00
Jaret Burkett (Ostris)
2a2ca26ee8
Merge branch 'main' into fix/dataloader-pin-memory
2026-09-05 08:08:39 -06:00
Jaret Burkett
ecb2bfbe5e
Fixed issue where some unmanaged modules could be lefs on cpu
2026-09-02 09:22:39 -06:00
Jaret Burkett
9d6a9a0803
Fixed embedding scale for offloading for ltx 2.3. Was a new bug added today.
2026-08-31 15:53:58 -06:00
Jaret Burkett
6940ebf533
Skip files where caching fails in a dataset instead of crashing.
2026-08-30 16:49:12 -06:00
Jaret Burkett
74ed5fddb0
Fix issue with ltx caching
2026-08-30 11:01:52 -06:00
Jaret Burkett
764b5064fb
Migrate to a new DTO for latents to carry more information that a normal tensor such as audio.
2026-08-30 10:30:21 -06:00
Jaret Burkett
be995185f5
Fixed issue with offloading with new model loader on nvfp4 weights.
2026-08-29 12:01:40 -06:00
Jaret Burkett
683fe8afc0
Stability improvements to model offloading. Added D-OPSD bleed loss as well.
2026-08-29 11:23:47 -06:00
Jaret Burkett
7195abc32b
Performance fixes and bug fixes
2026-08-28 22:21:09 -06:00
Jaret Burkett
5ddc5f8ca7
Optimize quantization path
2026-08-28 11:47:29 -06:00
hitoshic1982
5fc8fd76de
Fix cache_text_encoder device preset missing in BaseModel
...
BaseModel.set_device_state_preset() has no branch for the
'cache_text_encoder' preset, so when TextEmbeddingCachingMixin calls
set_device_state_preset('cache_text_encoder') the active_modules list
stays empty and the text encoder is placed on CPU. With
cache_text_embeddings enabled on a BaseModel arch (e.g. chroma), every
caption is then encoded on CPU — with a quantized T5-XXL this took
~6.5 minutes per caption (~21 hours for a 205-image dataset) while the
GPU sat idle. StableDiffusion.set_device_state_preset() already has
this branch; this adds the same two lines to BaseModel.
2026-08-29 01:37:39 +08:00
Jaret Burkett
hitoshic1982